Time-resolved functional genomics using deep learning reveals global hierarchical control of autophagy

Abstract

Recycling of cellular components through autophagy maintains homeostasis in changing nutrient environments. Although its core mechanisms are extensively studied, understanding of its systems-wide dynamic regulation remains limited, particularly regarding how autophagy is inactivated once nutrients are restored. Here we mapped the genetic network that controls activation and inactivation of autophagy during nitrogen changes by combining time-resolved high-content imaging, deep learning and latent feature analysis. This dataset, termed AutoDRY, categorizes 5,919 mutants based on nutrient response kinetics and their contributions to autophagosome formation and clearance. Integrating these profiles with functional and genetic network data uncovered hierarchical and multilayered control of autophagy and revealed multiple new regulatory pathways. Notably, we identified the retrograde pathway as a pivotal time-varying modulator that tunes the expression of core autophagy genes and plays a central role in autophagy inactivation. Together, this study establishes a systems-level resource to guide future investigations of autophagy.

Chica et al. combined genetic network mapping, time-resolved imaging and deep learning to generate a dataset called AutoDRY that describes multilayered control of autophagy and uncovers new regulatory pathways.

Main

Living organisms dynamically tune their metabolism to changes in nutrient availability. Macroautophagy, hereafter autophagy, sustains cellular homeostasis by adaptively degrading and recycling cellular components in response to nutrient scarcity, as well as other environmental cues, and its fine tuning is crucial for a healthy lifespan^1,2. This is evidenced by tight multilayered control of the autophagy core machinery and the association of several autophagy-related genes (ATGs) with human disorders^2–4. Characterization of autophagy execution has been the subject of intense research and several regulatory pathways have been identified^5–9. However, a generalized understanding of how living organisms dynamically control autophagic activity in the context of a broader genetic landscape is lacking. Along these lines, pharmacological manipulation of autophagy remains rudimentary and involves either complete inhibition of the process or strong activation through the use of TORC1 inhibitors, which have substantial side effects on growth and cell proliferation¹⁰. A systems-level understanding of the tuning of autophagic activity may help develop predictive models and support the development of effective treatments for autophagy-related conditions.

Whereas emerging studies utilizing multi-omics approaches, mathematical modelling and network analyses have begun to elucidate the complex landscape of potential autophagy regulators, systematic genome-wide data on causal influences on the autophagic nutrient response are limited^11–15. To date, the study of autophagy has largely been based on gain- or loss-of-function discoveries from autophagic flux assays, which measure the cumulative degradation or redistribution of cargo over a fixed time window^16,17. This approach has been successful in elucidating the central mechanisms orchestrating autophagy as well as proximal upstream switches acting in response to starvation. Genome-wide screens of autophagy-related mutants in yeast have identified only a limited number of new genes because they screened solely for loss in viability during starvation¹⁸ or for changes in autophagy flux at a single time point¹⁹. However, these assays do not fully capture the dynamic nature of autophagy regulation and may overlook important genomic factors influencing activation thresholds, sensitivity to nutrient shifts and more distant homeostatic tuning^17,20–23. These shortcomings also extend to a limited focus on the deactivation of autophagy and return to homeostasis following nutrient replenishment. Moreover, systematic information on the regulation of different stages of autophagy, such as the formation and clearance of autophagosomes, is lacking. To address these limitations, more sensitive real-time recordings of changes in autophagy during nutrient depletion and repletion are needed along with a systematic analysis of cellular features associated with distinct stages of the process.

In this study we performed a functional analysis of the genome-wide regulation of autophagy dynamics in Saccharomyces cerevisiae. Utilizing time-resolved high-content imaging combined with deep learning, we quantified the nutrient response kinetics and differential influences on autophagy execution in 5,919 yeast gene mutants (90% of the yeast genome) with high precision. By assembling quantitative regulatory profiles and integrating these with genome-wide resources that exist in yeast, we constructed a comprehensive map of the autophagy network response to shifts in nitrogen supply. Our approach reveals previously unexplored gene- and systems-level regulators of autophagy dynamics, including phase-dependent control of autophagic flux and multiple pathways for inactivation of autophagy.

Results

Genome-wide influences on autophagy dynamics

To map the genome-wide control of autophagy, we used an arrayed library of 4,760 deletion mutants and 1,159 decreased abundance by messenger RNA perturbation (DAmP) mutants expressing the autophagosome marker monomeric NeonGreen (mNG)–Atg8 and the vacuole-resident protease Pep4–mCherry. The mutants were cultured to logarithmic phase in nitrogen-containing medium (+N), and then imaged using high-content fluorescence microscopy every hour for 12 h during nitrogen starvation (−N) and for 7 h following nitrogen replenishment (+N; Fig. 1a). Each array included triplicate wells of wild-type (WT) cells and autophagy induction-deficient atg1Δ mutant cells as controls. We also included triplicates of vam6Δ mutant cells, which exhibit defects in Gtr1-dependent TORC1 activation and homotypic fusion with the vacuole, resulting in an abundance of uncleared autophagosomes^24–26.

Fig. 1. Deep learning-based profiling of autophagy nutritional response kinetics.

a, Study workflow overview. (i) Autophagic reporter integration into a yeast gene deletion library with hourly image capture during starvation (−N) and nitrogen replenishment (+N). (ii) Cell segmentation and extraction of 30 image features (p). (iii) Neural network training to predict autophagy probability (P0, non-autophagy; P1, autophagy) from control data. (iv) DNN output and latent space are used to predict autophagy response and states in mutants. (v) Computation of mutant perturbations as differentials from the plate median, where X refers to a kinetic parameter estimated from the response curves. P, parameters. b, Predicted autophagy response as a percentage of cells in autophagy for KO and DAmP collection strains. Median (solid line), negative plate control mean (dashed line) and s.d. (ribbon). c, Double-sigmoidal model fitting was used to measure kinetic perturbations and interpolate the autophagy responses. d, The HMP correlates with the minimum false discovery rate (FDR) over the set of hypothesis tests performed and is more sensitive than FDR for mutants with multiple significant perturbations. Distributions and percentages indicate the proportions of DAmP collection mutants; Fisher’s exact test (two-sided). e, Distribution of mutants with significant parameter perturbations with P ≤ 0.01. The dashed line indicates the separation between mutants with zero and with one or more significant perturbations. Inset: cumulative probability. Significant mutants (HMP ≤ 0.01 cutoff; blue) have multiple significant perturbations, with 50% having four or more (inset). f, Autophagy response dynamics for the six profiles. g, Distribution of significant parameter perturbations among the significant mutants. Groups of related parameters are colour-coded based on clustering (Supplementary Fig. 3d). Number of positive or negative significant perturbations per parameter (bottom). The radial bar charts (top) indicate log₂-transformed fold enrichment of significant perturbations with P ≤ 0.01 in the respective profile over the entire genome. h, Genome-wide proportions of mutant profiles (left) and representation of DAmP collection mutants in each class (right); χ² test. i, Distribution of reference mutants across response profiles; χ² test with Bonferroni-corrected post-hoc test on residuals; *P < 2.2 × 10⁻¹⁶. NS, not significant.

Source data

To robustly quantify the autophagy responses across the genome-wide library with minimal technical confounders and observation bias, we first assembled a large training dataset of extracted image features consisting of 1.34 × 10⁶ automatically annotated WT and atg1Δ cells from 70 independent experiments. We then applied a deep-learning approach to classify autophagic activity (Fig. 1a and Extended Data Fig. 1a–k). The image features included 31 single-cell parameters summarizing the distribution and intensity of the two primary markers as well as a third channel representing their pixel-wise co-occurrence. The output labels of the training data were determined by experimental control conditions where autophagy was fully activated (WT cells for 7 h in −N) or completely inactive (either atg1Δ across all time points or WT cells for 4 h in +N) in the cell populations, thus avoiding the limitations of human annotation. Subsequently, fully connected deep neural networks (DNNs) were trained to predict the activation state per cell based on the image features.

Extended Data Fig. 1. Robust quantification of autophagy using deep learning.

a, Overview of deep learning approach. Top left: Causal control conditions for training data. Non-autophagic states (0) were defined as WT cells at T0 of starvation and after 4 h of replenishment, and as atg1Δ cells for any time points. Autophagic states (1) were defined as WT cells after 7 h of starvation until T0 replenishment. Top right: Neural nets were trained on extracted single-cell signal distribution features to classify autophagic cells. Bottom left: A hyperparameter grid search was initially performed over different neural net models with varying activation functions, hidden layers, number of initial neurons, and decay in the number of neurons per layer. The initial collection of models was trained for 150 epochs with early stoppage on validation loss and subsequently evaluated for test predictions, validation accuracy loss, and overfitting. Top models with an elu, relu, or tanh architecture were selected based on the validation loss and retrained for 3,000 epochs with early stoppage on validation loss. Bottom right: A UMAP embedding of the last latent layer containing predictive autophagic features was used to evaluate the continuity of single-cell distributions between non-autophagic and autophagic states. For fast embedding, we used a parametric UMAP (pUMAP) approach where neural nets were trained to map the high-dimensional latent space features to a two-dimensional UMAP embedding. b, The average autophagy % in test data for neural nets, trained with various hyperparameters controlling the size and shape of the architecture, ranked by validation loss, consistently shows high separation between reference autophagy states for models with elu, relu, and tanh activation functions. c, Similar training and validation accuracies indicate low overfitting across most models. d, Sufficient model complexity is required to predict autophagy confidently. Only relu-based models show a positive association between increased model complexity and overfitting. Dashed lines indicate a complexity cutoff based on the model with the lowest complexity in top 1% (validation loss). Top 5 high and low complexity models (based on validation loss) were retrained and examined further. e, Top model architectures trained for a greater number of epochs show similar performances (accuracy and loss) and robustness (correlation of predicted autophagy dynamics of WT and vam6Δ between experiments). Dashed lines indicate separation between low complexity (left) and high complexity (right) models. For tanh and elu architectures, low complexity models perform slightly better on average. f, Different model architectures result in very similar predicted autophagy dynamics. g, Top model architectures show high robustness in dynamics of embedded latent autophagic features using UMAP (correlation of population mean for WT and vam6Δ between experiments), with higher experimental correlation for models based on elu and tanh. h, UMAP dynamics of latent autophagic features are similar and consistent with reference phenotypes across different model architectures, with some activation function-dependent distributional differences in mean and variance. Points and bars indicate the mean and standard deviation at given time points across all experiments performed. The two models with the least variation in non-autophagic features are indicated by the boxes. i, Latent-space UMAP embedding of control cells (model 30 and model 22). Points and bars indicate the population mean and standard deviation of an experiment at a given time point. j, Independent models yield highly consistent autophagy predictions across the genome-wide yeast library (left panel), with a similar correlation to the reference controls colour-coded as in h (right panel). k, Association between input features and predicted autophagy log-odds.

A hyperparameter grid search across 3,096 model architectures demonstrated that multiple DNNs could predict the rates of autophagy activation with high accuracy (>98%) and reproducibility across experiments (Extended Data Fig. 1b–f). Moreover, uniform manifold approximation and projection (UMAP) embedding of latent autophagic features represented in the DNNs revealed a continuous change in WT cells and intermediary phenotypes (vam6Δ) in which cells transition between distinct autophagic states with low levels of non-autophagic noise (variance in the atg1Δ embedding; Extended Data Fig. 1g–i). Of the top 30 model architectures, the DNN with the most consistent performance across several evaluation metrics (model 30) was selected to predict autophagy in the genome-wide screen (Extended Data Fig. 1e–h). To ensure robust analysis of latent-space features and avoid systematic measurement bias due to differences in activation functions, we compared the embeddings of two distinct DNNs: model 22 and model 30 (Extended Data Fig. 1h,i). This enabled reliable profiling of nutritional responses across plates, resulting in 5,678 unique starvation responses and 5,442 replenishment responses after the removal of mutants with low-confidence predictions and low cell counts (Supplementary Fig. 1a–e).

For most phenotypes, both the activation and inactivation phases in response to nutrient shifts displayed sigmoidal kinetics, with the inactivation phase exhibiting a much steeper decline (Fig. 1b and Supplementary Fig. 1a). Using a double-sigmoidal model, we extracted 15 parameters that captured different aspects of the response kinetics (Fig. 1c and Supplementary Fig. 2a,b)²⁷. These included the activation and inactivation rates in each treatment phase measured by the slopes or width of the dynamic ranges; the sensitivity to nutrient shifts measured by the response times (t_lag, transition midpoint (t₅₀) and t_final); and the autophagy activation potentials at the start, maximum and final points of the time course. Mutant perturbations were subsequently assessed by calculating the difference in each parameter from the corresponding array medians (Fig. 1a), along with the percentage of average autophagy perturbation, computed as the differential integral under the curve for each phase separately (denoted −N and +N) or the entire time course (denoted ‘overall’; Fig. 1c and Supplementary Fig. 2c–e).

To assess the statistical significance of individual parameter perturbations, we performed multiple hypothesis tests using an expected error model (H₀) derived from replicate WT and negative control measurements in each plate (Supplementary Figs. 2c–e and 3a). The global statistical significance of each mutant was scored by combining its test results using a harmonic mean P value (HMP) weighted by the genomic H₀-rejection rate under a 5% Benjamini–Hochberg (BH) rejection threshold for each parameter (Fig. 1d and Supplementary Fig. 3b)²⁸. A 1% HMP cutoff identified 1,613 mutants with significant perturbations in multiple parameters and a relative over-representation of essential genes (DAmP mutants; Fig. 1d,e, Supplementary Fig. 3c and ref. ²⁹, Data File S2). Grouping the spectrum of parameters into five distinct classes (Supplementary Fig. 3d) revealed that mutants with changes in starvation responses, particularly in the slopes, were more numerous and varied compared with those with changes in replenishment responses, suggesting a greater regulatory complexity in the activation phase of autophagy (Fig. 1f,g and Supplementary Fig. 3a).

Clustering the responses of significant mutants revealed six distinct autophagy perturbation profiles (Fig. 1f and Supplementary Fig. 3e), which were categorized based on the type of kinetic changes that characterized each cluster (Fig. 1g)²³. For instance, three major groups (‘ultrasensitive’, ‘hyposensitive’ and ‘hyperactive’) exhibited changes in temporal sensitivity and activation rates (Fig. 1g,h). The hyposensitive mutants displayed delayed responses to nutrient shifts, whereas the ultrasensitive and hyperactive mutants displayed increased activation and inactivation rates (slopes), with the hyperactive mutants further having elevated basal autophagy (at time (t) = 0, t₀) and higher sensitivity to −N. Three minor groups (‘insufficient activation’, ‘failed response’ and ‘null response’) exhibited impaired response potential at different degrees of severity (Fig. 1g,h). The null responses displayed minimal autophagy activation across the entire time course, whereas failed responses exhibited severe deficiencies in activation. Mutants with insufficient activation responded more normally to nutrient shifts but were unable to fully activate autophagy in −N. These mutants also displayed slight elevation in basal activity. Analysis of the composition of each cluster revealed that essential genes tended to be disproportionately represented among the hyperactive and insufficient activation profiles (Fig. 1h), reflecting either dynamic differences between the DAmP and knockout (KO) libraries (Fig. 1b) or suggesting a potential relationship between basal autophagy, starvation response potential and growth control (Supplementary Fig. 3a).

To inspect the profile distribution of well-characterized autophagy genes, we defined two ‘gold-standard’ reference sets of autophagy-related genes through literature curation and manual inspection of images (Extended Data Fig. 2a, Methods and Supplementary Table 1). A set of ATG genes essential for the induction of autophagy (ATG set, n = 16) was significantly over-represented among null responses and a set comprising genes involved in the maturation or clearance of autophagosomes (Fusion set, n = 19) was significantly over-represented among failed responses (Fig. 1i). These distributions indicate that the different response profiles meaningfully reflect the severity of known autophagy deficiencies.

Extended Data Fig. 2. Tests for detecting autophagy reference genes.

a, Model-fitted autophagy response curves for reference mutants. Repeated starvation measurements from a repetition screen are indicated in red for some genes. b, Area under the curve for PRC curves over ranked p-values for kinetic perturbation parameters comparing autophagy reference test sets versus equisized randomly sampled negative sets or ORF-negative controls (squares). The box plots indicate the test sets distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. c, PRC curves for perturbation parameters and HMP ranked by p-values comparing a combined autophagy reference test set (n = 35) versus equisized randomly sampled negative sets. The dashed lines represent the p-values and HMP for the recalled genes. d, PRC curves for perturbation parameters ranked by p-values comparing a combined autophagy reference test set (ATG: n = 16; Fusion: n = 19) versus an ORF-negative control set (n = 23). e, ROC curves for perturbation parameters ranked by p-values comparing a combined autophagy reference test set (ATG: n = 16; Fusion: n = 19) versus an ORF-negative control set (n = 23). f, ROC curve for perturbation parameters ranked by their signed p-values comparing ATG genes reference test set (n = 16) and fusion genes reference test set (n = 19) versus equisized randomly sampled negative sets. g, Area under the curve for ROC curves over ranked p-values (top panel) and over ranked signed p-values (bottom panel) for kinetic perturbation parameters comparing autophagy reference test sets versus equisized randomly sampled negative sets or ORF-negative controls (squares). The box plots indicate the reference tests, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers.

Accuracy and robustness of autophagy perturbation profiling

To evaluate the capacity of the perturbation statistics to identify true autophagy-associated genes, we analysed the precision-recall curves for the two autophagy reference sets, comparing them with a negative control set consisting of deletions of dubious open reading frames (ORFs) that did not overlap with verified genes (n = 23) or with random samples from the genome-wide distribution. For the combined ATG and Fusion reference sets, the HMP yielded the best precision-recall on average, with an area under the curve (PRC-AUC) > 0.97 (Extended Data Fig. 2b,c). The individual kinetic-parameter P values yielded PRC-AUCs that were in the 0.70–0.95 range when compared with genome-wide random samples and in the 0.8–1.0 range when compared with the negative control set (Extended Data Fig. 2b–f). Interestingly, although the ATG and Fusion mutants exhibited severe autophagic deficiencies in starvation, many Fusion mutants with delays in autophagosome clearance displayed slow deactivation in response to nitrogen replenishment (t₅₀ in +N; Extended Data Fig. 2a,f,g).

We also evaluated the reproducibility of the measured perturbation phenotypes. Repetition of the starvation protocol for the top mutants (selected based on the HMP; Extended Data Fig. 3) revealed that the dynamics of stronger autophagy perturbation phenotypes were more likely to replicate across experiments (Extended Data Fig. 3g). In addition, by assessing the representation of autophagy phenotypes among interacting genes in various yeast network databases based on physical and genetic interactions (Methods), we observed a genome-wide concordance between a gene’s autophagy perturbation statistic and enrichment of phenotypes among its nearest neighbours (Extended Data Fig. 3b–e)^30–32. Discordant outliers within this data, along with ‘autophagy’ gene ontology (GO)-annotated mutants yielding weak perturbations, were used to predict potential false positives (FPs) and negatives (FNs) that were subsequently retested (Extended Data Fig. 3d–g). This analysis identified only 56 potential FPs and 109 potential FNs. More than half of the predicted FPs produced replicable autophagy phenotypes, representing potentially novel autophagy-related genes, whereas only 27 mutants represented outliers that regressed in perturbation magnitude, resulting in weak or no correlation between experiments (Extended Data Fig. 3f,g). Among the predicted FNs, 42 mutants were significant under a more liberal threshold (HMP ≤ 0.05) and displayed a discrete but replicable autophagy perturbation phenotype (Extended Data Fig. 3h). These mutants were also supported by stronger network evidence (Extended Data Fig. 3i,j) compared with non-significant mutants (HMP > 0.05), which were more likely to have been selected based on GO annotations (Extended Data Fig. 3k). Further inspection revealed that many of these non-significant mutants represented genes with specialized roles, such as genes that are not essential for starvation-induced autophagy, including those involved in pexophagy (ATG36), the Cvt pathway (ATG20, ATG34 and ATG19) or mitophagy (ATG32 and ATG33) as well as genes with weakly penetrant but replicable phenotypes such as VPS15, FAR11, PEP12 and NPR3 (Extended Data Fig. 3l).

Extended Data Fig. 3. Reproducibility of top hits, and potential errors.

a, The first stage of reproducibility testing was done by selecting the top-ranked mutants based on the HMP, along with predicted false positives and negatives, and repeating the starvation protocol (Methods 1.3). Subsequently, the mutant-wise correlation in autophagy dynamics and perturbation dynamics was compared between the two screens. b, Prediction of false positives and negatives based on enrichment tests of significant phenotypes (generated by different HMP cutoffs) in the set of first-degree neighbours for different network sources. The enrichment test p-values were aggregated by taking their harmonic mean to an overall network enrichment p-value and compared with the autophagy perturbation statistics of the respective genes. c, Non-significant genes annotated with the indicated autophagy GO terms were predicted as false negatives. d, Network enrichment p-value correlates with autophagy perturbation FDR_min (minimum perturbation BH-FDR value). Potential false positives or negatives were predicted in the off-diagonal areas as indicated. R indicates the Pearson correlation, and p indicates their respective p-values. p < 2.2e-16. e, Network enrichment p-value is concordant with both autophagy perturbation FDR_min and HMP. f, Distribution of mutant replicate correlation of autophagy (top left) and autophagy perturbation (top right) for top hits and predicted false positives and negatives. Bottom right panel shows the replicate perturbation covariance. The box plots indicate the mutant replicate correlations, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR. g, Mutant replicate correlation of autophagy perturbations for top hits and predicted false positives and negatives versus average absolute autophagy perturbation for GW data (top panel) and validation data (bottom panel). The vertical lines represent the 2× median perturbation of the screen. R indicates the Spearman correlations, and p indicates their respective p-values, as shown in the figure. h, Stratifying the predicted false negatives based on less stringent HMP cutoffs shows that weakly significant screen mutants display weak but reproducible responses, in contrast to non-significant mutants. The box plots indicate mutant replicate correlations, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR. i,j, Weakly significant and reproducible have higher counts of nearest neighbours with autophagy phenotypes (i) that are more likely to be significantly enriched (adjusted p-value ≤ 0.01) within their most enriched network (j). k, Predicted false negatives that are non-significant under the HMP were disproportionately predicted based on GO annotation. l, Perturbation covariance of non-significant mutants with autophagy-related GO annotation.

Finally, to test the reliability of the genome-wide analysis, 33 strongly significant (HMP ≤ 0.001) deletion mutants representing different dynamic profiles and biological functions were reconstructed in triplicate and subjected to a set of validation experiments with variations in the growth protocol (Extended Data Fig. 4a,b and Methods). The reproducibility of phenotypes across independent clones and different protocols was similar to that of the same clones across all experiments, with a minor strain background effect and elevated noise in autophagy response caused by growing the cultures to the stationary phase (Extended Data Fig. 4c–g). Moreover, the time point-wise autophagy perturbations (Supplementary Fig. 4a–d) and response kinetic parameters (Supplementary Fig. 4e,f) showed strong correlations across mutants between the screens, with metrics for activation levels (Spearman correlation > 0.8), response times (Spearman correlation > 0.7) and slopes (Spearman correlation > 0.75) in −N being the most robust (Supplementary Fig. 4f). These observations indicate that the genome-wide variability in mutant-response kinetics is highly reproducible.

Extended Data Fig. 4. Reproducibility across experimental replicates, strain background, and experimental protocols.

a, The second stage of reproducibility testing was performed by selecting a sub-library of strongly significant (HMP ≤ 0.001) verified gene deletions with diverse perturbation phenotypes (dynamic variability and covering multiple gene classes for biological variability). For each mutant, three new clones were made in a new background strain, and the entire experimental protocol was repeated three times under two distinct growth preparation protocols (Methods 7.1 and 7.3). Reproducibility was assessed in a mutant-wise manner by comparing autophagy dynamics and perturbation dynamics, and in a screen-wise manner by comparing computed kinetic parameters. b, Workflow of experimental organization for comparing reproducibility. P1 refers to a growth preparation protocol where the cells were grown to log phase before being transferred to fresh media 8 hours before the experimental procedure. P2 refers to a growth preparation protocol where the cells were grown to stationary phase before being transferred to fresh media 8 h before the experimental procedure. The genome-wide screen R1 and replicate screen R2 were subjected to P1, and replicate screens R3 and R4 were subjected to P2. Coloured arrows indicate the different levels of replication tested. c, Comparison of autophagy perturbation and mutant-wise replication errors with the same colour coding as in b. The horizontal dashed line represents the 2× median perturbation of the screen. The box plots indicate mutant perturbations or replication errors, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. d, Mutant-wise replicate correlation of autophagy (left panels) and autophagy perturbation (right panels) for collection clones and new clones across biological replicates. e, Distribution of autophagy (top panels) and autophagy perturbation (bottom panels) correlation between collection clones and new clones for each experimental replicate. The distributions are stratified based on the ranked clone correlation from best to worst for each gene. The dashed horizontal lines represent the best, median, and worst correlation between WT control strains of the collection and new mutants. The median of all WT replicates was used as a plate control for computing autophagy perturbations, and a negative correlation for R2 indicates a distinct WT strain effect. f, Autophagy response dynamics for collection (Y7092) and new (BY4741) WT control strains. Dashed lines indicate the common median across all strains and experimental replicates for reference. g, Autophagy perturbation correlation between collection clones and new clones versus average absolute autophagy perturbation for new clones and collection clones across protocols. The red and black dots represent the best and median clone correlation for each gene, respectively. The dashed horizontal lines represent the best and median correlation between WT control strains of the collection and new mutants.

Cellular processes that impact autophagy dynamics

To identify pathways affecting autophagy dynamics, we first performed gene-set enrichment analysis (GSEA) on each of the individual response statistics (parameter-signed −log₁₀-transformed P values) for sets defined by GO biological processes (GO-BP). Significantly enriched terms (at least one enrichment P < 0.005) were analysed using principal component analysis (PCA) on the matrix of enrichment statistics (signed −log₁₀(P value)) to identify which kinetic parameters captured the most functional variation. This approach allowed us to examine the GO-term enrichment in a non-redundant manner by summarizing the functional variation along the primary principal components. Investigations of the contributions of each PCA loading (representing a kinetic parameter) revealed that the response potential and sensitivity to starvation aligned with the dominant direction of enriched GO-BP (Fig. 2a), with terms related to autophagy and vacuolar activity yielding the strongest associations (Fig. 2b). In contrast, sensitivity to nitrogen replenishment aligned more with the second principal component (Fig. 2a) and was enriched for processes related to membrane trafficking and fusion (Fig. 2b). Interestingly, these GO terms were highly decorrelated from the direction capturing mechanisms controlling the response potential (Fig. 2b). Finally, perturbations in initial and final autophagy levels, which were the least correlated with the other parameters (Supplementary Fig. 3d and Fig. 2a), were enriched for terms related to amino-acid and nucleoside metabolism, nitrogen utilization and mitochondrial activity (Supplementary Fig. 5a).

Fig. 2. Kinetic perturbation profiling is highly sensitive and potentiates deeper annotation of regulatory modules.

a, Principal component analysis loadings for kinetic parameters using a matrix of signed −log₁₀-transformed P values from GSEA of GO-BP. Parameter classes are colour coded as in Fig. 1f. b, Principal component analysis of GO-BP terms from a. Colours indicate the perturbation parameter class (Fig. 1g) with the most significant enrichment; GSEA-based permutation test (two-sided); PC, principal component. c, Enrichment map based on Jaccard similarity of enriched GO terms from GSEA over the HMP. Edges are drawn for terms with 10% overlap. The pie charts represent the composition of perturbation profiles among enriched genes (leading edge). P values were calculated using a GSEA-based permutation test (two-sided). MF, molecular function; CC, cellular compartment; pol., polymerase. d, Similarity in functional distribution of perturbation profiles across enriched GO terms, assessed by Spearman correlation of GSEA-leading edge counts from c.

We then identified enriched GO terms by GSEA using the combined HMP statistic (Fig. 2c). An enrichment map provided an integrated overview of the primary functional categories influencing autophagy with three large interconnected clusters representing autophagy and membrane trafficking, RNA metabolism and translation, and gene-expression regulation. This map highlighted a graded distribution of perturbation profiles with homogenous enrichment of severe autophagy deficiencies (null response and failed response) close to autophagy-essential GO clusters and less-severe deficiencies in neighbouring clusters (Fig. 2c). Quantification of the enrichment similarity between the different profiles revealed a gradual functional overlap between mutants causing partial loss-of-function phenotypes such as failed response, hyposensitive and insufficient activation (Fig. 2d). Interestingly, the gain-of-function profiles ultrasensitive and hyperactive co-occurred more frequently with each other as well as with the hyposensitive and insufficient activation profiles, respectively, indicating consistency in regulatory phenotype pairing along the dimensions of sensitivity and response potential (Fig. 2d).

Causal structure and network architecture of autophagy regulators

A phenotypic change in response to a genetic intervention does not necessarily imply that a gene is a direct regulator of the measured phenotype. We therefore examined the organization of autophagy-influencing genes and their relation to the autophagy core machinery in functional networks more closely. Previous work in yeast has shown that genome-wide networks can accurately capture the functional organization of the cell when constructed using similarities in genetic interactions, which are calculated using the Pearson correlation coefficients of interactions from synthetic genetic array data (SGA-PCC), or when constructed using protein–protein interactions (PPIs)^30,33,34. Using spatial analysis of functional enrichment (SAFE) for the different perturbation profiles, we found distinct genome-wide enrichment patterns for all the profiles except ultrasensitive in both genetic similarity and Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) PPI networks (Fig. 3a and Extended Data Fig. 5a)³⁵. It is noteworthy that the ATG set did not form a distinct cluster in the genetic similarity network but did so in the PPI network, where its spatial enrichment overlapped with that of null responses, and it co-localized with processes related to membrane fusion (Extended Data Fig. 5a).

Fig. 3. Genome-wide regulatory network architecture of autophagy-influencing genes.

a, SAFE analysis of the global genetic similarity network³⁰. Reference map of bioprocesses (left) as well as ATG and fusion sets (middle) to interpret the spatial enrichment of autophagy perturbation profiles (right). *ATG core enrichment. MVB, multivesicular body sorting. b, Shortest paths from the ATG core machinery. c, Average HMP per shortest path distances in different functional interaction networks, computed as indicated in b. The dashed line indicates HMP = 0.05. Data are the mean and 95% confidence interval, computed from n genes from Extended Data Fig. 5d, indicating the cumulative gene count per path distance. d, Shortest causal path length per autophagy perturbation profile relative to the non-significant genes (HMP > 0.01). The dashed line indicates the reference baseline (zero relative shortest causal path). Data are the mean and 95% confidence interval, computed from n genes per perturbation profile with the distribution for the Costanzo³⁰ network indicated in Extended Data Fig. 5b,f, indicating the absolute path distances per profile. e, Correlations in temporal autophagy perturbations or sets of perturbation parameters (signed −log₁₀-transformed P values). Correlations were considered for different HMP cutoffs (1.000, 0.010 or 0.001; set as significance thresholds in the screening). The mean, median and interquartile range (IQR; 25–75th percentile; q25, q75) were computed from n gene interactions as indicated in Supplementary Fig. 6b. f, Enrichment of positive and negative genetic interaction similarities (SGA-PCCs) per correlation interval of autophagy perturbations between pairs of strongly significant genes (HMP ≤ 0.001). Two-sided Fisher’s exact test; exact P values are provided in ref. ²⁹, Data File S4. g, Direction preferential enrichment (computed as the difference between the maximum and minimum of the GSEA normalized running-sum statistics, normalized enrichment score (NES); Supplementary Fig. 6c and Methods) of perturbation profiles along the SGA-PCC spectra for individual genes of the core autophagy machinery. GSEA-based permutation test (two-sided); exact P values are provided in ref. ²⁹, Data File S4. Screen phenotypes were manually categorized according to the expected or degree of perturbation. h, Illustration depicting hypothetical regulatory entry points within the modules of the autophagy core machinery, derived from the significant enrichment of different perturbation profiles along SGA-PCC spectra in g. NA, not available. *P ≤ 0.05.

Extended Data Fig. 5. Distribution of autophagy-regulating genes in genome-wide networks.

a, SAFE analysis of the genome-wide STRING network (version 10), with a combined score cutoff of 990 for edges. Reference map of bioprocesses (left) and of ATG and fusion set (middle) for interpreting the spatial enrichment of autophagy perturbation profiles (right). b, Shortest pathlength and shortest causal pathlength per autophagy perturbation profile. Causal paths were determined by requiring HMP ≤ 0.01 for all intermediate nodes. c, Shortest path length per autophagy perturbation profiles relative to the non-significant genes (ns). d, Cumulative count of significant genes (HMP ≤ 0.01) per shortest path or shortest causal path distances. e, Cumulative enrichment (log₂) of significant genes (HMP ≤ 0.01) per shortest path or shortest causal path distances. f, Enrichment of perturbation profiles per shortest causal path distances in the SGA-PCC network³⁰. Fisher’s exact test (two-sided); exact p-values as shown in the figures and Data File S4 (ref. ²⁹) (see distance per profile) g–i, Distribution of autophagy perturbations across gene modules spanning three regulatory network hierarchies. Node colours indicate the overall perturbation statistics (signed −log₁₀ p-values) and the size of the nodes is proportional to their HMP (−log₁₀). j, Hierarchy of autophagy regulatory network layers presented in g–i.

Interestingly, more severe autophagy phenotypes were enriched closer to the ATG core subgraph (ATG set, including ATG1 and ATG8). The proximity of autophagy-influencing genes to the ATG core in these networks was confirmed in two ways when analysing the shortest path lengths (Fig. 3b). First, the HMP was more significant for genes that were closer to the ATG core than for those that were further away (Fig. 3c). Second, the average shortest path length for each autophagy perturbation profile, except ultrasensitive, was significantly closer to the ATG core than the average shortest path length for genes that were not in any perturbation profile (HMP > 0.01; Extended Data Fig. 5b–f), suggesting that these phenotypes on average represent more direct influences on autophagy. This relationship became even stronger when path lengths were calculated using the shortest causal path (requiring all intermediate nodes to have a significant influence on autophagy in response to a gene deletion, HMP ≤ 0.01; Fig. 3b,d). In contrast, the 11% of mutations in the ultrasensitive profile, which only affect response sensitivity, might involve more distant fine tuning of the nutrient response, spanning through several intermediates as indicated by the observed path length.

Consistent with the clustering of autophagy-influencing genes in these networks, gene pairs with reported functional relationships (SGA-PCC, BioGrid PPIs and STRING)^30–32 or complexome associations³⁶ exhibited correlated autophagy perturbations (Fig. 3e and Supplementary Fig. 6a,b), suggesting that the dynamic profiles were consistent with functional gene relationships on a genome-wide scale. The signs of the pairwise SGA-PCC scores can distinguish between inhibitory and activating gene relationships^37–39. We examined whether these relationships would also be consistent with the signs of autophagy correlations by assessing whether gene sets with different SGA-PCC cutoffs were enriched in the perturbation dynamics correlations using Fisher’s exact test. Divergences in correlations between strongly significant genes indeed caused a shift in the representation of positive and negative SGA-PCC (Fig. 3f). Moreover, when testing for directional parametric enrichment of mutant profiles over the SGA-PCC of core autophagy genes by GSEA, hyperactive phenotypes (in contrast to loss-of-function phenotypes) were preferentially associated with negative SGA-PCCs for both ATG and fusion core genes, suggesting an inhibitory relationship (Supplementary Fig. 6c,d). Using this approach, we inferred differential regulation of the various subsystems executing autophagy, which cannot be measured from their null or deficient individual deletion phenotypes (Fig. 3g and Supplementary Fig. 6e). Here ultrasensitive responses were negatively related to genes involved in the early stage of autophagosome formation, particularly ATG1, suggesting that this may be the main limiting factor for the starvation response rate. Interestingly, failed response, which was positively related to the fusion core genes, was negatively related to some genes involved in autophagy induction (Fig. 3h). This could indicate the existence of negative-feedback circuits from stages involved in autophagosome clearance⁴⁰ or pleiotropic functions of gene products involved in amino-acid liberation and TORC1 reactivation (such as Vps41 and Vam6)^24,41.

To explore the regulatory relationships between specific autophagy-influencing genes, we mapped the autophagy perturbation phenotypes across curated complexome networks representing the three major regulatory cell layers: gene-expression regulation, RNA metabolism and translation, and membrane dynamics and trafficking, which exhibited systems-wide influences on autophagy dynamics (Extended Data Fig. 5g–j). This allowed us to observe the penetrance and distribution of phenotypes across well-characterized functional modules. For example, we observed clustering of multiple significant phenotypes, such as in the chromatin remodelling submodule, and consistency in the direction of autophagy perturbation for regulatory subgraphs (for example, Ure2 and Gln3, Gal80 and Gal3, and the CURI complex)^42–44. For RNA metabolism and translation in particular, we detected a systemic enrichment of stronger phenotypes along modules related to mRNA processing and translation decoding (for example, binding of mRNA by the 40S small ribosomal subunit), where perturbations in mRNA maturation and binding resulted in a reduced autophagic response, whereas perturbations in mRNA decay caused the opposite phenotype.

Dynamic latent-space analysis of autophagosome formation and clearance

We reasoned that the latent autophagic features generated by the DNN (Fig. 1a(iii)) potentially provided more information about cellular states that could be captured by the binary classifier used to determine the autophagy response (Extended Data Fig. 1i). We therefore projected these latent autophagic features onto two dimensions using UMAP (Fig. 4a,b and Extended Data Fig. 6a,b), where information about autophagy was conserved together with cell-to-cell variation and dynamic data from the DNN latent space (Extended Data Fig. 6c–f). Wild-type cells progressed from a UMAP region characterized by an absence of autophagosomes (−N, 0 h), through a region characterized by free autophagosomes (−N, early time points, 3–5 h) and eventually to a region characterized by cleared autophagosomes, where autophagosomes fused with the vacuole (−N, late time points, 5–11 h; Fig. 4a,b and Extended Data Fig. 6a,b). Following nitrogen replenishment, the WT cell trajectory reversed. As expected, the atg1Δ mutant cells’ trajectory never progressed out of the ‘no autophagosome’ UMAP region. In contrast, the vam6Δ mutant cells’ trajectory overlapped with the early WT cell trajectory but never advanced to the ‘cleared autophagosome’ UMAP region. It instead exhibited a distinct UMAP flux characterized by the progressive accumulation of free autophagosomes (Fig. 4a,b), which was statistically separable from that of WT cells in the UMAP within similar intervals of low-confidence DNN predictions (classification probabilities in the range 0.1 < P₁ < 0.9, where P₁ denotes the model-assigned probability of the autophagy class; Extended Data Fig. 6g,h). Given the substantial genome-wide variation in UMAP dynamics and many mutants exhibiting elevated classification uncertainty (Extended Data Fig. 6e,f), we tested the use of the vam6Δ latent space as an ‘out-of-distribution’ intermediate reference state in a two-step model framework for autophagy execution (Fig. 4c and Extended Data Fig. 7a–e). Here Bayes factors (BFs) were used to score the UMAP latent-space evidence of mutant cells executing autophagosome formation (VAM6:ATG1) or clearance (WT:VAM6) based on the position of cells relative to reference kernel densities computed for WT, atg1Δ and vam6Δ cells (Extended Data Fig. 7b,c; details in Methods). We employed a time-independent BF, computed using time-varying kernel densities, to classify the overall mutant behaviour between the three reference phenotypes (Extended Data Fig. 7b,d). This procedure effectively integrated out the time variable by averaging the BF scores per mutant, enhancing the classification accuracy and sensitivity for phenotype detection (Extended Data Fig. 6f–h). In addition, we scored the time-wise evidence for the state of the mutants as they moved between fixed reference kernel densities throughout the time course, indicating changes in autophagy execution activities (BF_t; Extended Data Fig. 7c,e and ref. ²⁹, Data File S7).

Fig. 4. Dynamic latent-space analysis of autophagy execution.

a, Latent-space model for autophagy execution. The distribution of atg1Δ overlaps with the region for no autophagosomes and the distribution of vam6Δ overlaps with the region for free autophagosomes. The arrow represents the dynamic of WT cells as they execute autophagy. b, Comparison of UMAP-embedded latent-space distributions for WT (ORF) controls (left) as well as vam6Δ (middle) and atg1Δ (right) cells with representative micrographs (top and bottom) for the indicated time points. The densities represent distributions from six independent plates, grouped according to time point and coloured according to the average autophagy prediction. Arrows indicate the position and direction of the average latent-space flux per plate. c, Bayes factor classification of autophagy execution phenotypes across the genome. Densities represent WT (blue), vam6Δ (orange) and atg1Δ (grey) cells. Reference ATG and fusion mutants are indicated and labelled. d, Colour-coded functional overview of fusion mutants in each complex, ranging from red to orange, based on severity as determined by BF WT:VAM6. e, UMAP-embedded latent-space distributions and autophagy response dynamics for fusion mutants with representative micrographs taken at the indicated time points. f, Correlations in genome-wide BF_t variations for different time lags (indicated by black arrowheads from BF_t to BF_t+l). BF_t+l represents the lagged BFs at time t+l, which was correlated with the BFs at time t, where l indicates different time lags. g, UMAP-embedded latent-space distributions and autophagy response dynamics for rtg1Δ, rtg2Δ and rtg3Δ (top) with representative micrographs from the indicated time points (bottom). White and grey arrowheads indicate the population-mean velocity vectors at each time point, with grey representing the plate control and black representing the mutant. h, Correlations between the average BF_t and autophagic flux measured as the average GFP–Atg8 cleavage from 0, 2 and 4 h in −N. The PCC (R) and exact P values are provided. i, Relative contribution of log₂(BF_t) scores in predicting autophagic flux per time point, estimated using mixed-model regression (with time as the grouping factor) based on a univariate (WT:ATG1) or bivariate (VAM6:ATG1 + WT:VAM6) model. b,e,g, White dashed circles mark the cell contour and serve as visual guides for identifying autophagy features. Scale bars, 3 µm (×60; 1 pixel represents 0.115 µm). h,i, Data are the mean ± s.e.m. (number of replicates per mutants and conditions are in ref. ²⁹, Data File S12).

Extended Data Fig. 6. Autophagy dynamics and noise in the latent space.

a, UMAP-embedded latent-space distributions for control cells from model 22 and model 30. The densities represent distributions from 6 independent plates, grouped per time point and coloured with the average autophagy prediction. Arrows indicate the position and direction of the average latent-space flux per plate. Left panel is presented with representative micrographs in Fig. 4b. b, Latent-space UMAP flux for control cells from model 22 and model 30. Arrows indicate the position and direction of the average latent-space flux per plate, and colour indicates the time for each population average. Note the slightly greater variability of non-autophagic features (atg1Δ) in model 30. c, Training and test scores for UMAP predictions, showing that predictive information from the DNN models is fully conserved. The DNN outputs, autophagy probability (P_a) or log-odds (LO), or the sample time per cell were predicted from bagged regression trees and scored using the Pearson correlation coefficient (PCC). Classifications of WT versus vam6Δ cells were performed with bagged classification trees and scored using an F1 score. The box plots indicate test prediction scores, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. d, Comparing multivariate properties of the UMAP embeddings and latent feature vectors from the DNN models of the control cells. Each point represents standard deviation (SD) and flux per cell population for a given experiment and time point. R indicates the Pearson correlations, and p indicates their respective p-values, as shown in the figure. e, Comparing phase-averaged UMAP deviation and flux, as well as prediction uncertainty (binary entropy) across models and datasets. The box plots indicate mutant distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. f, Comparing UMAP variance and flux between the two DNN models. Note that UMAP flux is more reproducible between the two models. R indicates the Pearson correlations, and p indicates their respective p-values, as shown in the figure. g, Distribution of WT and vam6Δ for different DNN prediction intervals. h, Training and test F1 scores for UMAP classification of WT versus vam6Δ cells per DNN prediction intervals. The classifications were performed with bagged classification trees.

Extended Data Fig. 7. Bayesian modelling of autophagy execution from out-of-distribution latent-space features.

a, Causal control conditions for learning intermediate states of autophagy execution. WT cells can execute autophagy based on nitrogen status. atg1Δ cells cannot produce autophagosomes. vam6Δ cannot clear autophagosomes and has de-regulated inhibition of autophagosome formation due to diminished TORC1 activation. The consequence is empirical latent-space distributions of vam6Δ cells that are populating an area between the non-autophagic and mature autophagic feature space of atg1Δ and WT cells. The latent-space distributions for the three mutants can be used as reference states in a Bayesian modelling framework using Bayes factors (BFs). b, BF scores for classifying mutant phenotypes where time is a marginalized variable by using time-wise reference probability densities. c, BF_t’s for quantifying latent-space movements between fixed reference states by using time-invariant probability densities. d, Time-independent log BFs from b for the reference mutants from 6 independent experiments. The box plots indicate reference mutant distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR. e, Time-wise log BF_t’s from c for the reference mutants from 6 independent experiments, quantifying the dynamics for autophagy execution. f, Comparison of log BF and BF_t between model 22 and model 30. R indicates the Pearson correlations, and p indicates their respective p-values, as shown in the figure. g, ROC-AUC scoring the separation of ATG and fusion set mutants versus equisized randomly sampled negative sets, versus ORF-negative controls, or the test sets versus each other. The BF and BF_t scores are compared with the predicted average autophagy perturbation and average binarized autophagy perturbation from the same neural classifiers. The box plots indicate test scores distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. h, log BF and BF_t z-scores for control cells and ATG and fusion set cells. Points and bars indicate the means and standard deviations from the collections of mutants. i, log BF z-scores per autophagy perturbation profiles. The box plots indicate mutant distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. j, Comparison of log BF for KO and DAmP collection. The box plots indicate mutant distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers.

The BFs accurately scored selective defects in autophagy execution, while also quantifying gradual increases in the severity of the mutant phenotypes (Fig. 4c,d, Extended Data Fig. 7f–j and Supplementary Fig. 7a). For example, proteins involved in vacuole tethering and the homotypic fusion and vacuole protein sorting (HOPS) complex exhibited a strong clearance defect and a high autophagosome burden, whereas retromer subunits and several other membrane trafficking proteins showed more moderate defects, with autophagosomes slowly docking and gradually resolving (Fig. 4c–e and Supplementary Fig. 7b). Analysis of correlations in genome-wide BF_t distributions between time points revealed highly nitrogen-dependent regulation of autophagosome formation, with a rapid decline in correlations between time lags (Fig. 4f). This indicated that genetic contributions to autophagosome formation varied across distinct time intervals and treatment phases (Fig. 4f(top left)). In contrast, genetic perturbations in clearance activity remained relatively constant across time (Fig. 4f(bottom right)), with clearance in −N negatively correlating with the autophagosome burden in +N (Fig. 4f(top right)). These differences underscore the dynamic sensitivity of autophagosome formation to nitrogen availability, while revealing a steady and robust regulation of clearance activity through different mechanisms.

Parametric GSEA revealed that perturbations in amino-acid synthesis and TORC1-mediated nutrient sensing were distinctly associated with elevated autophagosome formation, whereas disruptions in proteasomal activity and glucose uptake increased autophagosome clearance (Supplementary Fig. 7b). By grouping time-independent BFs over each treatment phase, we uncovered nutrient sensing functions of some of these gene sets, along with additional factors regulating autophagy execution in a nitrogen-dependent manner (Supplementary Fig. 7c,d). Notably, nucleosome disassembly and mitochondria–nucleus signalling (also known as retrograde or RTG pathway) were associated with a relative increase in autophagosome formation under nutrient-rich conditions (+N), whereas metabolism-related processes, including glucose-6-phosphatase activity and lipid homeostasis, contributed more prominently to autophagosome formation under starvation (–N). Interestingly, inspection of the UMAP distributions and fluorescent micrographs of the RTG deletion mutants confirmed a hyperactive phenotype with elevated autophagosome formation in +N and a strongly delayed clearance phenotype following replenishment (Fig. 4g). Although TORC1 signalling is known to play a central role in nutrient-dependent regulation of induction, these findings indicate that additional components are required to modulate autophagy dynamics at different stages throughout the process.

To assess how reliably our dynamic latent-space analysis captured true variation in autophagic activity, we compared the different BFs with an independent readout for autophagic flux, GFP–Atg8 cleavage, measured across a set of selected mutants (Extended Data Fig. 8), including known autophagy regulators such as TORC1-signalling components and enzymes in the spermidine biosynthesis (SPE) pathway. Compared with all the different screen measurements (including the kinetic parameters), the time-averaged BF_t WT:ATG1 yielded the best correlation (95%) with average cleavage percentage, providing support for its use as a proxy for autophagic activity (Fig. 4h and Supplementary Fig. 8a,b). Given that this measure is the sum of the BF_t metrics with VAM6 as an intermediate step, we next quantified the progressive contribution of formation (VAM6:ATG1) and clearance (WT:VAM6) activity on autophagic flux over time using a bivariate regression model (Supplementary Fig. 8c,d). This analysis highlights the importance of clearance regulation, which emerges as the dominant factor for predicting variation in Atg8 flux deep into starvation conditions and immediately following replenishment (Fig. 4i).

Extended Data Fig. 8. Experimental validation of autophagic flux perturbations.

a, UMAP-embedded latent-space distributions and autophagy response dynamics for gene deletions of the spermidine-synthesis pathway with representative micrographs from indicated time points. Scale bar, 3 µm (×60, 1 px = 0.115 µm). b, UMAP-embedded latent-space distributions and autophagy response dynamics for gene deletions of selected gene deletions related to TORC1 activation. c, Representative images of immunoblot analysis of GFP–Atg8 cleavage and Rps6 phosphorylation for WT (n = 7), spe1Δ (n = 3), spe2Δ (n = 3), spe4Δ (n = 3), tco89Δ (n = 2), gtr1Δ (n = 2), and lst4Δ (n = 2) in −N followed by +N for the indicated time points. ꞵ-Actin is used as a loading control. See Methods for replicas and quantifications.

Source data

The RTG pathway suppresses autophagosome formation independently of TORC1

As deletion of RTG genes led to excessive autophagosome formation and delayed clearance in nitrogen-rich conditions (+N), resembling the phenotype of some TORC1 pathway mutants (Figs. 4g and 5a and Extended Data Fig. 8), we speculated that the two pathways were functionally convergent. In a comparison of the treatment-interaction effects of RTG deletion mutants with established autophagy regulators—including deletion mutants of the SPE pathway (a TORC1-independent positive regulator), AMPK and WHI2, PSR1 and PSR2 (TORC1 suppressors), PIB2 (a TORC1 activator) and components of TORC1 itself—the RTG mutants displayed elevated autophagosome formation (BF VAM6:ATG1) similar to core TORC1 components and activators and opposite of TORC1 suppressors (Fig. 5b)^45–48. However, the RTG mutants had a more severe effect on clearance (BF WT:VAM6) and maintained their phenotype in both nutrient conditions, indicating an insensitivity to nitrogen replenishment (Fig. 5b(right)). In comparison, the autophagy-deficient phenotype of SPE mutants was highly starvation-specific.

Fig. 5. The RTG pathway suppresses autophagosome formation independently of TORC1.

a, Comparison of the average BF z-scores for GO terms per treatment phase, starvation and replenishment that had a significant (P ≤ 0.05; two-sided permutation test) differential enrichment between the phases. The diagonal dashed line is the identity line, indicating equal z-scores across conditions. The mitochondrial–nucleus signaling (the RTG pathway) and TORC1 signaling resemble similar autophagy phenotypes and are highlighted in red. b, BF z-scores for different components of the TORC1 signalling, RTG and spermidine-synthesis pathways. The BFs were computed for each phase (−N and +N; left) and as a differential between phases after standardizing the metrics (starvation versus replenishment; right). The dashed horizontal line indicates a z-score of 0 (the mean). The solid grey lines indicate z-scores of +2 and −2, corresponding to two standard deviations from the mean. RTG mutants are highlighted in red. c, Average autophagy (left) and BF_t (right) levels in response to starvation, rapamycin or both. The points and associated bars indicate the mean and s.d., respectively, for the plate controls (n = 27 WT, and 9 atg1Δ and vam6Δ). The box plots indicate the mutant distributions, showing the median (line), IQR (25–75th percentiles; box) and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. rap, rapamycin. d, Comparison of the BF responses to rapamycin treatment dependent on the nitrogen status. A linear regression line (dashed line within the ribbon) with the 95% confidence interval (grey ribbon) is shown to identify outliers with strong nutrient–gene interactions. The dashed line outside the ribbon denotes the identity line used to compare BF responses to rapamycin treatment across nitrogen conditions. The PCC (R) and exact P values are provided. Coloured points: black, WT controls; red, RTG mutants; blue, WT:ATG1; grey, VAM6:ATG1; orange, WT:VAM6. e, Pho8Δ60 activity as a measure of autophagic flux in WT (n = 5), atg1Δ (n = 5), rtg1Δ (n = 2), rtg2Δ (n = 3) and rtg3Δ (n = 2) cells in +N (before starvation) and followed by −N for 4 h. Individual data points are shown; the bars indicate the mean ± s.e.m. Analysis of variance with Tukey’s honestly significant difference test; exact P values are provided in ref. ²⁹, Data File S12. DKO, double knockout. f, Immunoblot analysis of GFP–Atg8 cleavage and RpS6-p for rtg1Δ, rtg2Δ, rtg3Δ and WT cells in +N and −N for the indicated time periods. +N 0 h, cells in the exponential phase before transferring to −N. β-Actin was used as a loading control. Mixed-model regression (n = 4); exact P values are provided in ref. ²⁹, Data File S12. t, time periods for the WT measurements. *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001.

Source data

To further dissect potential TORC1-independent regulation by the RTG pathway, we performed a secondary screen of selected mutants with rapamycin in both nitrogen-rich and nitrogen-starved conditions for 10 h and analysed them using our deep-learning approach (Extended Data Fig. 9). This procedure allowed us to identify nitrogen-sensing mutants whose influences on autophagy deviated from canonical TORC1–rapamycin-dependent regulation. As expected, rapamycin treatment boosted autophagy induction, elevating autophagosome formation (BF_t VAM6:ATG1) in +N and synergized with genetic disruptions of core TORC1 components (tor1Δ and tor2Δ, kog1Δ and pib2Δ; Fig. 5c and Extended Data Fig. 9a,f). Notably, rapamycin alone had a negligible effect on autophagosome clearance (BF_t WT:VAM6; Fig. 5c). To specifically score nitrogen-by-genotype interaction effects, while controlling for autophagy induction by rapamycin, we compared the differential BF_t response to rapamycin across all mutants and identified mutants that deviated from the expected effect (in +N) defined by a linear regression line (Fig. 5d and Extended Data Fig. 9g). Interestingly, this analysis revealed RTG and SPE mutants as consistent outliers, suggesting distinct nitrogen-responsive autophagy programmes outside TORC1 regulation. Specifically, rtg1Δ and rtg3Δ deletion mutants displayed rapamycin-independent elevation of autophagosome formation in +N, whereas spe2Δ and spe3Δ, and to a lesser extent spe1Δ deletion mutant, desensitized both autophagosome formation and clearance in response to −N (Fig. 5d and Extended Data Fig. 9f,g). These chemogenetic profiles further corroborate RTG signalling and spermidine as nitrogen status-dependent modulators of autophagy independently of TORC1-mediated nutrient sensing.

Extended Data Fig. 9. Rapamycin response screen.

a, Autophagy (%) response dynamic in response to starvation, rapamycin, or both for WT, ORF control, and selected components of the TORC1-signalling pathway, the RTG pathway, and the spermidine-synthesis pathway. b, Similarity in autophagy perturbations (%) between different treatment conditions. c, Similarity in BF responses between different treatment conditions. b,c, R indicates the Pearson correlations, and p indicates their respective p-values, as shown in the figure. d, Reproducibility in autophagy % and BFs between models and between experimental conditions. The box plots indicate distributions of replicate pairs, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR. e, Correlation between log BFs and overlap coefficient showing strong correspondence with overall autophagy (BF WT:ATG1) across all conditions. f, Comparison of the BF response to starvation dependent on the rapamycin treatment status. g, Comparison of the BF response to rapamycin treatment dependent on the nitrogen status. e–g, R indicates the Pearson correlations, and p indicates their respective p-values, as shown in the figure.

To independently validate these findings, we quantified autophagy using separate biological assays that tracked the delivery of distinct cargos, GFP–Atg8 and the modified phosphatase precursor Pho8Δ60, to the vacuole. Pho8Δ60 processing revealed elevated flux under both nitrogen conditions in all RTG mutants (Fig. 5e), with rtg1Δ and rtg2Δ in particular having a hyperactive phenotype that was attenuated by the deletion of ATG1 in rtg1Δ cells (Fig. 5e). These mutants also exhibited increased GFP–Atg8 cleavage under basal conditions (+N) but with a diminishing effect of starvation, consistent with a potential exhaustion-like phenotype (Fig. 5e,f). Importantly, RpS6 phosphorylation (RpS6-p) remained high in these mutants compared with WT cells, indicating a decoupling from TORC1 signalling (with elevated RpS6-p in +N; Fig. 5f). This contrasted with spermidine-synthesis mutants, which had a delay in autophagosome formation and clearance (Fig. 5d and Extended Data Fig. 8a) and elevated RpS6-p levels in −N (Extended Data Fig. 8c). Together, these data strongly indicate that the RTG pathway regulates nitrogen-responsive autophagy independently of TORC1, buffering both autophagosome formation and flux in nutrient-rich conditions. This suggests a broader architecture of nitrogen-sensing pathways that fine tune autophagy through parallel mechanisms.

Cross-omics inference reveals transcriptional mediators of autophagy regulation

Genome-wide quantification of dynamic signatures enables the construction of comprehensive systems-level models capable of predicting and simulating biological processes under different conditions without sacrificing structural complexity or causal sufficiency. This approach has been successfully demonstrated in yeast using biologically informed machine-learning models for growth and metabolism, leveraging the abundance of available multi-omics data resources^49,50. As a proof-of-principle, we applied random forest to predict various autophagy perturbation metrics from several yeast gene-deletion datasets. These datasets included transcriptomic, proteomic and amino-acid metabolomic features^51–53 as well as genetic (SGA) and STRING-based genome-wide correlation networks^30,32. All datasets were able to predict autophagy perturbations, with performance levels mirroring the experimental reproducibility of the different perturbation parameters (Extended Data Fig. 10a and Supplementary Fig. 4f). Notably, the genome-wide networks performed best overall, with test predictions achieving median correlations of more than 40% between the predicted and true outcomes for the BF_t metrics in −N (Extended Data Fig. 10a).

Extended Data Fig. 10. Cross-omics prediction of autophagy dynamics.

a, Benchmarking prediction of different autophagy metrics using 10-fold training and testing. Model performance is measured as the median test correlation between measured and predicted values. Transcriptome: Kemmeren et al.⁵¹; Proteome: Messner et al.⁵²; Metabolome: Mülleder et al.⁵³; SGA-PCC: Costanzo et al.³⁰. b, Benchmarking prediction of autophagy perturbation dynamics from TF expression induction profiles⁵⁴ using 10-fold training and testing, with random sampling of the entire TF time series (left panel) or random sampling of individual measurements (right panel). Model performance is measured as the median test correlation between measured and predicted values. Colour coding indicates different feature pre-selection criteria. c, Adjusted SHAP values for regulation of gene expression across models and autophagy metrics comparing feature classes based on autophagy perturbation profiles. Points and bars indicate mean and standard error. d, Adjusted SHAP values for regulation of gene expression comparing feature classes across models and autophagy metrics. F: formation genes; C: clearance genes with negative BF_t z-scores beyond two standard deviations from the mean. Points and bars indicate mean and standard error. e, Gene-wise SHAP values for differential regulation of gene expression predicting BF responses. f, qRT-PCR for expression of GCN4, ATG13, GTR1, GTR2, VAM6, VPS15, VPS34 and ATG9 in rtg1Δ, rtg2Δ, rtg3Δ, and WT cells subjected to 4 h of −N, followed by 1 h of +N. Graphs show mean and standard error (n = 3). *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001; two-sided t-test; exact P-values are provided in Data File S13. g, Correlations in differential expression between mutants per time point for the panel of gene targets analysed in Fig. 6e and Extended Data Fig. 10f. h, Median gene-expression change in response to TF induction in Hackett et al.⁵⁴. i, Representative electron micrographs showing autophagic bodies (AB) after 1 h of nitrogen starvation in WT and rtg1Δ cells (top panel), along with quantifications of AB area (a.u.; bottom panel). WT: n = 93 and rtg1Δ: n = 91. Scale bar, 500 nm; n.s., one-way ANOVA. The box plots indicate autophagic bodies distributions, showing the median (line), interquartile range (25th–75th percentile, box), and whiskers extending to 1.5× the IQR.

Given the prevalence of transcription factor (TF) deletions that strongly influenced autophagy dynamics and the broader challenge of identifying the causal mediators through which TFs regulate cellular processes, we performed a mediation analysis to identify gene-expression changes that predicted initial perturbations in autophagy dynamics. To achieve this, we analysed the gene-expression induction profiles of 186 overlapping TFs from the Induction Dynamics Gene Expression Atlas (Fig. 6a)⁵⁴.

Fig. 6. Cross-omics mediation analysis reveals transcriptional tuning of autophagy.

a, Transcription factor gene-expression induction profiles from Hackett and colleagues⁵⁴ were used to predict the corresponding autophagy perturbation dynamics of TF gene deletions using random forest (RF). Gene targets (model features) with high importance for predicting autophagy perturbations were evaluated as potential causal mediators of TF influence on autophagy. b, Adjusted Shapley additive explanations (SHAP) values for differential regulation of gene expression comparing feature classes across models and autophagy metrics (statistics in ref. ²⁹, Data File S14). Data are the mean ± s.e.m. a,b, F, formation genes; C, clearance genes with negative BF_t z-scores beyond two standard deviations from the mean. c, Gene-wise SHAP values for differential regulation of gene expression predicting autophagy responses. d, Hierarchical clustering of average TF-induced gene-expression changes (z-scores) for gene targets that were selected as important model features (using sumGain). e, Expression levels (determined using quantitative PCR (qPCR) with reverse transcription) of the ATG initiation genes ATG1, ATG13 and ATG14 in rtg1Δ, rtg2Δ, rtg3Δ and WT cells subjected to 4 h of −N, followed by 1 h of +N. Data are the mean ± s.e.m.; n = 3. Two-sided Student’s t-test; exact P values are provided in ref. ²⁹, Data File S13. f, Differential expression of various target genes in rtg1Δ, rtg2Δ and rtg3Δ cells per phase from e and Extended Data Fig. 10f; linear regression (n = 3); exact P values are provided in ref. ²⁹, Data File S13. g, Immunoblot analysis of GFP–Atg8 cleavage and RpS6-p for WT, rtg1Δ, gcn4Δ and gcn4Δ rtg1Δ cells (top) as well as WT, rtg1Δ, gln3Δ and gcn3Δ rtg1Δ cells (bottom) in −N for the indicated time periods. +N 0 h, cells before transferring to −N. Average quantifications and mixed-model regression are provided beneath the blots (n = 4 WT and rtg1Δ, and 2 for the double deletions); exact P values are provided in ref. ²⁹, Data File S12. h, Average number of autophagosomes per cell in WT, vam3Δ, rtg1Δ and rtg1Δ vam3Δ cells after 2 h of nitrogen starvation-induced autophagy measured at 15-min intervals. Data are the mean ± s.e.m. (n = 6). i, Area of autophagic vesicles in WT, vam3Δ, rtg1Δ and rtg1Δ vam3Δ cells (ref. ²⁹, Data File S18). Two-sided Student’s t-test (n = 6). Area in pixels; 1 pixel represents 70 nm. The box plots indicate the vesicle size distributions showing the median (line), IQR (25–75th percentile; box) and whiskers extending to 1.5× the IQR; points outside this range are plotted as outliers. j, Representative images for h and i taken 90 min after nitrogen starvation. Dashed circles mark the cell contour. Scale bars, 3 µm (×100; 1 pixel represents 0.07 µm). k, Immunoblot of Ape1 processing and RpS6-p in WT and rtg2Δ cells in −N for the indicated time periods, with or without Ape1 overexpression through CuSO₄ pretreatment for 20 h (n = 2). *P ≤ 0.05; **P ≤ 0.01; ***P ≤ 0.001.

Source data

Models built with only potential causal mediators (that is, significant genes, HMP ≤ 0.01, excluding the TFs themselves) were able to sustain predictive power, exhibiting only a moderate decrease in performance compared with genome-wide models (Extended Data Fig. 10b). Moreover, expression changes for gene targets associated with severe autophagy perturbations were estimated to mediate stronger influences across all models (Extended Data Fig. 10c), with targets involved in autophagosome formation or clearance acting as positive mediators of autophagy regulation overall (Fig. 6b and Extended Data Fig. 10d). Interestingly, only targets involved in clearance were predicted to mediate a positive influence on clearance dynamics, whereas targets involved in autophagosome formation were predicted to exert a negative influence (Fig. 6b and Extended Data Fig. 10d, BF WT:VAM6 panels). This suggests that transcriptional suppression of certain autophagy-induction genes may play a role in promoting autophagosome clearance. ATG1 and ATG14 were identified as major mediators of autophagy initiation control (Fig. 6c), whereas VAM10 and the ESCRT-0 subunit-encoding VPS27 were identified as key mediators of autophagosome clearance regulation (Extended Data Fig. 10e).

By clustering the average induction response of predicted mediators for select autophagy-regulating TFs, we found that RTG-mediated suppression of autophagy induction was probably mediated through Gcn4-regulated expression of ATG1 (Fig. 6d)^54,55. Quantitative PCR with reverse transcription revealed that ATG1 expression was substantially greater in rtg1Δ cells than WT controls under both −N and +N conditions. Moreover, the effects of different RTG subunits on expression of ATGs and upstream TORC1 regulators were decoupled during starvation (Fig. 6e,f and Extended Data Fig. 10f,g). When rtg1Δ was crossed with gcn4Δ strains, the hyperactive autophagy phenotype was readily rescued; however, this effect was not observed in the double-knockout rtg1Δ gln3Δ strains (Fig. 6g). This epistatic effect of rtg1Δ on gln3Δ was consistent with rtg1Δ suppressing GLN3 expression in +N, leading to decreased expression of Gln3-regulated gene targets in −N (for example, ATG14; Fig. 6f)⁴².

ATG1 deletion masked the autophagic flux phenotype in rtg1Δ cells, as measured by Pho8Δ60 activity (Fig. 5e). To test whether RTG1 affects autophagy through the induction stage, we crossed rtg1Δ strains with vam3Δ strains to block autophagosome fusion, and quantified the number and size of autophagosomes per cell using high-resolution imaging. Both rtg1Δ and rtg1Δ vam3Δ cells exhibited higher initial autophagosome counts than WT and vam3Δ cells (Fig. 6h). Following starvation, the rtg1Δ cells showed a transient increase in counts, which started to decline after 60 min, whereas the rtg1Δ vam3Δ cells continued to accumulate autophagosomes throughout the entire time course (Fig. 6h(right)). This difference in dynamics suggests that the larger autophagosome pools in rtg1Δ cells undergo a clearance burst following starvation, an effect that is abolished in rtg1Δ vam3Δ cells. In contrast, WT and vam3Δ cells displayed a stable difference in autophagosome counts, with an asymptotic increase towards a steady state in both strains (Fig. 6h(left)). Interestingly, the hyperactive autophagosome formation in rtg1Δ cells was also associated with slightly smaller vesicles (Fig. 6i,j and Extended Data Fig. 10i), suggesting potential exhaustion of resources. To investigate whether control of autophagy induction by the RTG pathway also influences cargo sequestration and clearance, we performed a giant Ape1 assay in which we overexpressed Ape1 from a copper-inducible promoter to challenge autophagic capacity with large protein aggregates⁵⁶. Because Ape1 overexpression had a detrimental effect on the growth of rtg1Δ strains, the assay was performed in rtg2Δ strains. Compared with WT cells, rtg2Δ cells exhibited an increased capacity to process Ape1 under both endogenous and overexpressed conditions (Fig. 6k). Together, these findings highlight the RTG pathway as a crucial modulator of autophagosome formation.

Discussion

The study of the dynamics of cell systems is becoming increasingly relevant as we strive to achieve predictability and guidance from our scientific models. This is particularly important for autophagy, which is intrinsically linked to an organism’s metabolism and growth and continuously fluctuates with environmental changes such as feeding cycles and circadian rhythms. To obtain a comprehensive understanding of autophagy activation during phases of declining nutrient levels requires deep insight into functional interactions with rate-limiting factors (Fig. 3g). Given that genetic perturbations of the core machinery severely compromise autophagy execution, high-quality temporal data from regulating genes and multi-omics strategies to model changes in autophagy dynamics are essential. Our observation that there is a concordance between autophagy perturbation profiles and the direction of functional similarities with the core machinery (Fig. 3g and Supplementary Fig. 6d), as well as a concordance in phenotypes between TF deletions and the predicted gene-expression mediators (Fig. 6b), provides strong support for the utility of the presented data.

The number of dynamic autophagy phenotypes detected with increased screening depth and sensitivity reflects the vast amount of cellular and environmental information integrated for optimal control of the process. The regulatory network complexity charted in this study is not surprising given the critical role of autophagy in maintaining cellular homeostasis. By decomposing phenotypes into groups with distinct kinetic influences, we uncovered a regulatory asymmetry between the activation and inactivation of autophagy (Fig. 1f,g). Notably, the majority of mutants (ultrasensitive, hyposensitive and hyperactive) influenced the timing of the starvation response. Although a variety of activatory signals may be crucial for metabolic adaptation in a fluctuating environment, the timely shutdown of the process may be under stricter control to prevent excessive degradation and cell death. It is also possible that autophagy is subject to auto-inhibitory regulation through the depletion of ATG proteins, thereby increasing sensitivity to suppressive signals from nitrogen replenishment.

The six autophagy profiles discovered were remarkably well-organized across the genome-wide landscape (Figs. 2c,d, 3 and Extended Data Fig. 5), with regulatory genes connected to the autophagy core machinery in a manner that reflected their tuning strength and functional characteristics (Fig. 3d,g). The main exception was the largest profile cluster, the ultrasensitive mutants, which exhibited weaker perturbation effects, no distinct network distribution and no significant connection to the core machinery as a group. Although this group includes several TOR signalling components, it probably also includes many genes with indirect effects on the starvation response (‘slope’) by influencing factors such as synchrony and cell-cycle stage within the population. In contrast, hyperactive mutants exhibited stronger perturbation effects and were significantly closer to the core machinery, with multiple strong associations with ATGs, HOPS components and vacuole tethering proteins (Fig. 3d,g). These findings suggest that further increasing overall autophagy during complete nitrogen withdrawal may require genetic interventions that engage multiple components of the process. Integration of phenotypic profiling data with network connectivity information could provide valuable design principles for modulating autophagy.

The genome-wide perturbations also provide a data source for parametric modelling of autophagy dynamics. In this study we have provided some starting points by showing the predictive power across multiple genome-wide yeast datasets (Extended Data Fig. 10a). By demonstrating the utility of this approach, we identified a novel phase-dependent tuning mechanism through the transcriptional repression of autophagy by the RTG signalling pathway. With Gcn4 positively regulating RTG gene expression (Extended Data Fig. 10h)⁵⁴, the retrograde pathway may function as a negative-feedback loop, buffering a positive interaction between Gcn4 and Gln3. This mechanism supports temporal coordination of ATG1 expression, limiting uncontrolled autophagy induction and promoting autophagosome clearance^54,55. Recent discoveries suggest that Atg1 plays a role in activity-dependent disassembly of several ATG subcomplexes involved in autophagosome formation⁵⁷. Thus, modulating the relative levels of ATG proteins may play a crucial role in tuning the strength and persistence of autophagy induction and the imbalanced expression observed in RTG deletion cells could explain the diminishing returns in autophagic flux from starvation (Figs. 4g, 5e,f and 6h). This adaptive control of autophagy dynamics may be evolutionarily conserved, as the yeast retrograde response and analogous pathways in higher organisms support cell survival in response to environmental perturbations or cellular challenges associated with ageing⁵⁸.

The versatility of our expectation-based deep-learning approach provided a richer characterization of the autophagic feature space with high accuracy. Human annotation of heterogeneous data is a fundamental challenge for the utility of machine learning in cell biology⁵⁹, while experimental generation of high-content reference data is becoming increasingly easy. Defining the ‘ground truth’ of a model using experimental control conditions can easily be extended to a range of other machine-learning tasks⁵⁹. When combined with a systematic analysis across functionally related gene sets, this approach revealed discrete phenotypes that might otherwise have gone unnoticed. Treatment with rapamycin also revealed further levels of variation and otherwise unobserved nutrient–gene interactions (Fig. 5c,d and Extended Data Fig. 9f,g). Hence, further manipulations of the culture environment and metabolic growth history can potentially unveil an even broader spectrum of unexplored autophagy regulatory mechanisms.

Due to the sensitivity of our approach, we were able to detect a vast number of genes with a significant influence on autophagy dynamics—approximately 30% of the investigated genome (Fig. 1). Our study also includes an analysis of essential yeast genes using the DAmP collection, which reduces protein expression by destabilizing mRNA molecules. Although this approach does not always guarantee a hypomorphic phenotype³⁰, it was preferred over conditional alternatives, such as temperature-sensitive alleles, which can introduce confounding effects on autophagy as a consequence of elevated temperatures^60,61. Across the dataset we observed library-wide differences in response kinetics and BFs, which suggests a potential confounding effect from strain background or DAmP treatment (Fig. 1b and Extended Data Fig. 7j). However, discrete differences also emerged between the KO library and various WT control strains, varying according to the growth protocol used (Fig. 1b and Extended Data Fig. 4f). Given the sensitivity of our methods and the multitude of factors influencing autophagy dynamics, these differences are unsurprising. Notably, essential genes were over-represented among hyperactive and insufficient activation mutants (Fig. 1h), suggesting that elevated basal autophagy (Supplementary Fig. 3a) may be a response to mRNA stress, potentially through stress granule formation, ribosome-associated quality control or TOR-independent nutrient sensing^62–64. Genes with these phenotypes were indeed enriched in functional clusters related to nucleotide metabolism, RNA processing and translation (Figs. 2c, 3a, Supplementary Fig. 5 and Extended Data Fig. 5a).

In conclusion, these data expand the knowledge base of autophagy-regulating genes, providing deeper insight into its phenotypic landscape within a broader regulatory context. The resulting genome-wide profiling repository not only represents a foundational reference for systems-level autophagy characterization but also serves as a powerful hypothesis-generating tool. It is publicly available via the Autophagy Dynamics Repository in Yeast (AutoDRY) web portal (https://cancell-apps.medisin.uio.no/AutoDRY/). Our study thus presents a time-resolved, genome-wide map of autophagy activation and inactivation, offering a comprehensive view of its dynamic regulation in cellular nutrient responses.

Methods

High-content screening

Details of genome-wide libraries

The yeast mutant libraries harbouring the autophagy reporter (autophagy library KO-A and DAmP-A) were constructed based on the collection of non-essential deletion mutants (YKO collection; Open Biosystems) and the collection of essential genes with a decreased abundance by mRNA perturbation (DAmP collection; Open Biosystems). The DamP mutations were chosen over conditional temperature-sensitive mutations to eliminate temperature as a variable that could impact the autophagy response^60,61. The mutant strains were originally distributed across 56 KO-A and 12 DAmP-A 96-well plates. To introduce control strains into each plate, the mutant strains in row A were transferred to new plates designated as Mixed plates. The mutants with significantly reduced growth rates, which may affect the cell count for imaging, were re-inoculated at higher density into independent plates that were annotated as Recovery plates. A subset of mutants was rearranged into Repetition plates, which were used for validation and reproducibility experiments. Rapamycin plates were prepared for the analysis of rapamycin response.

Each plate included three control strains in triplicate, specifically the SGA starting strain WT (Y7092)⁶⁵, and the deletion mutants vam6Δ::kanMX4 (KO-A) and atg1Δ::kanMX4 (KO-A). The libraries Mix, Recovery, Repetition and Rapamycin included the deletions of three dubious ORFs—ydr535cΔ::kanMX4 (KO-A), ycr102w-aΔ::kanMX4 (KO-A) and yel074wΔ::his3MX6 (this study)—in addition to atg1Δ and vam6Δ to control for a potential effect of the kanamycin and histidine selectable markers on the autophagy response. The complete list of mutant strains and controls distributed into the arrays KO-A, DAmP-A, Recovery, Mixed, Repetition and Rapamycin is provided in ref. ²⁹, Data File S1.

Preparation of genome-wide libraries

To prepare genome-wide libraries for high-throughput screening, duplicates of the yeast libraries were created in 96-well plates (Nunc Microwell, Thermo Fisher Scientific). In each well, 1 µl of both control strains and mutant strains was inoculated into 150 µl synthetic medium containing amino acids (SC medium; 6.9 g yeast nitrogen base with ammonium sulfate and without amino acids, 20 g glucose anhydrous and 0.790 g Complete Supplement Mixture (Formedium) per litre). To promote optimal growth and a low cell death rate, the duplicate plates were incubated at 25 °C for 72 h before being stored at −80 °C with 20% glycerol.

A working batch of 16 plates from duplicates was prepared five days before an imaging round. In each plate, 1 µl of both mutant and control strains was inoculated into 200 µl SC medium. Following incubation at 25 °C for 48 h, these plates were stored at 4 °C with a cover film. The batch was divided into four sets and each set underwent pre-culturing and imaging on consecutive days.

Cell growth and starvation protocol for high-content screening

The growth protocol was optimized to enable high-throughput cell growth while inducing an optimal autophagy response in yeast cells. This involved periodic medium replacement, which kept cells metabolically active and prevented premature autophagy initiation due to nutrient unavailability during prolonged incubations without shaking. It ensured that yeast cells were in the logarithmic phase and minimized population response variability before the onset of starvation.

After incubation of the previous batch for 48 h, a 1 µl volume of strains with a normal growth rate and 4 µl of strains with a reduced growth were transferred into 200 µl fresh SC medium. These plates were then incubated at 25 °C for an additional 48 h, followed by reinoculation and another 24-h incubation at 25 °C. For time-lapse imaging, the cells (4–6 µl) were inoculated into 96-well black plates (Corning Black microplate, Sigma-Aldrich) coated with 0.25 mg ml⁻¹ concanavalin A and incubated for 8 h at 25 °C to ensure the cells were in the exponential phase. To induce autophagy, the cells were washed twice with sterile MilliQ water and transferred to nitrogen-starvation medium for 12 h (SD-N medium; 1.9 g yeast nitrogen base without ammonium sulfate and amino acids plus 20 g glucose anhydrous (Formedium) per litre). Subsequently, the same cells were transferred to SC medium (at t = 12 h) to record autophagy inactivation following nitrogen replenishment for 7 h. Live imaging was performed at 30 °C, capturing images every hour for 19 h. The cells were maintained at room temperature during the intervals between each time point of image acquisition.

Autophagy reporter details

To assess autophagy activity, translocation of fluorescent protein-tagged Atg8 from the cytoplasm to the vacuole was monitored using fluorescence microscopy⁶⁶. For time-lapse experiments, we used optimized mNG fluorescent protein fused to the amino terminus of Atg8 under the control of the ScTDH3 (GPD) promoter. To visualize the vacuole, mCherry fluorescent protein was fused to the carboxy terminus of the vacuolar proteinase A Pep4.

Plasmid construction

The pYM-N natNT2-GPD1pr–mNG plasmid was generated by replacing the yeGFP coding sequence from pYM-N17 (ref. ⁶⁷) by the codon-optimized mNG DNA sequence⁶⁸. The mNG sequence was amplified from the pFA6-mNG-his3MX6 plasmid using the primers mNG-Fwd-XbaI and mNG-Rv-EcoRI, which included XbaI and EcoRI sites for the tag replacement (ref. ²⁹, Data File S1).

Integration of autophagy reporter into libraries and SGA query construction

The autophagy reporter was crossed into the collections of non-essential and essential mutants to create the YKO-A and DAmP-A collections, respectively. Triple allele strains harbouring the autophagic reporter and the deletion of non-essential or DAmP alleles were obtained by adapting the SGA analysis method^69,70.

The query bait containing autophagy reporter was created using the SGA host strain MATα haploid Δcan1::STE2pr-Sp_his5 Δlyp1 (Y7092; a gift from C. Boone, University of Toronto, Canada). Endogenous ATG8 and PEP4 were tagged with mNG and mCherry using a standard PCR-based approach⁶⁷. Briefly, natNT2:GPD1pr-mNG-ATG8 was constructed by PCR amplification of a natNT2-GPD1pr-mNG fragment from the pYM-N17–mNG plasmid using the primer pair ATG8-S1 and ATG8-S4, whereas PEP4–mCherry:hphMX6 was constructed by integrating an mCherry:hphMX6 PCR fragment amplified from plasmid pBS35 using the primer pair PEP4-S3 and PEP4-S2 at the 3′ end of the PEP4 locus (ref. ²⁹, Data File S1). This resulted in the construction of JEY11511 (MATα can1Δ::STE2pr-Sp_his5 lyp1Δ his3Δ1 leu2Δ0 ura3Δ0 met15Δ0 natNT2:GPDpr-mNG–ATG8 PEP4–mCherry:hphMX6). The SGA bait strain JEY11511 was crossed with a target array of 4,760 viable MATa deletion mutants and 1,159 conditional alleles of essential genes (KO and DAmP mutants, respectively) in the BY4741 background (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0), wherein each ORF or 3′ untranslated region was disrupted by a kanMX4 resistance cassette marker. Crossing of JEY11511 was performed in quadruplicate by pinning onto fresh yeast extract peptone dextrose (YPD) agar plates using a Singer RoToR system (Singer Instruments), followed by growth for two days. The cells were subjected to two rounds of pinning onto diploid selection medium (YPD agar containing 200 μg ml⁻¹ geneticin (Thermo Fisher Scientific, G418) and 200 μg ml⁻¹ hygromycin (Gibco)) and cultured at 30 °C for 1–2 days. The cells were then pinned onto pre-sporulation medium (15 g Difco nutrient broth (BD Biosciences), 5 g bacto-yeast extract (Thermo Fisher Scientific), 10 g bacto-agar (Thermo Fisher Scientific) and 62.5 ml of 40% glucose (Thermo Fisher Scientific) per 500 ml) and cultured at 30 °C for three days. Cells from the pre-sporulation medium were then pinned onto sporulation medium (10 g potassium acetate, 0.05 g zinc acetate and 20 g bacto-agar per litre supplemented with a final concentration of 50 μg ml⁻¹ G418 and 50 μg ml⁻¹ hygromycin) and incubated for seven days at 30 °C. The resulting spore-containing cells were then subjected to two rounds of pinning onto diploid killing medium (1.7 g yeast nitrogen base without amino acids and without ammonium sulfate (BD Biosciences), 1 g L-glutamic acid monosodium salt, 2 g SC dropout mix without lysine, canavanine and histidine (US Biological), 20 g bacto-agar and 50 ml of 40% glucose per litre supplemented with a final concentration of 50 μg ml⁻¹ thialysine (Sigma-Aldrich), 60 μg ml⁻¹ canavanine (Sigma-Aldrich), 200 μg ml⁻¹ G418 (Gibco), 200 μg ml⁻¹ hygromycin (Invitrogen) and 100 μg ml⁻¹ nourseothricin (Jena Bioscience), followed by culturing at 30 °C for five days for the first pinning and two days for the second pinning. The cells were then subjected to two rounds of pinning and growth on haploid selection medium (1.7 g yeast nitrogen base without amino acids and without ammonium sulfate, 1 g L-glutamic acid monosodium salt, 2 g SC dropout mix without histidine, 20 g bacto-agar and 50 ml of 40% glucose per litre supplemented with a final concentration of 200 μg ml⁻¹ G418, 200 μg ml⁻¹ hygromycin and 100 μg ml⁻¹ nourseothricin) and cultured at 30 °C for two days. The cells were then pinned and cultured on YPD agar, followed by storage at −80 °C in YPD medium containing 20% (vol/vol) glycerol. The resulting MATa haploid triple mutants were used for high-content screening.

High-content fluorescence imaging

The high-content screening for autophagy analysis was carried out in a 96-well-plate format using an ImageXpress Micro Confocal microscope equipped with a temperature control unit, light-emitting-diode light source and specific filter sets (Molecular Devices). The images were acquired using automated acquisition in wide-field, utilizing the Nikon Plan Apochromat ×60/0.95 Air objective coupled to a 16-bit 4MPix scientific CMOS camera and a laser-based autofocusing unit. Image acquisition was controlled using the ImageXpress Analysis Software (Molecular Devices). Live-cell imaging was conducted at 30 ± 2 °C. Fluorescence signals were captured using specific filter sets for fluorescein isothiocyanate and TexRed, with exposure times of 200 and 800 ms, respectively. For transmitted-light images, an exposure time of 10 ms was used. The same cells were imaged every hour for 19 h with two or three images per well by selecting adjacent imaging sites near the centre of the well and employing modus focus on the plate and bottom to maintain the same plane.

Automatic image analysis using Fiji/ImageJ for high-content screening

We developed an automated image analysis tool using Fiji/ImageJ to extract features from the genome-wide dataset and evaluate changes in the autophagy reporter along the time-lapse experiments. The tool employed a macro to extract parameters for each image categorized by well (row + column), time point (starvation/replenishment), image within well (s1, s2, s3) and channel (w1 = BF, w2 = green, w3 = red). The extracted features included parameters related to both fluorescence signals and cell morphology. Parameters related to fluorescence signals were the mean and s.d. of signal intensity, modal and minimum signal values, and integrated median signal intensity. Parameters related to cell morphology included cell area, centroid, centre coordinates, perimeter, bounding box, best-fit ellipse dimensions, shape descriptors, ferret dimensions, skewness and kurtosis.

For image segmentation, we first enhanced contrast using a saturation value of 0.3 and then sharpened it to improve clarity and pixel contrast. We further enhanced contrast with the same saturation value to refine the image after sharpening. We then applied the ‘Unsharp Mask’ plugin with a radius of five and a mask value of 0.70. We used the ‘Auto Threshold’ plugin with the default method and set the image to white. To finalize the segmentation and make objects white on a black background, we inverted the image and applied the ‘Minimum’ filter with a radius of 0.5. To eliminate small holes or gaps in the image, we used the ‘Close-’ plugin and followed it with another inversion of the image.

To identify and count the objects in the image, we used the ‘Analyze Particles’ plugin with a size range of 400–2,500 and circularity range of 0.50–1.00. After correcting the background using a rolling value of 50 pixels and measuring the signal intensity of regions of interests in both channels, we quantified co-localization by multiplying the pixel values of the two channels with the ‘multiply create 32-bit’ command.

Deep-learning analysis

Data preparation for model training

All neural net models were trained using the Keras application programming interface (version 2.2.5.0) for Tensorflow (version 2.00) in R (version 3.6.3)⁷¹. The training data with associated state labels were compiled from atg1Δ or WT cells under the experimental conditions where autophagic or non-autophagic features were uniformly maximized across the cell culture populations (Extended Data Fig. 1a). Wild-type cells in t₀ or from 5 h in nitrogen replenishment and atg1Δ cells across the entire experimental protocol were assigned state 0 (non-autophagic), and WT cells from 8 h to 12 h of nitrogen starvation were assigned state 1 (autophagic). Although autophagic features, including free mNG–Atg8 puncta and co-localization of mNG–Atg8 and Pep4–mCherry, could be observed earlier in WT cells, 8 h of starvation represented a demarcation where mNG–Atg8 and Pep4–mCherry clearly overlapped homogeneously across all WT cells, representing that autophagosomes had been delivered to the vacuole. Experimental outliers visibly deviating from these standards were manually removed from the training data. Furthermore, cells with low mNG–Atg8 signal (mean intensity < 150) or either low (mean intensity < 120) or high (mean intensity > 5,000) Pep4–mCherry signal were removed from the training data.

Single-cell parameters for variations in signal distributions from both colour channels and the co-occurrence were used as input features. Before training, we computed the mean and s.d. feature vectors from the entire WT dataset, including all time points, which were used to transform the training data features and transform all future new data. The resulting training set included 1,073,981 observations in state 0 and 269,655 observations in state 1.

Model training, hyperparameter search and model selection

To train the models binary, cross-entropy was used as loss and the models were trained using the AdaGrad algorithm, with a batch size of 50,000, learning rate of 0.01 and patience of five epochs. In the hyperparameter search, 10% of the data were kept out for testing. Of the remaining dataset, 80% of the data were used for training the models and 20% were used as validation data for early stoppage. To find the optimal model architecture a grid search was performed over all combinations of activation functions (relu, elu, tanh and sigmoid) with networks of different sizes (Extended Data Fig. 1a). The networks were grown automatically by defining a number of layers (1–5 and 6–8, respectively) and neurons in the first layer (60–120 and 80–120, respectively, with increments of five), and adding the layers with a decay rate in the number of neurons from the former layer (0.5–1.0 and 0.7–1.0, respectively, with increments of 0.05), with a minimum layer size of ten neurons. This resulted in 3,096 models that were trained for a maximum of 150 epochs (with early stoppage). The models were evaluated based on the test-set prediction averages compared with the respective state labels, training and validation set loss and accuracy, and model overfitting (Extended Data Fig. 1b–d). With the exception of sigmoid-based models, which underperformed compared with the other models, most architectures resulted in similar test performances (Extended Data Fig. 1b). Thus, the sigmoid-based models were excluded from further analyses.

Model complexities were quantified from the number of model weights. Given that the model test and validation performances scaled asymptotically with model complexity, whereas more complex models also tended to have more overfitting for some architectures (relu), we defined a model complexity cutoff at the smallest model in the top 1% (validation loss; Extended Data Fig. 1d). Moreover, network size may affect the rate of convergence and generalization of a model for complex function approximations, where 150 epochs may not be enough. Subsequently, we selected the five top-performing models (validation loss) among the least and most complex models (given the complexity cutoff) for relu, elu and tanh activation functions, respectively (Extended Data Fig. 1e). These 30 models were re-evaluated by re-training them for a maximum of 3,000 epochs (with early stoppage) using all of the data in an 80/20 training and validation split. The models were evaluated based on training and validation set loss and accuracy as well as the between-experimental correlation of predicted autophagy dynamics (plate averages for WT, atg1Δ and vam6Δ data points combined). On average, the less complex models slightly outperformed the more complex ones for architectures using tanh and elu but not relu activation functions (Extended Data Fig. 1e).

In addition, the predicted last latent layers (autophagic feature vectors) for the top two models (validation loss) for each activation function were extracted for 500,000 randomly sampled WT, atg1Δ or vam6Δ cells from any time point and embedded with UMAP projection (Extended Data Fig. 1a) using the umap package (version 0.2.7.0) in R⁷². The UMAP was used to force a low-dimensional and regularized representation of the predictive information in the neural networks and to investigate the relative contribution of noise and extracted autophagic features between three distinct autophagy phenotypes. The UMAP distance metric was set to Pearson correlation, and n_neighbors were set to 10, 15 or 30. The between-experimental correlation of predicted UMAP dynamics (coordinate-wise plate averages for WT and vam6Δ data points combined) was used to test latent-space robustness (Extended Data Fig. 1e). The two highest-performing and most robust models across multiple metrics (elu model 30, 7 layers, 90 neurons in layer 1 and 66 neurons in layer 7; tanh model 22, 8 layers, 115 neurons in layer 1 and 37 neurons in layer 8) were deployed for genome-wide (GW) predictions and used to corroborate each other (Extended Data Fig. 1h). Statistical analyses of per cent autophagy response dynamics (Figs. 1–3) were based on predictions from model 30. For latent-space analyses, model 22 had less noise and was primarily used to visualize the UMAP dynamics (Fig. 4), but all statistical analyses and inferences from the UMAP distributions were averaged or corroborated by predictions from both models (Fig. 4).

Parametric UMAP

Given that transforming new data with UMAP is slow, we implemented a fast parametric transformation of the inputs (autophagic feature vectors) to the embedded coordinates using a neural net (parametric UMAP: fully connected, 8 layers, 120 neurons per layer, tanh activation functions, with bivariate linear output and trained using the adam optimizer with mean absolute error loss; Extended Data Fig. 1a). The specific implementation here was conceived before the publication by Sainburg et al.⁷³ but follows the same essence.

Genome-wide predictions and quality control

New screen data were filtered for cells with low (mean intensity < 120) or high (mean intensity > 2,000) Pep4–mCherry signal. A more stringent upper bound was used for the red channel in order to ensure the removal of dead autofluorescent cells. The input features were z-transformed using the feature mean and s.d. from the training data, and the probability for cells activating autophagy (P_a) was predicted using the neural nets. Subsequently, the expected probability of autophagy (predicted autophagy percentage), $A_t\%=\frac{100}{n}{\sum }_{i=1}^n{P}_{\mathrm{a},i,t}$, was computed for the mutant cell population at every time point, where i indicates cell identities and t indicates time points. Alternatively, we also computed the autophagy classification percentage, $A_t^{\mathrm{binary}}\%=\frac{100}{n}{\sum }_{i=1}^{n}I\left(P_{\mathrm{a},i,t}>0.5\right)$, which strongly correlated with $A_t\%$, but was not used as it was less robust for smaller cell populations. Here, I is the indicator function. The expected s.d. per cell was computed as ${\sigma }_{A,t}=\frac{100}{n}{\sum }_{i=1}^n\sqrt{P_{\mathrm{a},i,t}\left({1-P}_{\mathrm{a},i,t}\right)}$ to assess the prediction uncertainty. We also computed the classification uncertainty using the entropy, $S_A=-\frac{1}{n}{\sum }_{i=1}^n\left[P_{\mathrm{a},i}\log \left(P_{\mathrm{a},i}\right)\right)$$\left(+\left({1-P}_{\mathrm{a},i}\right)\log \left({1-P}_{\mathrm{a},i}\right)\right]$ (ref. ²⁹, Data File S2).

Outlier plates were identified by computing the IQR from all plate median responses and finding plates that had a mean absolute deviation greater than 2IQR over the time series. These plates were subsequently repeated. We also identified mutant outliers based on the 99% high-density region of average cell numbers and expected s.e.m. over the time course (Supplementary Fig. 1e), which were subsequently repeated in new experiments with higher initial cell seeding concentration. Four DAmP plates that did not have the replenishment response were manually removed from genome-wide analyses and clustering procedures.

Statistical analyses

Double-sigmoidal model fitting

Double-sigmoidal models were fitted to the data using the sicegar package (version 0.2.4) in R (version 4.2.1)²⁷. This model includes six parameters (maximum autophagy, final autophagy, two transition midpoints (t₅₀) and two slope parameters; Fig. 1c). Tangent lines crossing the transition midpoints were computed to extract additional parameters describing the dynamic part of the curve. To de-noise, interpolate potential missing time points and standardize responses with minor shifts in imaging times, the models were also used to predict the response for all mutants for the entire time series with 30-min increments. The goodness-of-fit was assessed using the log-likelihood (on normalized data) and root mean squared error (ref. ²⁹, Data File S2).

Hypothesis testing and perturbation parameter statistics

Using the double-sigmoidal model predictions, autophagy perturbations were computed from the difference between the interpolated curve fits ($A^f\%$) for all mutants ($m$) and the corresponding plate medians (ctr),

$${\mathrm{\Delta }A}_{m,t}^f\%=A_{m,t}^f\%-A_{\mathrm{ctr},t}^f\%$$

and subsequently averaged per phase (perturbation −N or +N) or over the entire time series (perturbation overall; Supplementary Fig. 2d). Equivalently, all of the mutant parameters were centred by subtracting the corresponding plate medians to compute parameter (parm) perturbations,

$${\mathrm{\Delta }\mathrm{parm}}_m={\mathrm{parm}}_m-{\mathrm{parm}}_{\mathrm{ctr}}$$

(Supplementary Fig. 2e). Autophagy perturbation/parameter errors were computed for each of the plate controls (WT, atg1Δ and vam6Δ) by subtracting their respective plate-wise average. The WT errors were then used as H₀ references for each parameter by fitting the error distributions using the locfit.raw function of the locfit package (version 1.5-9.8) in R⁷⁴ and assessing significance of the parameters by computing P values from the re-sampled fitted H₀ densities. To control for the FDR, the parameter P values were adjusted using the BH method (ref. ²⁹, Data File S2).

P-value aggregation and mutant statistics

Due to the high correlation between the different perturbation parameters, causing a lack of independence between the individual hypothesis tests (non-independent P values), we used an HMP technique to aggregate the P values and score the overall significance of the mutants while controlling for the FDR²⁸. The HMPs were computed as $\mathrm{HMP}$$={\sum }_j{w}_j/{\sum }_j\frac{w_j}{{P\mathrm{value}}_j}$, where weights (w_j) corresponded to the H₀ rejection rate at a 5% BH-correction cutoff, $w_j=F_j({\mathrm{FDR}}_{\mathrm{BH}}\le 0.05)$, to favour the more likely alternative hypotheses based on the genome-wide data. Here F_j(FDR_BH ≤ 0.05) is the fraction of mutants with BH-adjusted P value q_i,_j ≤ 0.05 for parameter j, that is, F_j(FDR_BH ≤ 0.05) = 1/N ${\sum }_i{l}_{\left(q_{i,j}\le 0.05\right)}$, where I is the indicator function, i indexes mutants, j indexes parameters, and N is the total number of mutants tested. This grouping of hypothesis tests maintained sensitivity of the least stringent BH-adjusted P (FDR minimum) or improved the sensitivity for mutants with multiple significance tests. We also computed HMPs based on equal weights but, in comparison, these did not improve the power over individual BH-FDR corrections per mutant (Fig. 1d, Supplementary Fig. 3b and ref. ²⁹, Data File S2).

Mutant profiling and clustering

Statistically significant mutants (HMP ≤ 0.01) were extracted for clustering. The parameter perturbations were clustered and grouped by computing the log₁₀-transformed P values and performing a hierarchical clustering using Pearson correlation as a similarity metric and ward.D2 clustering. The parameters were split into five groups (Supplementary Fig. 3d) using the cutreeDynamic function from the dynamicTreeCut package in R⁷⁵.

To group the mutants into distinct dynamic profiles, the response predictions from the double-sigmoidal models for mutants with HMP ≤ 0.01 were subjected to hierarchical clustering using Euclidean distances and ward.D2 (Supplementary Fig. 3e). For mutants with multiple significant independent experimental measurements, the median responses were computed before the analysis. Groups were identified using the cutreeDynamic function.

Screen quality assessments

Recall of reference mutants

Gold-standard test sets for evaluating the sensitivity and accuracy for retrieving true autophagy perturbation phenotypes were defined through manual inspection of images and literature curation (Supplementary Table 1)^{5,25,26,61,76–83}. This included two true positive test sets, including genes involved in autophagosome formation (ATG set, n = 16) and clearance (Fusion set, n = 19), respectively. A negative control set was also developed from several independent replicates of two dubious ORFs⁸⁴ that produced WT autophagy responses (n = 23; Supplementary Table 1). None of the mutants included in the test sets were part of the training data or used as reference data for the UMAPs.

Precision-recall and receiver-operating-characteristic (ROC) curves were computed from the P-value rank orders for the different perturbation parameters, comparing the true positive sets (ATG set, Fusion set or combined) with the negative control set (Extended Data Fig. 2d,e), and AUCs were computed to score the precision and accuracy of recalling true positives (Extended Data Fig. 2b,g). Alternatively, negative sets with equal size to the true positive sets were bootstrapped by repeated random sampling with replacement from the GW population performing b = 1,000 number of bootstraps for a more statistical assessment of the different perturbation parameters with respect to the average deletion mutant phenotype (Extended Data Fig. 2c,f). For mutants with multiple independent experimental measurements, the median P values or HMP were computed before the analysis. To use ROC curves for a direction-specific separation of the test sets from the GW population average, the P values were log-transformed and signed by the direction of the perturbation (Extended Data Fig. 2f,g).

Gene-set enrichment analysis

GSEA was performed using the gseGO function of the clusterProfiler package (version 4.4.4) and org.Sc.sgd.db (version 3.15.0) in R (version 4.2.1)⁸⁵, with 1 × 10⁵ permutations, and the minimum and maximum gene-set sizes set to 5 and 200, respectively. To analyse the directional association of GO-BP terms for the individual parameter perturbations, we used the log-transformed P values signed by the direction of the perturbation. For mutants with multiple independent experimental measurements, the median P values were computed before the analysis. Subsequently, enriched GO-BP terms with at least one test P < 0.005 were analysed across the multivariate parameter space using PCA (Fig. 2a,b) and hierarchical clustering (Supplementary Fig. 5a). In both cases, we used a matrix with the signed log-transformed enrichment P values for the selected GO-BP terms for each parameter (ref. ²⁹, Data File S3). Hierarchical clustering was performed using Euclidean distances and ward.D2.

To generate an enrichment map (Fig. 2c), we performed the same GSEA over the −log-transformed HMP values that were zero-centred around −log(0.01). For mutants with multiple independent experimental measurements, the median HMP was computed before the analysis. Here all GO categories—biological processes, molecular function and cellular compartment—were included. GO terms with P < 0.05 and NES > 0 were graphed based on the Jaccard correlation coefficient for scoring the GO-term similarities and a minimum Jaccard correlation coefficient > 0.01 for each edge, using the enrichplot (version 1.16.2) and clusterProfiler packages in R. The Spearman correlation over the profile count matrix for the GO terms was used to quantify the functional co-occurrence and purity of the different profiles in the enrichment map (Fig. 2d).

Network analyses

Spatial analysis of functional enrichment

The previously published by Costanzo and colleagues³⁰ was used to perform SAFE was performed on the SGA-PCC network (Fig. 3a). The SGA-PCC network was projected as in Costanzo et al. using the edge-weighted, spring embedded implementing Kamada–Kawai force-directed algorithm and loaded in Cytoscape (v.3.9.1)⁸⁶. We also prepared a STRING network, based on a combined_score ≥ 990 of the STRING (version 10) database, and projected using the unweighted spring embedded layout algorithm (Extended Data Fig. 5a and ref. ²⁹, Data File S4). Both networks were annotated with the GO-BP reference data from Costanzo et al. using SAFE^30,35. The SGA-PCC network displayed a coverage of 84% and an attribute coverage of 56%, and the STRING ≥ 990 network displayed a coverage of 99% and an attribute coverage of 66%. The enrichment landscape was built with a Jaccard correlation value of 0.75 and the resulting reference map labels were manually curated according to the functional domains identified. We also used SAFE to annotate the ATG core subgraph using the ATG set including ATG1 and ATG8, and the fusion core subgraph using the entire tethering complex (including VTI1 and YKT6), the HOPS complex, and VPS17, VPS29, VPS5, VPS35, VAM10, VPS1, VSP45, VPS3, VPS36 and VPS26. The profiles were enriched using SAFE and the enrichment colours were scaled with the SAFE score ranging from min = 0.35 to max = 0.35 (ref. ²⁹, Data File S4).

Network distance analyses

The shortest network path from any gene to the ATG core machinery (ATG set) was computed by finding the set of genes one degree away from any gene in the ATG set and iteratively finding the set of genes one degree away from any gene within the previous set until all possible paths had been found (Fig. 3b). For the shortest causal network path, connector genes within the previous set were required to have a significant phenotype given by HMP < 0.01 or an autophagy perturbation P_min < 0.01.

Gene-pair correlations

Correlations in phenotypes between genes were computed using the Pearson correlations of autophagy perturbation dynamics, $\mathrm{P}\mathrm{C}\mathrm{C}\mathrm{\Delta },\{m1,m2\}=\mathrm{c}\mathrm{o}\mathrm{r}({\mathrm{\Delta }}_{m1,t}\mathrm{\%},{\mathrm{\Delta }}_{m2,t}\mathrm{\%})$, or equivalently using signed −log₁₀-transformed P values over the set of kinetic-parameter hypothesis tests performed. Here m1 and m2 indicate a compared mutant gene pair. The correlations were subsequently grouped based on network interaction or complexome association status of the gene pairs as well as significance status (HMP ≤ 1, 0.01 or 0.001).

For strongly significant gene pairs (HMP ≤ 0.001), the correlations in perturbation dynamics were stratified into intervals for SGA-PCC concordance testing. The three intervals considered positive correlations (PCC_Δ > 0.25) representing similar phenotypes, negative correlations (PCC_Δ < −0.25) representing opposite phenotypes and no correlations. Enrichment of SGA-PCCs was computed for each interval given different positive or negative SGA-PCC thresholds (0.100, 0.125, 0.150, 0.175 and 0.200). Enrichment was calculated as $E_{k,l}$$=\frac{N({\mathrm{PCC}}_{\mathrm{\Delta }}=k,\mathrm{SGA}\cdot \mathrm{PCC}>l)}{N({\mathrm{PCC}}_{\mathrm{\Delta }}=k)}\times \frac{N}{N(\mathrm{SGA}\cdot \mathrm{PCC}>l)}$, for a PCC_Δ interval $k=(\mathrm{lower},\mathrm{upper}]$ and an SGA-PCC threshold $l$ (here indicated as greater than, but lesser than for $-l$). Here N is the total number of compared gene pairs and N(⋅) is a function that gives the number of compared gene pairs that satisfy conditions specified within the function. A Fisher’s exact test was used to compute the P values for over- or under-representation.

Directional GSEA along SGA-PCC spectra

Directional, positive or negative associations of autophagy perturbation profiles with SGA-PCC spectra for specific core machinery genes were computed using a modified GSEA algorithm (Supplementary Fig. 6c). A positive or negative enrichment score (ES₊ and ES₋, respectively) was computed as:

$${\mathrm{E}\mathrm{S}}_{+}=\max \left({F_c}^{\mathrm{GS}}-{F_c}^{\overline{\mathrm{GS}}}\right)\mathrm{a}\mathrm{n}\mathrm{d}{\mathrm{E}\mathrm{S}}_{-}=\min \left({F_c}^{\mathrm{GS}}-{F_c}^{\overline{\mathrm{GS}}}\right).$$

Here the $F$ represent the cumulative distribution functions for the ranked SGA-PCC scores (x)

$${F_c}^{\mathrm{GS}}=\sum _{i=1}^c\mathrm{abs}\left(x_i\right)1_{\left\{m_i\in \mathrm{GS}\right\}}/\sum _{i=1}^N\mathrm{abs}\left(x_i\right)1_{\left\{m_i\in \mathrm{GS}\right\}}\mathrm{and}{F_c}^{\overline{\mathrm{GS}}}=\sum _{i=1}^c{1}_{\left\{m_i\in \overline{\mathrm{GS}}\right\}}/\sum _{i=1}^N{1}_{\left\{m_i\in \overline{\mathrm{GS}}\right\}},$$

where 1 represent an indicator function for whether a mutant $m_i$ is present in the perturbation profile gene set (GS) or among the non-gene set mutants ($\overline{\mathrm{GS}}$). Direction preferential enrichment was computed as $\mathrm{\Delta }\mathrm{ES}={\mathrm{ES}}_{+}+{\mathrm{ES}}_{-}$. For statistical evaluation, we used a permutation test where the distribution of GS membership along x were randomized 500 times to generate respective H₀ distributions of ${r\mathrm{ES}}_{+}$, ${r\mathrm{ES}}_{-}$ and $\mathrm{\Delta }r\mathrm{ES}$, which were used to compute the respective P values, as well as the normalized enrichment scores ${\mathrm{NES}}_{+}=\frac{{\mathrm{ES}}_{+}}{⟨r{\mathrm{ES}}_{+}⟩}$, ${\mathrm{NES}}_{-}=\frac{{\mathrm{ES}}_{-}}{⟨r{\mathrm{ES}}_{-}⟩}$, and $\mathrm{\Delta }\mathrm{NES}={\mathrm{NES}}_{+}-{\mathrm{NES}}_{-}$. The NES values were grouped for genes belonging to the ATG core and the fusion core machinery respectively, and profile enrichment per group was tested using a one-sample Student’s t-test.

Module networks

The module networks were generated by identifying complexes through the Complex Portal³⁶ or applying the MCODE algorithm⁸⁷ to the STRING network (combined_score ≥ 990)³², with multiple significant genes. These complexes were then manually curated for PPIs using cross-references from SGD, STRING and BioGRID databases^31,32,88. The networks were embedded in Cytoscape using the prefuse force-directed open layout algorithm and subsequently manually organized with node colours scaling with the signed −log₁₀(P value) for the overall perturbation and size of the nodes scaling with their HMP (−log₁₀) (ref. ²⁹, Data File S5).

Latent-space and Bayes factors analysis

Statistical analysis of the UMAP latent space

Parametric UMAP predictions were performed on the latent autophagy feature vectors for all cells. The embedded cell densities were plotted by grouping populations per time points and colour coding either for time or predicted autophagy percentage ($A_t\%$). Moreover, the population means (${\mu }_{U1,t},{\mu }_{U2,t}$) and standard deviations (${\sigma }_{U1,t},{\sigma }_{U2,t}$) were computed for each UMAP coordinate ($x_{U1}$ and $x_{U2}$) per time point.

The UMAP embeddings were analysed for representation of predictive information, population dynamics and noise. To test whether the single-cell UMAP embeddings conserved the information about autophagy from the DNNs, we trained a bag of regression trees to predict the DNN outputs, either represented as the probability of autophagy (P_a) or its log-odds, $\mathrm{LO}=\log \frac{P_{\mathrm{a}}}{(1-P_{\mathrm{a}})}$. For this we used the XGBoost package (version 1.7.5.1) in R to train 50 parallel trees with subsampling = 0.63, depth = 10 and min_child_weight = 3 (ref. ⁸⁹). The regression was performed per plate, where the data were split to 70% training data and 30% test data, and a PCC between the predicted and true values was reported. For comparison, the same procedure was used to test representation of time. We also compared these results with the ability of the UMAPs to separate WT from vam6Δ cells by using bags of classification trees (with the same hyperparameter settings) subjected to tenfold training and testing over the entire dataset, and scoring the concordance between the predicted class and the true label using an F1 score. To assess whether the information distinguishing these phenotypes was due to latent information about intermediate autophagic states in the UMAP or simply due to distributional differences based on the overall DNN prediction for the two strains, we stratified the UMAP distributions into percentiles based on the predicted autophagy $P_1$ or LO, and then performed the same training and testing procedure within each interval.

To measure UMAP latent-space noise and flux, we first calculated the population bivariate s.d.

$${\sigma }_{U,t}=\sqrt{{{\sigma }^2}_{U1,t}+{{\sigma }^2}_{U2,t}}$$

and the average population displacement

$${\mathrm{\Delta }}_{U,t}=\sqrt{{{\mathrm{\Delta }}_{U1,t}}^2+{{\mathrm{\Delta }}_{U2,t}}^2}$$

for each time point. Here the population displacement per coordinate, for instance, ${\mathrm{\Delta }}_{U1,t}$, is given as ${\mathrm{\Delta }}_{U1,t}={\mu }_{U1,t+1}-{\mu }_{U1,t}$. We compared ${\mathrm{\Delta }}_{U,t}$ and ${\sigma }_{U,t}$ with equivalent metrics computed from the unembedded latent feature vectors directly for populations of more than 100 cells (Extended Data Fig. 6d). For this comparison, outliers for the latent-space metrics beyond the range given by the quartiles Q1 and Q3 ±1.5× the IQR were filtered out.

As a measure of signal-to-noise we also computed the normalized flux:

$${n\mathrm{\Delta }}_{U,t}=\sqrt{{\left(\frac{{\mathrm{\Delta }}_{U1,t}}{{\sigma }_{U1,t}}\right)}^2+{\left(\frac{{\mathrm{\Delta }}_{U2,t}}{{\sigma }_{U2,t}}\right)}^2}$$

The overall population variability and flux was assessed by averaging ${\mathrm{\Delta }}_{U,t}$, ${n\mathrm{\Delta }}_{U,t}$ and ${\sigma }_{U,t}$ for each mutant time series per treatment phase. For this comparison only observations with an average of 50 cells were evaluated (ref. ²⁹, Data File S6).

Bayes factor computation and assessment

Reference UMAP states of WT, atg1Δ and vam6Δ cell populations were collected from 21 independent time series across seven plates (Mix8, Mix10 and Rec1–5), where the WT cells comprised the dubious ORF-negative controls in order for the latent-space distributions to match that of unperturbed H₀ outcomes of the KO collection. Kernel density estimators (KDE) were fitted using the kde function of the Kernel Smoothing (ks) package (version 1.13.2) in R (version 4.2.1)⁹⁰. The KDEs were computed for the reference states based on their overall distributions independent of treatment phase ($K_{\mathrm{WT}}$, $K_{\mathit{atg}1\mathrm{\Delta }}$, $K_{\mathit{vam}6\mathrm{\Delta }}$) or their distributions stratified per time point ($K_{\mathrm{WT},t}$, $K_{\mathit{atg}1\mathrm{\Delta },t}$ and $K_{\mathit{vam}6\mathrm{\Delta },t}$) to perform time-wise phenotype comparisons. Using the dkde function, we computed the probabilities $P\left(D|K_S\right)$, which represented the marginal likelihoods for UMAP observations $D$ (U1, U2) given a reference state $S$. Thus, ratios of the marginal likelihoods comparing any two given states represented BFs, $\mathrm{BF}(\mathrm{WT}:\mathit{atg}1\mathrm{\Delta })$, $\mathrm{BF}(\mathrm{WT}:\mathit{vam}6\mathrm{\Delta })$ and $\mathrm{BF}(\mathit{vam}6\mathrm{\Delta }:\mathit{atg}1\mathrm{\Delta })$ that summarized the evidence favouring a given autophagic reference state in a specific UMAP (Extended Data Fig. 7a,c). For time-independent phenotype classification of mutants between reference states $A$ and $B$, the time-wise KDEs were paired with data for the corresponding time points to compute time-wise likelihoods. The individual KDEs were computed with equal matrix bandwidths generated by taking the medians of the bandwidth selection results using the Hpi.diag function on each time point. Subsequently, the log-likelihood ratios were averaged per mutant over all cells and all times points:

$$\mathrm{logBF}\left(B:A\right)=\frac{1}{n}\sum _{i=1}^n\sum _{t=1}^T\log \frac{P\left(D_{i,t}\mid K_{B,t}\right)}{P\left(D_{i,t}\mid K_{A,t}\right)}.$$

This population average or expected log(BF) can be viewed as a time-independent overall phenotype comparison per mutant, where dynamical effects have been ‘integrated’ out with the moving KDEs. This averaging was done for the entire experimental time course (overall) as well as for each treatment phase individually. On the other hand, to quantify the execution ‘activity’ as a time-wise transition between temporally fixed reference states, the expected log-likelihood ratio was computed with the time-invariant KDEs over a cell population per time point:

$$\log {\mathrm{BF}}_t\left(B:A\right)=\frac{1}{n_t}\sum _{i=1}^{n_t}\log \frac{P\left(D_{i,t}\mid K_B\right)}{P\left(D_{i,t}\mid K_A\right)}.$$

Within this BF framework, given the execution model $\{A\to B\to C\}$ ($\mathit{ATG}1\to \mathit{VAM}6$: autophagosome formation; $VAM6\to \mathrm{WT}$: autophagosome clearance), the overall activity for the transition $\{A\to C\}$ can be factorized into the intermediate transitions ($\{A\to B\}$ and $\{B\to C\}$) as

$$\log {\mathrm{BF}}_t\left(C:A\right)=\log {\mathrm{BF}}_t\left(C:B\right)+\log {\mathrm{BF}}_t\left(B:A\right).$$

For comparison with the time-independent BFs, and to score the expected activity over a given time frame, these time-wise BFs were averaged per mutant for all cells over a given time frame as well. This was done for the entire experimental time course (overall) as well as for each treatment phase individually (ref. ²⁹, Data File S7). Because these BFs represent comparisons between fixed reference states, they are more sensitive to dynamical variation among the mutants and hence, indicated with the $t$ subscript. All time-averaged BFs were corrected for distributional biases across plates by subtracting the plate-wise median difference from the screen median:

$$\log \mathrm{BF}(B:A{)}_m=\log \mathrm{BF}(B:A{)}_m-\left[\log \mathrm{BF}\right(B:A{)}_{\mathrm{plate}}-\log \mathrm{BF}(B:A{)}_{\mathrm{screen}}].$$

Analysis was restricted to mutants with at least n = 100 cells for time-averaged BFs overall and n = 50 cells when averaging per treatment phase. For mutants with multiple independent experimental measurements, the median BFs were computed. ROC curves and AUCs for analysing retrieval of test gene sets (ATG set and Fusion set) were conducted as described above. Potential overfitting of the density kernels to the reference distributions was assessed by comparing the Bayes factors between atg1Δ and the ATG set, and between vam6Δ and the Fusion set. To compare the performance (ROC analysis) of the BFs directly with the predicted autophagy perturbations (not subjected to sigmoidal curve fitting),

$$\mathrm{\Delta }{A}_{m,t}\%=A_{m,t}\%-A_{\mathrm{ctr},t}\%$$

and

$${\mathrm{\Delta }A}_{m,t}^{\mathrm{binary}}\%=A_{m,t}^{\mathrm{binary}}\%-A_{\mathrm{ctr},t}^{\mathrm{binary}}\%$$

were computed per mutant and averaged over the entire time series. Comparison with the fluorescent signal co-occurrence per experimental measurement was done by computing the mean co-occurrence early (0−1 h) and late (>7 h) in the starvation protocol (Supplementary Fig. 7a). In Extended Data Fig. 9e, the overlap coefficient was computed by dividing the cell’s average co-occurrence by the average signal intensity of each channel multiplied ($C=\frac{{\mu }_{\mathit{GxR}}}{{\mu }_G{\mu }_R}$), and computing the average across each cell population and time course. Functional analyses of BFs were performed over the KO library only. We removed BFs for experimental measurements where the mean red or green fluorescent signals early (0−1 h of starvation) were below the 25th percentile—1.5× the IQR of the GW screen distributions (ref. ²⁹, Data File S7). For functional analyses, we also computed the mean of the Bayes factor from the two neural latent-space models (model 22 and model 30). To assess the concordance in phenotype classification between treatment phases, we paired the treatment phase-specific BFs and computed the fraction of deletion mutants with substantial evidence for the same reference state (either logBF > 5 or logBF < −5) across both phases (Supplementary Fig. 7c). Analysis of lagged correlations for time-wise BFs ($\log {\mathrm{BF}}_t$) were computed using the KO library only.

In Fig. 5b,c and Supplementary Fig. 7e, where DAmP mutants were included, a harmonization procedure was performed where the BFs were adjusted to the global mean and variance of the two libraries using the time-invariant BFs:

$$\log \mathrm{BF}{\left(B:A\right)}_{m,\mathrm{adj}}=\left(\log \mathrm{BF}{\left(B:A\right)}_{m,\mathrm{library}\mathrm{type}}+{\mathrm{\Delta }}_{\mathrm{library}\mathrm{type}}\right)R_{\mathrm{library}\mathrm{type}},$$

where m indicates the mutant, ${\mathrm{\Delta }}_{\mathrm{type}}={\mu }_{\log \mathrm{BF}\left(B:A\right),\mathrm{screen}}-{\mu }_{\log \mathrm{BF}\left(B:A\right),\mathrm{library}\mathrm{type}}$ is the difference in screen average from the library specific average (either KO or DAmP) and $R_{\mathrm{type}}={\sigma }_{\log \mathrm{BF}\left(B:A\right),\mathrm{screen}}/{\sigma }_{\log \mathrm{BF}\left(B:A\right),\mathrm{library}\mathrm{type}}$ is the ratio of standard deviations for the screen over the specific library (either KO or DAmP).

Parametric gene-set enrichment analysis

Parametric GO (BP and CC) enrichment analyses of BFs were performed using the piano package (version 2.8.0) in R (version 4.2.1)⁹¹, using the gene sampling permutation test (nPerm = 1 × 10⁵) of the mean gene-set statistics. The gene sets were retrieved and annotated using the GOdb package (version 2.9) and org.Sc.sgd.db (version 3.15.0). Only the KO library phenotypes were analysed, including the median of all atg1Δ and vam6Δ phenotypes observed. The BFs or their respective treatment phase differentials were standardized over the GW population (to compute log-transformed BF z-scores) before the analyses (ref. ²⁹, Data File S8).

For the analysis of BF_t enrichment, the directional statistics (representing the average log-transformed BF_t z-score per gene set) were visualized for gene sets with P < 0.05 for at least one BF. To analyse BF differential enrichment, the directional statistics were visualized for gene sets with a differential enrichment P < 0.05. For the heat map, we selected the gene sets with a differential enrichment P < 0.01 for BF VAM6:ATG1 or BF WT:VAM6 and a minimum absolute statistic of one. Euclidean distance and ward.D2 were used for the clustering.

Experimental validation of autophagy predictions

Yeast and growth condition details

S. cerevisiae strains were cultured at 30 °C until mid-logarithmic phase in standard YPD medium or in synthetic medium supplemented with the relevant amino acids (SC) before proceeding with genetic manipulation and experimental assays. The strains used in this section are listed in ref. ²⁹, Data File S1.

Reproducibility library and strains construction for autophagy assays

All strains used in this section were derived from the S288c strain BY4741 (ref. ⁹²) using either PCR-targeted gene-replacement methods or crossing. Standard cell culture and genetic techniques were employed according to previously published protocols^93,94. Unless otherwise noted, all genetic modifications were carried out chromosomally. Deletions and promoter replacements were performed as described previously⁶⁷. For creation of the ‘reproducibility’ set, Atg8 was first N-terminally tagged by inserting mNG under control of the ScTDH3 (GDP) promoter into the 5′ end of ATG8. Pep4 was C-terminally tagged with mCherry by inserting the mCherry coding sequence into the 3′ end of PEP4 using the plasmid pBS35. Finally, all deletion strains and ORF control YEL074W were deleted using the HIS3MX6 selection marker amplified from plasmid pFA6a-his3MX6 (ref. ²⁹, Data File S1).

For the quantitative Pho8Δ60 assay in Fig. 5e, the 60 N-terminal amino-acid residues of the phosphatase Pho8 were removed by deleting 180 nucleotides from the 5′ end of PHO8 using the plasmid PYM-N15, rendering expression of the truncated ORF under the control of the ScTDH3 (GPD) promoter (ref. ²⁹, Data File S1)^95,96.

For the GFP–Atg8 processing assay in Figs. 5f, 6g and Extended Data Fig. 8c, yeast cells were transformed with a plasmid expressing GFP–Atg8 under the control of the endogenous ATG8 promoter (pRS416 GFP–ATG8/AUT7; ref. ²⁹, Data File S1)^97,98. All strains were confirmed by PCR and fusion proteins were additionally validated by fluorescence microscopy and western blot analysis. The primer sequences are provided in ref. ²⁹, Data File S1.

For the pr-Ape1 processing assay in Fig. 6k, cells were transformed with a CUP promoter-driven Ape1 plasmid (pCK782, a gift from C. Kraft).

Growth and starvation protocols for screen reproducibility analysis

The screen reproducibility analysis comprised two approaches as described in the next section. In the first, aimed at replicating starvation observations in the initial screening, strains were cultured sequentially to logarithmic phase, following an identical protocol as in the initial screening (‘Cell growth and starvation protocol for high-content screening’ section).

In the second approach, focused on evaluating the uniformity of autophagy phenotypes across various experiments and strain backgrounds, the growth protocol was adjusted to induce autophagy with cells in stationary phase. Cells were cultured in SC at 25 °C for three consecutive periods of 48 h each, followed by 8 h at 25 °C in the imaging plates before being transferred to SD-N to induce nitrogen starvation for 12 h and replenishment for 7 h. Live imaging was performed at 30 °C; images were captured every hour for 19 h.

Screen reproducibility analyses

Nearest-neighbour analysis and repetition set selection

Autophagy perturbation gene sets of different sizes were generated by setting various HMP cutoffs (≤0.01, 0.005, 0.001 or 0.0005; Extended Data Fig. 3b). Subsequently, we performed a Fisher’s exact test for over-representation of autophagy perturbations among the nearest neighbours using different GW yeast network data, including BioGrid PPIs, SGA-PCCs (top 1%) as well as all STRING interactions or only STRING-PPIs with a combined score of >500. These over-representation tests resulted in eight P values from which a harmonic mean was computed to yield an overall network enrichment significance (network P value). Potential FPs were defined as mutants with significant autophagy perturbations (perturbation FDR_min < 0.01; min referring to the lowest value between −N and +N) but non-significant network enrichment (network P > 0.1). Potential false negatives (FNs) were defined as mutants with non-significant autophagy perturbations (perturbation FDR_min > 0.1) and significant network enrichment (network P < 0.01). Potential FNs were also selected based on genes with key autophagy GO-term annotations (Extended Data Fig. 3c) and non-significant phenotype (FDR_min > 0.1 and HMP > 0.01). Top hits were selected by rank-ordering mutants based on the HMP (ref. ²⁹, Data Files S9 and S10).

Replication of repetition set and outliers

Only mutants from the KO library were used for the validation screen due to the lack of WT DAmP controls. This included 109 potential FNs, 56 potential FPs and 75 top hits (ref. ²⁹, Data Files S1 and S10). For the validation screen perturbations in predicted autophagy ($\mathrm{\Delta }{\mathrm{\%}}_{m,t}$) were computed from the mean WT ORFs included in each plate (as the plate medians would not be valid WT references due to the enrichment of mutant phenotypes). The replicability was tested by repeating the GW screen protocol up to 8 h of starvation and computing the correlation in predicted autophagy dynamics (${\mathrm{PCC}}_{A,m}=\mathrm{cor}\left(A_{m,t,R1}\%,A_{m,t,R2}\%\right)$) and autophagy perturbation dynamics (${\mathrm{PCC}}_{\mathrm{\Delta }A,m}=\mathrm{cor}\left({\mathrm{\Delta }A}_{m,t,R1}\%,{\mathrm{\Delta }A}_{m,t,R2}\%\right)$) for each mutant between replicates $R$. The dynamics correlations were compared with the mean absolute autophagy perturbations, $\mid \mathrm{\Delta }A\mathrm{\%}{\mid }_m=\frac{1}{T}{\sum }_{t=1}^T\mathrm{abs}\left(A_{m,t}\%-A_{\mathrm{ctr},t}\%\right)$. For mutants with multiple independent experimental measurements in the GW screen, the median responses were computed before the comparisons.

Reproducibility across experiments, strain background and conditions

To assess the reproducibility of autophagy phenotypes across strain background, independently generated clones and growth conditions, 33 mutants covering different cellular functions and with different autophagy dynamic profiles were picked manually. All mutants were strongly significant (HMP ≤ 0.001). Three independently generated new clones were included for each mutant, along with the collection clone, and three new screens were done independently (ref. ²⁹, Data File S10 and Extended Data Fig. 4a,b).

For evaluation of mutant replication error, the mean absolute perturbations were computed per mutant and compared with the mean absolute errors, ${ϵ}_{A,m}=\frac{1}{T}{\sum }_{t=1}^T\mathrm{abs}\left(A_{m,t,R1}\%-A_{m,t,R2}\%\right),$ between different types of replicates $R$. These included technical replicates (same mutant clone, different wells in the same plate), experimental replicates (same mutant clone, different plates/experiments), replication of mutant phenotypes generated in a new background strain and replication between independent clones generated in the same background strain. Correlations in autophagy dynamics (${\mathrm{PCC}}_A$) and autophagy perturbation dynamics (${\mathrm{PCC}}_{\mathrm{\Delta }}$) were also computed for each mutant across experimental replicates, and between collection clones and clones generated in a new background strain. Mutant perturbations in autophagy on the new screen plates were computed with respect to the average from 12 independent WT controls, six based on the collection background strain and six based on the new background strain.

For assessing screen-level experimental reproducibility of autophagy perturbations and kinetic parameters, Pearson and Spearman correlations (due to large outliers for some parameters) were computed from global pairwise comparisons of the screen metrics across all mutant clones. Comparisons were done between the GW screen and the new screens for autophagy perturbations, ${\mathrm{\Delta }\mathrm{\%}}_t$, grouped on treatment phase (−N, +N or both) and strain background (collection clones, new clones or both) or between the new screens across all clones for the parameters (ref. ²⁹, Data File S10).

Rapamycin-dependent autophagy response and analysis

Experimental cell growth and conditions

Cells were cultured in SC at 25 °C for two consecutive periods of 48 h each, followed by a period of 24 h and a last of 8 h at 25 °C in the imaging plates. The cells were then washed twice with sterile MilliQ water and transferred to independent plates containing 200 µl SC + vehicle, SC + rapamycin, SD-N + vehicle and SD-N + rapamycin. Live imaging was performed at 30 °C; images were captured every hour for 10 h. A solution of 1 mg ml⁻¹ rapamycin dissolved in 9:1 (vol/vol) ethanol:Triton X-100. Cells were treated with rapamycin at a final concentration of 440 nM. As a vehicle control, the rapamycin solvent was added in the same proportions to cells. Each plate included duplicate experimental conditions of every strain (triplicates for the controls), and the complete experiment was repeated three times.

Data analysis

Autophagy ($A_t\%$) and UMAP BFs (BF_t) were computed for each well (ref. ²⁹, Data File S11). Autophagy perturbations ($\mathrm{\Delta }{\%}_t$) were computed with respect to the mean of the ORF plate controls; the time series averages were computed for all metrics. Reproducibility between experiments was assessed by computing the PCC for each metric.

Yeast growth conditions for protein and RNA analysis

To correspond to library growth conditions, cells were serially diluted in fresh medium to maintain them in logarithmic phase before inducing autophagy.

For protein analysis in the GFP cleavage assay, the cells were first cultured overnight in SC-URA and then transferred to fresh SC-URA (6.9 g yeast nitrogen base with ammonium sulfate and without amino acids, 20 g per glucose anhydrous, and 0.770 g single drop-out -ura (Formedium) per litre) and cultured to mid-logarithmic phase in three sequential cultures at 30 °C for 24 h. The cells were then washed five times with SD-N medium with membrane filtration and transferred to SD-N at an optical density at 600 nm (OD₆₀₀) of 0.25, followed by incubation for 4 or 8 h at 30 °C, depending on the experiment. In the experiments with prolonged starvation (8 h), the cells were transferred to SC-URA medium and incubated for a further 4 h. The cells were collected at specified time points during nitrogen starvation and replenishment, and subjected to protein analysis. For protein analysis in the assay of the precursor form of Ape1 (pr-Ape1) processing, cells were cultured in SC-LEU (6.9 g yeast nitrogen base with ammonium sulfate and without amino acids, 20 g per glucose anhydrous, and 0.690 g single drop-out -leu (Formedium) per litre) at 30 °C to mid-logarithmic phase for 24 h. The cultures were then split into two conditions: one supplemented with 50 µM CuSO₄ to induce Ape1 overexpression for 20 h, and a control condition without CuSO₄. To induce starvation, the cells were washed five times with SD-N with membrane filtration and resuspended in SD-N at an OD₆₀₀ = 0.25. The cells were then incubated at 30 °C for 4 h in SD-N, and samples were collected at the indicated time points according to the experiment.

For RNA analysis, cells were sequentially grown in SC at 30 °C to mid-logarithmic phase, transferred to SD-N, collected at specified time points during nitrogen starvation and replenishment, and subjected to protein analysis. Samples at t₀ were collected from a portion of the culture that was transferred to fresh SC once more and incubated at 30 °C for an additional hour.

GFP–Atg8 and pr-Ape1 processing assay

The GFP–Atg8 and pr-Ape1 processing assays were performed and interpreted as described previously^66,98,99. According to each experiment, cells were collected at the indicated times and protein extracts were prepared and resolved by SDS–PAGE. Full-length GFP–Atg8 and free GFP were detected by western blotting using an anti-GFP and pr-Ape1 processing using a yeast anti-Ape1.

Protein extraction, immunoblotting and antibodies

Cell lysis and immunoblotting were performed as previously described¹⁰⁰. For protein analysis, 2 × 10⁸ cells were collected by filtration, washed with STOP buffer (150 mM NaCl, 50 mM NaF, 10 mM EDTA and 1 mM sodium azide) and frozen on dry ice before extraction. Total protein extracts were prepared by precipitation with trichloroacetic acid as previously described¹⁰¹. Following gel electrophoresis in 4–20% Criterion TGX precast midi gels (BioRad), the proteins were transferred onto a polyvinylidene fluoride membrane using a semi-dry blotting system (Trans-Blot Turbo, BioRad). The membrane was blocked in 5% skimmed milk with TBS–0.1% Tween and incubated with primary antibodies in 5% BSA with TBS–0.1% Tween (for antibody to phosphorylated residues) or 5% milk with TBS–0.1% Tween (all other antibodies). The western blots were developed using Pierce and SuperSignal Dura ECL detection reagents (Thermo Fisher) and imaged using a Chemidoc MP Imaging system (BioRad Laboratories). The following primary antibodies were used in this study: rabbit anti-phospho-(Ser/Thr) Akt Substrate Antibody to detect phosphorylated RpS6 (1:1,000; Cell Signaling Technology, 9611), goat anti-GFP (1:1,000; Abcam, 5440) to detect GFP–Atg8, rabbit yeast anti-Ape1 (1:10,000; a gift from C. Kraft)⁵⁶ and mouse anti-β-actin (1:1,000; Abcam, 8224). The following secondary antibodies were used: horseradish peroxidase (HRP)-conjugated goat anti-rabbit IgG (1:5,000; Sigma-Aldrich), HRP-conjugated bovine anti-goat IgG (1:5,000; JIR) and HRP-conjugated goat anti-mouse IgG (1:10,000; JIR). The western blot experiments were performed in replicates as indicated in ref. ²⁹, Data File S12 and representative images have been provided.

Quantifications of western blots

Degradation of GFP-fused Atg8 in the vacuole, processing of pr-Ape1 in the vacuole and phosphorylation of RpS6 in relation to the β-actin loading control were estimated by calculating the densitometric values of the bands in the raw digital western blot images using the Image Lab software (version 6.0.1; BioRad Laboratories). The western blot experiments were performed for multiple independent replicates as indicated in ref. ²⁹, Data File S12. All quantifications and numerical transformations are in ref. ²⁹, Data File S12.

Autophagy flux by assessing GFP cleavage and RpS6-p comparison with BFs

Western blot quantifications were averaged over two exposure times for every experiment before computing the mean and s.e.m. from the independent biological replicates (ref. ²⁹, Data File S12). GFP cleavage was computed from the adjusted band volume (intensity) as the percentage of free GFP signal divided by the total GFP signal (free GFP + GFP–Atg8) in each lane using densitometric analysis (BioRad Image Lab Software). The relative RpS6-p levels were computed from the absolute band volume by dividing the RpS6-p signal by the β-actin signal and normalizing each experiment to their respective WT t₀ value.

The time-averaged BFs, kinetic parameters and average screen metrics were compared with the autophagic flux by computing the mean GFP cleavage from 8 h of starvation (0, 2, 4, 6 and 8 h). The similarity in autophagy phenotypes measured across mutants were assessed using Pearson’s correlation. For time-series comparison of BF_t with GFP cleavage, we matched the time-wise measurements for −N (0, 2, 4, 6 and 8 h) and +N (0, 1 and 4 h) for the two time series experiments (number of replicates per mutants and conditions in ref. ²⁹, Data File S12). Note that +N for the BF_t occurred from 12 h in −N, whereas for GFP cleavage +N occurred from 8 h in −N. Moreover, the 5 min time points in +N for GFP cleavage were used to match with t₀ replenishment for BF_t. The time-wise similarity was then assessed using a mixed model based on BF_t WT:ATG1 alone (model 1) or BF_t WT:VAM6 and BF_t VAM6:ATG1 (model 2) with random intercept and slope for each time point:

$$\mathrm{Model}1:y_{\mathrm{GFP}\mathrm{\%},t}={\beta }_0+b_{0,t}+\left({\beta }_{\mathrm{BF}{1}_t}+b_{\mathrm{BF}{1}_t,t}\right)x_{\mathrm{BF}{1}_t,t}+ϵ$$

$$\begin{array}{c}\mathrm{Model}2:y_{\mathrm{GFP}\mathrm{\%},t}={\beta }_0+b_{0,t}+\left({\beta }_{\mathrm{BF}{2}_t}+b_{\mathrm{BF}{2}_t,t}\right)x_{\mathrm{BF}{2}_t,t}\hfill \\ +\left({\beta }_{\mathrm{BF}{3}_t}+b_{\mathrm{BF}{3}_t,t}\right)x_{\mathrm{BF}{3}_t,t}+ϵ\hfill \end{array}$$

Here $\beta$ represent fixed effects, $b$ represent random effects per time point $t$, $ϵ$ is the residual error, and $x_{\mathrm{BF}{1}_t}$, $x_{\mathrm{BF}{2}_t}$ and $x_{\mathrm{BF}{3}_t}$ represent the log-transformed BF_t WT:ATG1, BF_t WT:VAM6 and BF_t VAM6:ATG1 values, respectively. Goodness-of-fit was assessed by measuring the Pearson correlations between predicted and observed values across mutants for each time point and across time points for each mutant. The time-wise variable (BF_t) contribution to autophagic flux point was computed as $({\beta }_{{\mathrm{BF}}_t}+b_{{\mathrm{BF}}_t,t}){\sigma }_{{\mathrm{B}\mathrm{F}}_t,t}$, where ${\sigma }_{{\mathrm{B}\mathrm{F}}_t,t}$ is the s.d. of the log-transformed BF_t sample space at time point $t$.

For comparisons of the BFs with the RpS6-p status, the time-averaged BF_ts were correlated with RpS6-p at t₀, average RpS6-p in −N (2, 4, 6 and 8 h) and average RpS6-p in +N (15 and 30 min, and 1, 2 and 4 h; number of replicates per mutant and condition in ref. ²⁹, Data File S12).

Statistical evaluation of genetic interactions effects on GFP cleavage between rtg1Δ and gcn4Δ or gln3Δ were done using a mixed-effects model with random intercept on experiment identity, rtg1Δ as the baseline and interaction between the mutants and time (ref. ²⁹, Data File S12).

Quantitative Pho8Δ60 assay

The Pho8Δ60 assay was performed as previously described^96,102. For the analysis of alkaline phosphatase, cells were first cultured in SC medium to mid-logarithmic phase (5 × 10⁶ cells ml⁻¹) at 30 °C, diluted to 2.5 × 10⁶ cells ml⁻¹ and transferred to SC for 1 h and SD-N for 4 h. For sampling, 1.00–1.25 × 10⁶ cells were collected by filtration, washed with 1 ml PBS pH 7.2 and the pellets were frozen on dry ice. For protein extraction, the cell pellets were dissolved in 500 µl lysis buffer (pH 6.8) containing 1 M PIPES, 10% Triton X-100, 1 M KCl, 1 M potassium acetate, 1 M MgS0₄, 10 mM ZnSO₄ and 2 mM phenylmethanesulfonyl fluoride. Cell disruption was performed by agitation and homogenization with a cryogenic bead beating grinder (FastPrep-24 5G). For the phosphatase assay, 0.1–0.5 mg total protein was used as the phosphatase source, with 125 mM p-nitrophenyl phosphate serving as the substrate in the reaction buffer. The reaction buffer was prepared with 1 M Tris–HCl pH 8.6, 10% Triton X-100, 1 M MgSO₄ and 10 mM ZnSO₄. The mixture was incubated at 37 °C for 5 and 7 min, allowing the phosphatase present in the total protein to dephosphorylate the p-nitrophenyl phosphate substrate. The enzymatic reaction was then stopped by the addition of 1 M glycine pH 11. The alkaline phosphatase activity was calculated as the amount of p-nitrophenol (nmol) produced per minute per milligram of total protein. Both incubation times were measured using independent reactions. The Pho8Δ60 assay experiments were performed for multiple independent replicates as indicated in ref. ²⁹, Data File S12. Data are presented as the mean ± s.e.m. Significant differences were calculated using a one-way analysis of variance with a Tukey’s honestly significant difference post-hoc test. Significance is reported for mutant comparisons with the negative control within each treatment phase (ref. ²⁹, Data File S12).

RNA extraction and quantitative PCR

RNA extraction and qPCR were performed as previously described¹⁰⁰. For qPCR analysis, 8 × 10⁷ cells were collected by centrifugation, washed in diethyl pyrocarbonate-treated water and frozen in dry ice. Total RNA preparation was performed using a MasterPure yeast RNA purification kit (Epicentre) following the manufacturer’s instructions. Complementary DNA synthesis was performed using 1 µg RNA and SuperScript III reverse transcriptase (Invitrogen). Quantitative PCR was performed using SYBR select master mix (Applied Biosystems). The primer sequences (forward and reverse) are listed in ref. ²⁹, Data File S1. The RNA analysis experiments were performed in triplicate as indicated in ref. ²⁹, Data File S13.

Quantitative PCR analysis

Three independent experiments were performed, each with two technical replicates, for all targets per treatment condition. The mean C_t value was calculated for each target per treatment condition and subsequently normalized by subtracting the median value of the actin control to compute a ΔC_t value, which was transformed to the signal $S=2^{-\mathrm{\Delta }{C}_{\mathrm{t}}}$. The complete dataset was batch-corrected using the ComBat algorithm from the sva package (version 3.44.0) in R¹⁰³, using parametric adjustments with experimental replicate as batch variable and genotype as the model variable. The $S$ value was log-transformed before the batch correction and then exponentiated to generate $S_{\mathrm{adj}}$. Finally, the $S_{\mathrm{adj}}$ value was normalized to the mean $S_{\mathrm{adj}}$ of the WT controls at t₀ and log₂-transformed (ref. ²⁹, Data File S13).

Significance for individual time points was assessed by comparing the mutants with the WT using a two-sided Student’s t-test. For statistical evaluation of the overall differential regulation effect by the mutants on every target, a regression was performed per treatment phase (t₀ +N, −N and +N), with the WT as baseline and the genotype and time point as independent variables. The reported values are the coefficients for the genotype effects (ref. ²⁹, Data File S13).

Live-cell imaging of autophagosome dynamics

Experimental cell growth and imaging

For quantification of autophagosome dynamics in WT, rtg1Δ, vam3Δ and DKO cells (Fig. 6h–j), cells were imaged every 15 min for 2 h, capturing 3–6 images per well. Experiments were conducted in three replicates, each with two independently generated clones harbouring the autophagy reporter GPD1pr-mNG–Atg8 and Pep4–mCherry. The cells were cultured sequentially following the protocol in ‘Cell growth and starvation protocol for high-content screening’. For time-lapse imaging, cells (4–6 µl) were inoculated into 96-well plates (Nunc Microwell, Thermo Fisher) containing 200 µl SC per well and incubated for 8 h at 25 °C to ensure the cells were in exponential phase. To induce autophagy, the cells (400 µl) were inoculated into ibidi µ-Slide eight-well untreated plates (Ibidi) coated with concanavalin A (0.25 mg ml⁻¹), washed twice with 400 µl SC medium and transferred to SD-N medium (400 µl) for 2 h. Live imaging was performed using a Nikon ECLIPSE Ti2-E inverted microscope equipped with a temperature control unit, CrestOptics X-Light V3 spinning disk confocal module and Lumencor Celesta multiline laser (Nikon Crest System). Images were acquired using automated acquisition in wide-field, utilizing a CFI Plan Apo l D ×100 oil numerical aperture 1.45 objective coupled to a 16-bit back-illuminated sCMO camera, and Z-drive and PFS autofocusing mechanisms. Image acquisition was controlled using the NIS Elements AR software (Nikon Instruments Inc.). Live-cell imaging was conducted at 30 ± 2 °C. Fluorescence signals were captured using specific filter sets for wide-field GFP and wide-field RFP, with exposure times of 30 ms and 100 ms, respectively, using z-sections acquired at 0.2-µm steps over a range of 0.8 µm. For bright-field images, an exposure time of 100 ms was used to capture a single focal plane.

Cell and vesicle segmentation and feature extraction

We performed image analysis using a custom FIJI/imageJ pipeline to quantify autophagy dynamics, including the formation of autophagosomes, represented by changes in the intensity and distribution of the green signal and the co-localization between the green and red signals. We categorized the images by position (xy), channel (BF, GFP and RFP), focal plane (z) and time point (t). We generated maximum-intensity z-projections for each position, channel and time point. We then assembled sequential time points for each channel into a time-lapse stack, followed by image-processing steps, including applying an unsharp mask filter (radius = 1.0, weight = 0.60) to enhance edges and performing background subtraction (rolling radius = 50 pixels) to reduce noise. For optimal visualization, we applied alignment corrections across time points using the SIFT algorithm (rigid, 2,048 pixels) for each channel and we merged the GFP and RFP time-series stacks into a single RGB composite image. For cell segmentation, we applied the ‘Huang Dark’ threshold to convert the non-aligned processed t-stacks into binary masks. We then used the Analyze Particles plugin with a minimum particle size of 20 pixels and a circularity range of 0.10–1.00 to extract measurements, including area, perimeter, centroid, mean, median, bounding box, Feret’s diameter, shape, integrated intensity, skewness, kurtosis, spatial coordinates and stack position. For autophagosome segmentation, we applied the ‘MaxEntropy’ threshold and used the Analyze Particles plugin with a minimum particle size of 1 pixel and a circularity range of 0–1.00, restricting the analysis to autophagosomes within the segmented cell mask. The same measurements used for cell segmentation were applied to vesicle tracking over time. Finally, we quantified co-localization by multiplying the pixel values of the GFP and RFP channels using the ‘Multiply create 32-bit’ command.

Data analysis

Cells were filtered for an area between 1,400 and 6,000 pixels and roundness > 0.7; vesicles were filtered for an area below 750 pixels. The size of the autophagosomes was reported by their area in pixels with one pixel unit corresponding to 70 nm.

Electron microscopy

For harvesting, cells were first grown in YPD medium and transferred to grow in three sequential cultures to mid-logarithmic phase in SC medium at 30 °C for 24 h. The cells were then washed five times with SD-N medium using membrane filtration and transferred to SD-N at OD₆₀₀ = 0.25. The cells were then incubated for 1 h, washed with 1 ml PBS pH 7.2 and resuspended in 30 μl of the remaining buffer. The cells were then immediately high-pressure frozen using a Leica HPM 100 system. Freeze substitution was performed as follows: cryo-vials were filled with freeze substituent (acetone containing 1% (wt/vol) osmium tetroxide, 0.1 % (wt/vol) uranyl acetate, 0.25 % glutaraldehyde and 1% water), and placed in a temperature control AFS2 system (Leica) and cooled to −90 °C. Sample carriers containing yeast cells were transferred into cold cryo-vials and left at −90 °C for 48 h before the temperature was raised to −45 °C over 9 h. The samples were kept at −45 °C for 5 h before being raised stepwise to −20 °C (within 1 h) and then +4 °C (within 30 min). The samples were removed from the AFS2 once the temperature had reached +4 °C, washed three times in acetone and then infiltrated with increasing concentrations of epon (25%, 50%, 75% and twice in 100%; 1 h each) mixed with acetone. Finally, the cells were infiltrated a third time with epon (100%) and left at room temperature overnight for infiltration. Polymerization was conducted at 60 °C for 72 h. Serial sections (50 nm thickness) were cut on an Ultracut UCT ultramicrotome (Leica), collected on formvar-coated slot grids and post-stained with Reynold’s lead citrate and 4% uranyl acetate solution. Electron micrographs of individual yeast cells were obtained on a Thermo Scientific TM TalosTM F200C microscope. Vacuolar inclusions that were not continuous with the cytoplasm were annotated manually and measured and analysed using the FIJI/ImageJ software. A one-way analysis of variance comparing WT (n = 93) and rtg1Δ (n = 91) inclusion body areas was concluded as non-significant.

Cross-omics inference for autophagy predictions

Cross-omics prediction benchmarking

Cross-omics modelling and prediction benchmarking were performed against the time-averaged BFs and kinetic-parameter statistics (signed −log₁₀-transformed P values) for the KO library only. Three omics datasets from a genome-wide deletion library were tested as well as the SGA-PCC dataset and a STRING protein similarity network (STRING-PCC). The individual datasets were matched for overlapping deletion mutants with autophagy phenotypes as the output and subsequently tested with random forest and tenfold cross-validation using all features as model inputs. The proteomics dataset was retrieved from Messner et al.⁵² and we used the complete imputed dataset. Each protein abundance value was normalized to the median value of the WT controls and subsequently log₂-transformed. The dataset overlapped with 4,189 deletion mutants and included 1,850 proteins as model features.

The transcriptomic dataset was retrieved from Kemmeren et al.⁵¹ and represented microarray-based log₂-transformed fold changes from the WT controls. The dataset overlapped with 1,419 deletion mutants and included 6,112 transcripts as model features. The expression profiles for each mutant were standardized to have a zero mean and unit variance.

The metabolomics dataset was retrieved from Mülleder et al.⁵³ and represented concentration changes in deletion mutants for 19 amino acids. The dataset overlapped with 4,501 deletion mutants.

The SGA-PCC was retrieved from Costanzo and colleagues³⁰. Because this dataset includes some correlation with multiple deletions for the same gene, we averaged the correlations per ORF over the columns and rows subsequently. The resulting dataset overlapped with 4,565 deletion mutants and included 5,707 gene correlations as model features.

The STRING-PCC dataset was retrieved from the STRING database version 10 (ref. ³²). A binary interactome matrix for combined_sore ≥ 200 was generated and subsequently used to compute the pairwise PCC between each column. The resulting correlation matrix overlapped with 4,684 deletion mutants and had 6,385 gene correlations as model features.

The response variables were standardized before training. Random forest was executed using the regression implementation of the XGBoost package (version 1.7.5.1) in R (version 4.2.1)⁸⁹ using squared error as loss and learning_rate = 1. The number of parallel trees was set to 1,000, subsample = 0.63 and colsample_bynode = $\sqrt{p}/p$, where $p$ indicated the number of features and varied per dataset. The number of splits were set to max_depth = 20 and we used a min_child_weight = 3. The Pearson correlation coefficient between observed and predicted values was computed for each test fold, and the median across folds was reported.

Random forest mediation analysis

Discovery of transcriptional mediators of regulatory effects of TFs on autophagy responses were executed by matching the time series of the Induction Dynamics Gene Expression Atlas (IDEA) by Hackett et al.⁵⁴ with the initial starvation responses dynamics of 186 overlapping TFs using a 1 h time shift. The rationale for using the IDEA dataset is that the gene-expression profiles presented in IDEA are directly causally driven by the acute activation of the TFs. Moreover, because of the abundance of autophagy perturbations caused by TF deletions and the relatively low variation in autophagy at the onset of the starvation protocol, it is probable that a lot of the dynamic variation is related to loss-of-functions of crucial gene-expression changes acutely involved in the starvation response. Gene-target induction profiles that strongly predict the autophagy perturbation responses generated across multiple TF deletions, where the targets themselves are genes with a measured influence on autophagy, represent probable candidates as targets for transcriptional modulation of autophagy.

In the IDEA datasets we used the ‘log2_cleaned_ratio’ variable for differential gene expression time series ($X$), which are normalized to the t₀ baseline of each induction protocol. To compensate for varied efficacies on gene-expression regulation from induction for different TFs, as well as batch effects, we performed the harmonization procedure $X_{\mathrm{TF},\mathrm{adj}}=X_{\mathrm{TF}}\cdot \left({\sigma }_X/{{\sigma }_X}_{\mathrm{TF}}\right)$, where ${\sigma }_X$ is the s.d. of $X$ for the entire screen and ${{\sigma }_X}_{\mathrm{TF}}$ is the standard deviation of $X_{\mathrm{TF}}$ for a specific TF. The reported log-transformed fold change (log(FC)) values in the study are based on these adjusted values. Missing values in the time series were imputed using the na_interpolation function with the stine method from the imputeTS package (version 3.3) in R¹⁰⁴, and time series values from 30, 45, 60 and 90 min after TF induction were extracted for the remaining analysis. These were matched with 90, 105, 120 and 150 min after starvation for $A_t\%$ and the ${\mathrm{BF}}_t$ time series, where the intermediate missing values were imputed using the na_interpolation function with the spline method. Thus, the resulting dataset had 744 samples (186 TFs × 4 time points). Before the regression, the response variables were centred around the median of each time point, to compute the perturbations, and then standardized. The dataset had 6,175 gene targets as features, 5,989 after the removal of the 186 TFs and 4,508 that were present in the autophagy dataset, where 1,354 targets also had a significant autophagy phenotype (HMP ≤ 0.01).

Random forest was executed as described above using the XGBoost package (version 1.7.5.1) in R (version 4.2.1), with the number of parallel trees set to 5,000. For model evaluation, tenfold cross-validation was performed where each fold was defined by withholding a random sample of individual data points or a random sample of TF time series. The r² was computed for each test fold and the median value was reported. To score the feature importance, we used the sumGain, which was computed using the importance function in the XGBoost package. For clustering, selected TFs with potential mediators, target genes with a sumGain > 2,200 against one of the response variables and a HMP < 0.001, were considered (ref. ²⁹, Data File S14). The sumGain cutoff was estimated as 2× the average sumGain of features from control models trained with randomized sample IDs. SHAP values were used to assess the predictive effect from gene-expression changes on autophagy perturbations. To compute SHAP values, we used the SHAPforxgboost package (version 0.1.3)¹⁰⁵. The individual SHAP values were grouped on whether the corresponding rfvalue (raw feature value; corresponding to the log-transformed fold changes in expression, $X_{\mathrm{TF}}.\mathrm{adj}$) were positive or negative and then computing the mean per group for each feature (ref. ²⁹, Data File S14). Thus, this yielded average SHAP values for upregulated and downregulated gene expression, respectively. Differential SHAP values were computed by subtracting the SHAP value for downregulation from the SHAP value for upregulation. To compare the SHAP values for models trained with different number of features, they were multiplied with $p/p_{\max }$, where $p$ indicates the number of features for the specific model and $p_{\max }$ indicates the maximum number of features.

Statistics and reproducibility

The numbers of biological and technical replicates are indicated in the figure legends, main text and methods. All measurements were taken from distinct samples. Statistical analyses included multiple hypothesis testing, Fisher’s exact test, χ² test, permutation-based GSEA, two-sided Student’s t-tests and one-sample Student’s t-tests, as indicated in figure captions and described in Methods. Standardized procedures including protocols, instruments, measurement settings and analysis workflows were applied to minimize bias and are reported across the Methods sections. Full statistical parameters are reported, including exact and adjusted P values for multiple hypothesis testing as well as slopes, interaction terms and regression coefficients for mixed models. Exact P values are reported in the figure legends. When there are more than five significant comparisons, the exact p values are provided in the source data associated with the figure. No data points were excluded from the analyses, except for outliers in signal for the genome-wide BF (Supplementary Fig. 7 and Methods). For Bayesian analyses and hierarchical or complex designs, appropriate statistical tests were selected and all outcomes are fully reported in the corresponding data files (10.5061/dryad.cfxpnvxdh). Additional covariates, including time and conditions, were tested.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41556-025-01837-0.

Author contributions

Study conception and funding acquisition: J.M.E. Conceptualization: N.C., A.N.A. and J.M.E. Methodology—construction of genome-wide libraries: I.G., A.N.P., P.A.-D. and C.D.P. Methodology—optimizing of cell growth and starvation protocols: N.C. Methodology—development and optimization of high-content screening protocols: N.C., S.O.-M. and A.N.P. Methodology—construction of validation libraries: N.C., S.O.-M. and L.H. Methodology—image segmentation and feature extraction analysis: N.C. and I.G. Methodology—development and optimization of deep-learning pipeline: A.N.A. and I.G. Investigation—performance and validation of high-content screening: N.C. and L.H. Investigation—optimization of autophagy assays: N.C., L.H., S.O.-M., A.N.P. and E.R. Investigation—performance of RTG, TOR and SPE experiments: N.C., L.H. and E.R. Investigation—electron microscopy and analysis: S.W.S. Data curation and analysis: N.C. and A.N.A. Formal analysis and statistics: A.N.A. Network analysis: N.C. and A.N.A. Visualization: N.C. and A.N.A. Web portal: S.N. and A.N.A. Writing—original draft: N.C. and A.N.A. Writing—review and editing: N.C., A.N.A., C.D.P., M.Z., T.E.R. and J.M.E. Project coordination: N.C. and J.M.E. Supervision: N.C., M.Z., T.E.R. and J.M.E. All authors discussed, revised and approved the paper.

Peer review

Peer review information

Nature Cell Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available.

Data availability

Genome-wide profiling repository: all mutant profile autophagy responses are publicly available on the AutoDRY web portal: https://cancell-apps.medisin.uio.no/AutoDRY/. High-content Image availability: all high-content images will be publicly available in the BioImage Archive at EMBL-EBI (https://www.ebi.ac.uk/bioimage-archive/) under accession number S-BIAD2338. Additional information can be obtained from N.C. or J.M.E. Data Files S1–S18, including files for libraries, strains and primers as well as the source data supporting the main, extended and supplementary figures, are available on Dryad at 10.5061/dryad.cfxpnvxdh. Source data are provided with this paper.

Code availability

All code is freely available at https://github.com/Enserink-lab/AutoDRY and 10.5061/dryad.cfxpnvxdh. For any additional information required for reanalysis of the data reported in this study, please contact the corresponding authors.

Competing interests

The authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41556-025-01837-0.

Supplementary information

The online version contains supplementary material available at 10.1038/s41556-025-01837-0.