Characterizing gene perturbations in single cells via network divergence analysis

Abstract

Functional perturbations of genes do not always cause expression changes, but can manifest through network rewiring or context-specific shifts in regulatory activity. However, inferring functional shifts in genes and linking them to the specific cell populations remains challenging, as most current scRNA-seq data analysis focuses either on differential gene expression or on cell abundance/state changes, but rarely associate gene perturbations with particular cell populations. Here we present scDNS, a framework that quantifies gene-specific functional perturbations by measuring information-theoretic divergence between condition-specific gene interaction network configurations. In simulated stress tests and multiple experimental datasets, scDNS prioritizes key regulators and perturbed cell populations, even when expression changes are minimal but network rewiring is pronounced. Applications to immunodeficiency mutations, stimulus responses, and viral infection reveal hidden regulatory programs and heterogeneous responder cell states. In pancreatic cancer, scDNS nominates TIMM44 as a mitochondrial sensitizer enhancing gemcitabine efficacy. Together, scDNS provides a powerful tool for inferring dynamic gene perturbations in single cells.

Gene functional shifts can rewire molecular networks beyond expression changes. Here, authors present scDNS, which quantifies network divergence to infer gene perturbations and projects them to single cells, enabling sensitive detection of hidden regulation and heterogeneous responder states.

Introduction

Gene interaction networks (GINs) integrate multiple layers of molecular interactions, such as transcription factor-DNA binding, protein-protein interaction, and signaling pathways, to coordinate gene expression programs that define cellular identity, function, and adaptation to environmental stimuli. These networks exhibit inherent dynamism and modularity, wherein perturbations—including genetic mutations, infections, environmental stressors, or pharmacological agents—can alter regulatory relationships and modular structures, thereby reprogramming cellular states and phenotypes^1,2. For example, the T95R mutation in IRF4 modifies its DNA-binding affinity, thereby restructuring its target gene network and resulting in defective B-cell maturation, reduced immunoglobulin class switching, and immunodeficiency³. Mutations in TP53 reorganize downstream transcriptional modules, yielding lineage-specific oncogenic outcomes that differ between epithelial and hematopoietic cells⁴. Pyruvate kinase M2 (PKM2), conventionally regarded as a glycolytic enzyme, undergoes striking remodeling of its interaction network upon activation of the EGFR–ERK pathway. Phosphorylation-dependent nuclear translocation enables PKM2 to relinquish its metabolic role and act as a transcriptional coactivator of β-catenin⁵. These cases highlight that perturbations extend beyond changes in gene expression, fundamentally restructuring the GINs that govern cellular processes. Thus, quantifying such network reorganizations will provide a valuable approach to deciphering how perturbations modify molecular circuits, identifying critical regulatory nodes, and elucidating adaptive mechanisms in disease progression and therapeutic responses. Importantly, further mechanistic interpretation requires more than ranking “perturbed genes” alone: it also demands resolving in which cell states these regulators exert their effects—particularly when the perturbed program is confined to a specific cell subpopulation or manifests heterogeneously across cell states.

Advances in single-cell RNA sequencing (scRNA-seq) have revealed the heterogeneous and dynamic nature of gene regulation across diverse cell types and states. This inherent complexity poses a significant challenge for comprehensively identifying functionally perturbed genes by simply reconstructing universal GINs and analyzing network reconfiguration^6,7. Current strategies for scRNA-seq analysis still predominantly rely on differential expression analysis (DEA) methods coupled with pathway enrichment analysis. Methods, such as Wilcoxon rank-sum or DESeq2-based methods, assess differences in mean expression or variance but fail to capture alterations in regulatory connectivity^8,9. As a result, genes with unchanged expression yet profoundly modified regulatory associations often evade detection. Emerging DEA approaches, such as waddR¹⁰ and covariance-oriented frameworks like Memento¹¹, capture distributional changes and gene–gene correlations. However, they remain insufficient for quantifying functional perturbations that specifically modulate gene–gene interactions. Similarly, regulon-centric tools like VIPER¹² and SCENIC¹³ depend on predefined, static regulatory models. These methods often underperform in dynamic contexts where regulators and targets exhibit condition-specific rewiring, as exemplified by the IRF4 pT95R mutation³. Critically, even when a candidate regulator is nominated, existing pipelines rarely provide a principled way to map that regulator’s functional perturbation back onto individual cells, leaving the downstream cellular “footprint” of rewiring ambiguous. This disconnection limits mechanistic follow-up, because it becomes unclear whether a putative driver reflects a broad state transition, a rare responder subset, or a context-restricted program within a specific lineage.

Detecting cell populations subject to condition-specific perturbations represents an additional analytical challenge. Current strategies, including Milo¹⁴ (which evaluates differential abundance via k-nearest neighbor graphs), scDist¹⁵ (which quantifies donor-specific variability using mixed-effects models), MELD¹⁶ (which computes perturbation probabilities from density variations in latent space), and Augur¹⁷ (which employs machine learning to prioritize cell types based on separability across conditions), offer insights at the population level but primarily focus on compositional or broad transcriptional shifts rather than the underlying rewiring of interaction networks driving functional alterations. As a result, cell- and gene-focused analyses are often performed in parallel but remain decoupled: cell-state changes may be identified without linking to upstream regulators, and regulators may be prioritized without resolving their cell-state–specific effects. Consequently, there is a pressing need for a comprehensive computational framework that quantifies perturbation-driven GIN reorganization at both gene and cellular resolutions, incorporating prior biological knowledge while accounting for the heterogeneity characteristic of scRNA-seq data.

Here, we present scDNS, a computational platform that frames GIN remodeling as an information-theoretic divergence between network architectures across contrasting conditions. scDNS first assesses a gene’s functional perturbation by quantifying dynamic alterations in its gene–gene dependency patterns, subsequently inferring the resultant effects on individual cells to detect the perturbed cell populations. By coupling regulator-level rewiring with cell-resolved perturbation readouts, scDNS bridges upstream molecular drivers and downstream cellular consequences within a unified framework. In both simulated gene and cell perturbation tests, scDNS demonstrated superior sensitivity and robustness in detecting predefined perturbed genes and cells compared to current methods. Furthermore, application across diverse biological contexts—including DSP knockdown, concurrent mutations in TERT, TP53, and PDGFRA, IRF4^p.T95R-associated immunodeficiency, IFN-β stimulation in lupus-derived T cells, COVID-19 lung single-nucleus RNA-seq, and gemcitabine treatment in pancreatic cancer—consistently revealed previously overlooked perturbed regulators and delineated their cell-type-specific effects. We discovered the multifaceted effects of IRF4^p.T95R on B cells, natural killer (NK) and natural killer T (NKT) lineage development, upstream mediators of IFN-β-induced T-cell tolerance, alveolar cell susceptibilities in COVID-19 with therapeutic implications, and TIMM44 as a mitochondrial factor enhancing gemcitabine cytotoxicity. Collectively, these findings establish scDNS as a robust and versatile platform for quantifying functional perturbations within GINs. This capability significantly enhances our power to unravel the molecular basis of cellular responses, reveal previously latent regulatory processes, and ultimately accelerate the identification of therapeutic candidates across a wide array of biological systems.

Results

Overview of scDNS

GINs, comprising multifaceted molecular interactions such as protein-DNA binding, protein-protein interactions (PPIs), and signaling cascades, underpin the orchestration of transcriptional programs in cellular systems. Perturbations, whether experimental or pathological, can profoundly remodel these networks by modulating gene–gene dependencies, inducing transcriptional reprogramming, and reorganizing modular connectivity. Quantifying such reconfiguration provides a rigorous approach to elucidating perturbation effects on cellular phenotypes and revealing core regulatory principles. We introduce scDNS, a computational framework that conceptualizes network reconfiguration as the information divergence between two conditions. It offers a robust metric to evaluate gene-level functional disruptions and to characterize the cellular consequences of perturbations under contrasting conditions. The methodology proceeds through three sequential stages.

The first step reconstructs a context-specific GIN by integrating prior biological knowledge with single-cell transcriptomic data to capture condition-dependent interactions (Fig. 1a). We employed a Graph Attention Network (GAT)¹⁸, a specialized Graph Neural Network, to refine an initial reference network assembled from curated regulatory interactions (TRRUST¹⁹ and Dorothea²⁰), PPIs (IntAct²¹ and STRING²²), and signaling pathway databases (KEGG²³ and SuperPathway²⁴) (see “Methods”). This integrated prior network covers >90% of genes detected in the human single-cell datasets analyzed in our study, and ablation analyses confirm that incorporating these curated databases improves scDNS’s ability to infer biologically meaningful perturbation signals (Supplementary Fig. 1a, b). The model represents the network as a graph G = (V, E), where vertices V correspond to genes and edges E denote regulatory interactions. Vertex features are encoded using node2vec²⁵ to capture network topology, while edge features are represented by embeddings that integrate three attributes of neighborhood relationships. To train the GAT, we adopted a semi-supervised contrastive learning strategy by leveraging high-confidence positive and negative edge labels to guide model optimization. Positive edges were defined as those involving housekeeping genes (e.g., consistently expressed across cell types) or cell-type-specific markers (e.g., derived from scRNA-seq cluster annotations), enabling the model to learn both universal and cell-type-specific interactions. Negative edges were assigned to pairs with zero geometric mean expression in scRNA-seq data or randomly sampled gene pairs lacking supporting evidence in prior knowledge databases. GAT is optimized using a composite loss function comprising: (1) an edge classification loss, which evaluates attention scores against labeled positives to prioritize biologically relevant interactions, and (2) a graph information loss based on Deep Graph Infomax (DGI)²⁶, which maximizes the mutual information (MI) between the encoder representations of the original graph and a corrupted graph generated by randomly shuffling vertices and edges. To ensure robustness, the model is trained ten times, with attention scores aggregated to rank edge reliability. Edges consistently receiving high attention scores (top 80% by ranking aggregated score) were retained as strong, context-specific interactions, while the lowest 20% were pruned to eliminate noisy or spurious connections. This GAT-based network inference is stable, reproducible, and scalable, yielding robust attention patterns and reliable convergence across marker-gene choices, high-confidence edge construction, stochastic training runs, and upstream clustering/annotation variability (Supplementary Note 1). It effectively distinguished true interactions from random ones, achieving a high area under the Receiver Operating Characteristic (ROC) curve of 0.99 (Supplementary Fig. 1c, d). This process yielded a refined GIN that balances universal regulatory interactions with cell-type-specific and condition-dependent edges, providing a robust foundation for downstream analyses.

Fig. 1. Overview of the scDNS approach.

a Context-adaptive GINs reconstruction. Single-cell data were grouped into different cell clusters. For each cluster, an initial cell cluster–specific GIN was constructed using prior knowledge networks (GRIs, PPIs, and pathways). High-confidence positive and negative gene–gene interactions were annotated based on gene expression patterns. Node embeddings (via Node2Vec) and edge embeddings (combining expression, co-expression, and network context) were derived for each cell type. All networks were jointly encoded by GATv2 trained with a hybrid objective combining unsupervised contrastive learning (Deep Graph Infomax) and supervised edge classification. After iterative training, low-confidence interactions in all cell types were removed, yielding refined, context-adaptive GINs that integrate both structural and functional constraints. b Network-based quantification of gene perturbation. For each gene pair (G₁–G₂) across two conditions, gene expression values were projected onto a two-dimensional space to estimate local interaction densities. Joint and conditional probability distributions were constructed to characterize interaction dynamics: the joint distribution effectively captures perturbations within high-density cell populations, whereas the conditional distribution sensitively reflects subtle shifts in low-density regions. For each gene pair, Jensen–Shannon divergence (JSD) was computed between the two conditions for both distributions and integrated to quantify interaction-level perturbations. Gene-level perturbation scores were then derived using a network ensemble approach that aggregates JSDs across all interaction partners of a gene and normalizes them into Z-scores. This ensemble strategy mitigates node-degree bias and yields robust quantification of gene perturbation. c Cell-specific perturbation profiling. For each single cell, we extracted the expression levels of a gene pair (G₁,G₂) and assigned the corresponding point-wise JSD as the contribution of that interaction to the cell-specific perturbation. The perturbation score of a target gene within a cell was obtained by summing the contributions from all its interacting partners. This procedure generated a cell-specific gene perturbation profile, which, when integrated with gene expression profiles, enables the identification of cells exhibiting perturbed molecular states. Source data are provided as a Source Data file.

The second step involves quantifying perturbation-induced alterations in gene–gene relationships within the refined GIN to identify regulatory genes across two contrasting conditions (Fig. 1b). To address data sparsity and heterogeneity inherent in single-cell data, we model the relationship between gene pairs using two complementary probability distributions: (1) a joint probability distribution that captures universal expression relationships prevalent in abundant cell populations, and (2) a conditional probability distribution that enhances sensitivity to regulatory shifts in sparse or transitional cell subpopulations^27,28. Together, this dual-model framework encapsulates both global and context-specific gene–gene relationships (Supplementary Fig. 1e–h). The Jensen–Shannon divergence (JSD) measures the information-theoretic distance of these distributions between the two conditions, enabling the detection of both broad and localized perturbations in gene expression patterns, and also providing insights into changes in cell population abundance or cell states. For each gene pair, the JSD value is weighted by the geometric mean of their expression levels to emphasize changes in transcriptionally active regions. The weighted JSD values are aggregated for each gene across its interacting partners to derive a functional perturbation score. To mitigate bias from heterogeneous regulatory module sizes, raw scores are transformed into degree-normalized Z-scores based on a null distribution derived from randomized networks.

The final step distributes each gene pair’s JSD value to individual cells based on its pointwise divergence contribution (Fig. 1c). By summing the pointwise divergence contributions of all associated gene pairs for each gene, scDNS generates cell-specific perturbation scores that measure the impact of gene-level functional changes within each cell. All gene perturbation scores are then combined to compute a composite cellular effect score, which reflects the overall impact of perturbations on cell phenotypes, explaining changes in either cellular states or population proportions. In summary, scDNS quantifies functional perturbations at both gene and cellular levels by measuring network divergence across contrasting biological conditions, identifying key regulators and their responsive cell populations.

Systematic assessment of scDNS’s performance

We evaluated the performance of scDNS using both simulated and real scRNA-seq datasets. For stress testing, three simulated datasets were generated from iPSC-derived type 2 alveolar epithelial cells (iAT2)²⁹, H1 human embryonic stem cells (H1-hESCs), and PANC-1 cells³⁰, respectively, with cells randomly assigned to control or in silico–perturbed groups. We perturbed each simulated dataset by modifying a subset of 100 genes under three biologically motivated regimes: (i) differential gene expression only (DGE), (ii) gene regulatory rewiring (DGR), and (iii) their combination (DGE + DGR); importantly, the simulation engine introduces these changes as structured, context-dependent perturbations to better mimic diverse, biologically realistic scenarios (Fig. 2a and Supplementary Note 2). To comprehensively characterize the problem difficulty and assess the power of our method to detect the perturbation signal, we generated ground truth case-control datasets by varying two key aspects. First, to further assess the effect of perturbation strength, we defined perturbed cells as hybrids of control cells and perturbed cells of variable relative weight (“Methods”). Decreasing the strength of the transcriptional difference between perturbed and unperturbed cells increased the difficulty of identifying the perturbed cell cluster. Second, perturbations were applied to varying fractions of cells, ranging from 10% to 100%. In these challenging scenarios, scDNS achieved accuracy exceeding 80% for either slight perturbation strength (>0.1) or a low fraction of perturbed cells (>10%) (Fig. 2b, c and Supplementary Fig. 2a). These results highlight the robustness of scDNS in detecting functionally perturbed genes, whether arising from altered expression levels or perturbations in gene–gene relationships.

Fig. 2. Evaluation of scDNS using simulated and experimental datasets.

a Schematic illustration of in silico perturbation design. Gene expression matrices were perturbed by modifying gene–gene relationships (DGR), gene expression levels (DGE), or mixed (DGR + DGE). Perturbation strength was modulated by α (0–1), defined as the hybridized expression between control and perturbed states, and perturbed cell proportions ranged from 10% to 100%. b, c Benchmark accuracy of scDNS in identifying perturbed genes under varying perturbation strengths (b) and proportions (c) in the iAT2 dataset. n = 1253 independent simulations. d Comparison of original cell embeddings (right) and embeddings integrated with perturbation scores (left) for simulated datasets. e scDNS performance (ROC AUC) in detecting perturbed cells across DGR, DGE, and mixed perturbations in the iAT2 dataset. n = 50 independent experiments. f UMAP plots of experimental datasets showing wild-type (WT) and perturbed cells under triple-mutant (TERT^C228T + TP53^G743A + PDGFRA^D842V) and DSP knockdown conditions. g Volcano plots of log₂ fold change versus scDNS-predicted perturbation scores. Lowly expressed TERT (1.1% of cells) was excluded from downstream analyses. h Comparison of scDNS with existing perturbed-gene detection methods in identifying perturbed genes of PDGFRA, DSP, and TP53. Note that PDGFRA and TP53 are not evaluable by MAST, and PDGFRA is not evaluable by memento_de and memento_dv. Significance determined per method; P values adjusted via Benjamini–Hochberg. Large circles = BH-adjusted p < 0.01; small = non-significant. memento_de and memento_dv denote differential mean expression and differential expression variability under the Memento noise model, respectively. i Comparison of scDNS with VIPER and expression-based metrics in identifying perturbed cells for single-gene perturbations (TP53 and DSP). j Performance comparison of scDNS with MELD, Milo, Augur, and scDist in identifying perturbed cells. scDNS integrates perturbation scores from multi-gene perturbations to infer single-cell perturbations. Milo infers perturbed cells based on averaged fold changes across neighboring cells (Nhood), whereas scDist and Augur evaluate perturbation magnitude at the cluster level without single-cell resolution. Box plots (b, c, e) show the median, IQR, and whiskers extending to 1.5 × IQR. Source data are provided as a Source Data file.

We next assessed scDNS’s ability to recover perturbed cell subpopulations driven by diverse gene perturbations. In this validation setting, we artificially dysregulated 30 genes across three distinct subpopulations derived from the major cell clusters of the perturbed group. As an illustrative example, the perturbed subpopulations were not visually separated in the original UMAP embedding; however, incorporating the scDNS-derived perturbation scores (P-Score_cell) into the PCA embeddings successfully and clearly distinguished them (Fig. 2d and Supplementary Fig. 2b). The quantified stress tests across all three simulated datasets showed that scDNS maintains high accuracy even under weak and sparse perturbations (α = 0.25; 25% of cells perturbed), achieving AUC > 0.9 for perturbed-cell identification (Fig. 2e and Supplementary Fig. 2c).

To evaluate scDNS in real biological systems, we analyzed two scRNA-seq datasets: iAT2 cells with or without Desmoplakin (DSP) knockdown²⁹, and H1-hESCs either wild type or carrying triple mutations (TERT^C228T + TP53^G743A + PDGFRA^D842V)³¹ (Fig. 2f). In both datasets, DSP (log₂FC = −0.27, false discovery rate (FDR) = 2.6 × 10⁻⁵, Wilcoxon test) and PDGFRA (log₂FC = 0.49, FDR < 6.4 × 10⁻³⁸⁴, Wilcoxon test) showed significant expression changes and therefore could be identified by standard DEA (e.g., Wilcoxon test), distributional differential expression methods (i.e., waddR), and Memento-de for evaluating differences in gene expression levels (Fig. 2g, h). However, the TP53 mutation did not significantly alter its expression (log₂FC = 0.05, FDR = 1, Wilcoxon test) but reconfigured its network of interacting genes (Supplementary Fig. 2d), rendering it undetectable by most conventional differential expression analyses. While Memento’s differential-variability module identified TP53, it failed to detect DSP and PDGFRA. Remarkably, scDNS was the only method capable of identifying all perturbed genes, including both differentially expressed and non-differentially expressed but network-perturbed genes. Furthermore, scDNS achieved superior accuracy compared with VIPER’s inferred protein activity analysis (99.8% vs. 70.5% for PDGFRA; 95.1% vs. 78.5% for DSP; 91.4% vs. 18.1% for TP53) (Supplementary Fig. 2e). Additional perturbed genes identified by scDNS were strongly enriched for biological processes associated with the phenotypes affected by these perturbed driver genes (Supplementary Fig. 2f). Despite its reliance on the reconstructed gene relationship network, scDNS exhibited remarkable robustness to false-positive interactions, with a notable decline in accuracy only observed when the network contained more than 50% spurious interactions (Supplementary Fig. 2g). Moreover, scDNS demonstrated stable performance across genes with diverse degrees of network connectivity (Supplementary Fig. 2h, i).

We extended our validation by benchmarking scDNS against existing algorithms for detecting perturbed cells influenced by DSP knockdown or triple-gene mutations. For analyzing a single gene’s perturbation effects, scDNS demonstrated superior discrimination between contrasting groups compared to metrics derived from inferred protein activity (VIPER) and raw gene expression levels (Fig. 2i). Furthermore, when assessing the overall perturbation effects across all genes, scDNS achieved the highest classification performance among all comparable methods (AUC = 0.99 for triple mutations dataset and 0.95 for DSP knockdown dataset) (Fig. 2j). Notably, Milo emerged as the second-best method, showing only slightly lower accuracy and specificity than scDNS (AUC = 0.92 for triple mutations dataset and 0.94 for DSP knockdown dataset). Collectively, these results robustly establish scDNS as a precise and powerful approach for identifying functionally perturbed genes and their affected cell populations across both simulated and real-world scRNA-seq datasets.

scDNS uncovers pleiotropic effects of IRF4 mutation on immune cells’ regulatory program and development

scDNS was applied to a single-cell RNA-seq profile of 18,640 B cells from healthy donors and individuals heterozygous for the IRF4^p.T95R mutation (Fig. 3a). The IRF4^p.T95R variant, which causes autosomal-dominant combined immunodeficiency, preserves normal IRF4 mRNA and protein expression yet alters its canonical DNA-binding specificity, leading to a global rewiring of its regulatory network³. Such mutation-induced topological reconfigurations of gene–gene interactions exemplify the type of functional network perturbation that scDNS was designed to detect and quantify. Accordingly, scDNS pinpointed IRF4 as a significantly perturbed regulator (top 2%), characterized by significantly altered interaction relationships within its regulatory network (Supplementary Fig. 3a). This sensitivity contrasted with conventional single-cell analytical methods—such as Memento, waddR, and DEA—which either ranked IRF4 poorly or failed to detect its dysregulation, since the T95R mutation does not alter IRF4 mRNA abundance (log₂FC = −0.12, FDR = 1, Wilcoxon test) (Fig. 3b and Supplementary Fig. 3b). Even protein-activity inference methods like SCENIC only detected IRF4 dysregulation modestly (79% accuracy) (Supplementary Fig. 3c).

Fig. 3. scDNS identifies functional perturbations of the IRF4^p.T95R mutation and its impact on B cell differentiation.

a UMAP of B cell subsets from wild-type (WT) and IRF4^p.T95R donors, colored by cell type (inset) and genotype (main plot). b Accuracy comparison of detecting the driver gene of IRF4^p.T95R by scDNS and other methods. Bars represent the proportions of significant (sig) and non-significant (non-sig) genes, respectively. Statistical significance was determined per method; P values were adjusted using Benjamini–Hochberg, with adjusted P < 0.01 considered significant. c UMAP of scDNS-derived IRF4^p.T95R perturbation scores (|sc-Zscore|). d Pseudotime trajectory of B cell differentiation. e Mean dynamics of scDNS sc-Zscore, IRF4 expression (log₂ fold change), and SCENIC activity (log₂ fold change) across 10 pseudotime bins per B cell subtype. Bar colors reflect the mean absolute value of each metric. diff. differentiation. f Correlation between IRF4^p.T95R perturbation scores and pseudotime (two-sided Pearson test). Cells divided into high- and low-perturbation groups based on the density distribution. g, h Comparison of low- versus high-perturbation B cells: g expression of IRF4 target genes quantified by VIPER, h differentiation potential assessed by cytoTRACE. Significance determined via two-sided Wilcoxon test. Cell numbers: WT = 9346; Low = 6588; High = 1904. i Violin plots of transcriptional repressors (NOTCH2, BACH2, PAX5, FOXO1) across B cell subsets. j Gene Ontology enrichment analysis of dysregulated genes identified by scDNS. k Correlation and network analysis of dysregulated immunodeficiency-related genes. Left: Spearman correlations with IRF4^p.T95R sc-Zscores; OMIM-annotated genes highlighted. Right: STRING-based protein–protein interaction network including IRF4 and immunodeficiency genes. l Conditional-density–rescaled heatmap of IRF4^p.T95R and its functional partner IRF8 cell perturbation scores. m UMAP of all PBMCs from WT and mutant donors, colored by cell type (left) and IRF4 scDNS score (right). n Violin plots showing the distribution of IRF4^p.T95R perturbation scores across all identified PBMCs lineages. o Functional shift in NK cells induced by IRF4^p.T95R mutation. UMAPs are colored by cell density, SLEC signature, and MPEC signature. Box plots (g, h) show the median, IQR, and whiskers extending to 1.5 × IQR. Source data are provided as a Source Data file.

We observed that the perturbation score of IRF4 was markedly elevated in plasmablast and plasma-cell subsets (Fig. 3c), consistent with the differentiation blockade reported in IRF4^p.T95R immunodeficiency³. However, Milo, MELD, and Augur were unable to identify this remarkable perturbation in plasma cell development except scDist (Supplementary Fig. 3d–f). To contextualize these perturbations, we reconstructed the B-cell differentiation trajectory (Fig. 3d) and mapped the scDNS-derived IRF4 perturbation scores across pseudotime (Fig. 3e). The scDNS method successfully captured functional perturbations even in early B cells, while metrics based on IRF4 gene expression and SCENIC activity were insensitive to these early changes, exhibiting weak changes (log₂FC < 0.1) (Fig. 3e). For quantitative validation, the IRF4 perturbation scores exhibited a strong correlation (R = 0.86, p = 2.2 × 10⁻¹⁶) with the aggregate expression of established IRF4-bound target genes identified by ChIP-seq data analysis, substantially exceeding the correlations obtained from gene expression (R = 0.49, p = 8.4 × 10⁻⁵) or SCENIC activity (R = 0.42, p = 9 × 10⁻⁴) (Supplementary Fig. 3g). Analysis of B cells before they commit to becoming plasmablasts and plasma cells revealed two distinct subpopulations (Fig. 3f), exhibiting differential expression of IRF4-bound target genes during B cell differentiation (Fig. 3g and Supplementary Fig. 3h). These two groups displayed an inverse relationship between IRF4 perturbation score and differentiation pseudotime, though they showed different perturbation levels (Fig. 3f). Specifically, one subpopulation underwent advanced differentiation despite minimal IRF4 perturbation (Fig. 3h). In contrast, the second group exhibited significantly stronger IRF4 dysfunction, leading to restricted differentiation. This suppression of differentiation was further supported by upregulation of key transcriptional repressors of B cell differentiation—including NOTCH2³², BACH2³³, PAX5³⁴, and FOXO1³⁵ (Fig. 3i). This pattern ultimately indicates that B cells with mild perturbation proceed toward progressive differentiation, while those with strong perturbation remain arrested in immature states.

Beyond IRF4 itself, scDNS identified the top 250 dysregulated genes, implicating pathways central to B-cell immunity and antiviral defense (Fig. 3j). Seventeen of these genes are annotated in the Online Mendelian Inheritance in Man (OMIM) database³⁶ as causative for immunodeficiency syndromes. Significantly, all but one (SEC61A1) showed a positive correlation with the IRF4 perturbation score and were components of a relevant protein-protein interaction network (Fig. 3k). These results suggest that interacting proteins undergo co-modulatory or synergistic functional changes, implying shared underlying pathogenic mechanisms. This is exemplified by IRF8, a known IRF4 functional partner during early B-cell development³⁷, which showed coordinated functional perturbations with IRF4 in the progenitor stages (Fig. 3l).

Extending beyond the B-cell compartment, analysis of the whole peripheral blood mononuclear cells (PBMCs) dataset revealed that IRF4 dysfunction also perturbs the NK and NKT lineages, as evidenced by significantly elevated IRF4 perturbation scores (Fig. 3m, n). This finding contradicted the original clinical report, which found no alteration in NK-cell activation or abundance³. Our analysis further delineated the impact of IRF4 perturbation in the NK lineage, revealing a functional shift from short-lived effector cells (SLEC) toward memory-precursor cells (MPEC). This shift was transcriptionally characterized by the corresponding downregulation of SLEC and upregulation of MPEC gene signatures (Fig. 3o and Supplementary Fig. 3i). Collectively, these results demonstrate that scDNS is highly effective at pinpointing gene dysfunctions arising from pathogenic mutations and delineating their diverse impacts on cell differentiation across various immune lineages.

scDNS uncovers upstream regulators driving heterogeneous cellular responses to external stimuli in PBMCs

To evaluate the value of scDNS in uncovering regulatory mechanisms of cell response to external stimuli, we applied it to a single-cell RNA-seq dataset of 13,999 PBMCs from 8 lupus patients treated or untreated with recombinant human IFN-β³⁸ (Fig. 4a). IFN-β plays a central role in immune activation by interacting with the interferon alpha/beta receptor (IFNAR); however, its functional impacts across diverse cell types and developmental states remain incompletely understood. For perturbed genes identification, scDNS detected 2784 genes significantly responsive to IFN-β (adjusted p < 0.05) (Fig. 4b). Notably, IFNAR1—whose expression change was minimal (log₂FC = −0.13, Supplementary Fig. 4a)—was ranked much higher by scDNS (66th, 99.4%) than by conventional differential expression approaches. The best-performing DEA was memento’s differential-variability analysis (accuracy 88%), while differential analysis of VIPER-inferred protein activity reached 73% accuracy. These results underscore scDNS’s superior sensitivity for identifying direct treatment targets (Fig. 4c). Gene set enrichment analysis further showed that scDNS-identified genes were strongly enriched in interferon-related pathways, including “cellular response to type I interferon” and “type I interferon signaling pathway” with stronger significance levels compared to alternative methods (Fig. 4d). This improvement arose because scDNS not only captured canonical transcriptional responses to IFN-β but also revealed upstream regulators of the signaling cascade that were missed by conventional gene expression analysis (Fig. 4e). For example, JAK1 and WTAP—despite showing negligible transcriptional changes under IFN-β stimulation (log₂FC = 0.01, Wilcoxon rank = 7551; log₂FC = 0.06, Wilcoxon rank = 8011) (Supplementary Fig. 4b)—were prioritized by scDNS (JAK1: scDNS rank = 2263, FDR = 0.02; WTAP: scDNS rank = 1692, FDR = 0.007). This observation suggests that IFN-β modulates their activity via non-transcriptional mechanisms. Indeed, IFN-β stimulates JAK1 activation through receptor-mediated autophosphorylation within the IFNAR–JAK–STAT signaling axis³⁹. For WTAP, a recent study demonstrated that IFN-β induces its phase separation into liquid-like condensates, thereby modulating its role in m⁶A RNA methylation⁴⁰. Consistently, we observed significant enrichment of scDNS-identified genes in related pathways governed by JAK1 and WTAP activity (Supplementary Fig. 4c).

Fig. 4. scDNS reveals upstream regulators and heterogeneous cellular responses to IFN-β in PBMCs.

a UMAP visualization of PBMCs from lupus patients under control (green) and IFN-β-stimulated (purple) conditions, with annotated cell types: B cells, natural killer cells (NK), dendritic cells (DC), Eryth erythrocytes, Mono monocytes, CD4⁺ T cells, CD8⁺ T cells, and megakaryocytes (MK). b Volcano plots of log₂ fold change versus scDNS-predicted perturbation scores. c Comparison of IFNAR1 perturbation detection. Bar plots show the ranking of IFNAR1 by scDNS and other differential expression analysis methods. The heights of purple and gray bars indicate the proportions of VIPER-predicted significant (sig) and non-significant (non-sig) genes, respectively. Statistical significance was determined using the statistical framework implemented in each method (two-sided where applicable). Where necessary, P values were adjusted for multiple comparisons using the Benjamini–Hochberg procedure. Genes with adjusted P < 0.01 were considered statistically significant. d Comparison of the ability of different methods to identify IFN-β–associated pathway genes. Bar plots show the significance of enrichment for perturbed genes identified by each method. e Schematic of the IFNAR–JAK–STAT signaling pathway, highlighting scDNS-detected upstream regulators whose expression changes are not significant in response to IFN-β stimulation. f UMAP overlay of integrated perturbation scores across cell types, showing heterogeneous IFN-β responses. g Swarm plots of sc-Zscore distributions across cell types, with CD4⁺ T and monocytes showing high variability. Cell numbers: Monocytes = 5406; DC = 604; CD4⁺ T = 4899; CD8⁺ T = 814; Megakaryocytes = 236; B cells = 1366; NK cells = 619; Erythrocytes = 55. h UMAP of sc-Zscore (Integrated) for CD4⁺ T cell subsets (Naïve, Memory, Activated), with color intensity indicating response strength. i Heatmap of perturbation scores for perturbed genes across CD4⁺ T cell subsets, with functional annotations for genes specific to each cell type. Box plot (g) shows the median, IQR, and whiskers extending to 1.5 × IQR. Source data are provided as a Source Data file.

We next examined cell-type-specific effects of IFN-β treatment across 8 annotated clusters. By mapping scDNS-derived integrated sc-Z-scores onto the UMAP embedding, we observed heterogeneous responses, indicating distinct sensitivities of individual cell states to IFN-β stimulation (Fig. 4f). CD4⁺ T cells and monocytes exhibited relatively stronger responses than other cell types, with high variability (Fig. 4g). In monocytes, scDNS revealed a significant and stronger response to IFN-β in CD14⁺ (classical) subpopulations than in CD16⁺ (non-classical intermediate) monocytes, aligning with prior studies⁴¹, whereas MELD and Milo did not detect this difference (Supplementary Fig. 4d–h). Within CD4⁺ T cells, memory and naïve subsets were the most responsive, while activated states showed less perturbation (Fig. 4h). Beyond shared responsive genes, scDNS identified distinct mechanisms across three subsets, with memory cells upregulating genes associated with antigen processing and presentation, while naïve cells responded primarily through T cell proliferation (Fig. 4i). Importantly, scDNS revealed that activated T cells responded through the induction of cell tolerance, a key mechanism in lupus treatment. The majority of these responsive genes were not detected by standard DEA. Collectively, these findings demonstrate that scDNS is a valuable tool for discovering key upstream regulators of cascade signaling pathways and heterogeneous cellular perturbations in response to external treatments.

scDNS identifies cellular vulnerabilities and therapeutic targets in COVID-19

To further assess scDNS’s accuracy in the presence of cohort effects (i.e., individual-to-individual variation), we applied it to lung single-nucleus RNA-sequencing (snRNA-seq) data from 19 COVID-19 patients who succumbed to severe infection and 7 healthy donors⁴² (Fig. 5a). scDNS assigned higher ranks to 1738 host genes known to interact with SARS-CoV-2 RNA and viral proteins, with their perturbation scores exceeding those of other genes (Fig. 5b). Based on the aggregated perturbation score of these host genes, we found that SARS-CoV-2 exerted a significantly high perturbation effect on alveolar type 1 (AT1) and alveolar type 2 (AT2) cells (Fig. 5c and Supplementary Fig. 5a). This finding aligns with the well-established role of alveolar epithelial cells as the primary host cells for SARS-CoV-2⁴³. Notably, scDNS-prioritized genes showed significant enrichment in phosphoproteomic and proteomic signatures detected from SARS-CoV-2–treated Vero E6 cell lines in vitro⁴⁴, with enrichment scores rising with treatment duration (Supplementary Fig. 5b). These findings demonstrate scDNS’s ability to identify genes with functional roles at the translational and post-translational levels.

Fig. 5. scDNS identifies cellular vulnerabilities and therapeutic targets in COVID-19.

a UMAP visualization of single cells from COVID-19 patients and healthy donors, colored by cell type (left) and condition (right). b Violin plots of perturbation scores for host genes interacting (TRUE) or not interacting (FALSE) with SARS-CoV-2 proteins (n = 322 and 15769, respectively) or RNA (n = 241 and 15850, respectively); significance determined via two-sided Wilcoxon test. c Boxplots of average scDNS sc-Zscores for host genes interacting with SARS-CoV-2 across cell types. Each box represents all cells within a given cell type (n per cell type shown in Source Data). Bar plots above show gene expression, indicating that scDNS identifies disease-relevant genes independently of high expression. FB fibroblast. d UMAP of AT2 cells colored by perturbation score, revealing a gradient from low to high. e Volcano plot of differentially expressed genes in high-perturbation AT2 cells versus other cells, defining a severe-infection signature (two-sided Wilcoxon test, BH-adjusted). f Severe-infection signature enrichment in AT2-enriched spots increases toward the bronchus. Correlation with distance was assessed using a two-sided Pearson test. Spot numbers per distance bin: Near = 198; Middle (Mid.) = 190; Far = 67. g Spatial expression map of SLPI (left), showing elevated levels in bronchus-proximal regions. Violin plot (right) confirms higher SLPI expression in cells with high perturbation scores. h Bar plot showing the success rate (frequency of tests with antiviral activity) for drugs targeting different gene sets. Success rates increase substantially when prioritizing targets using scDNS scores. HP-pathogen, host proteins interacting with SARS-CoV-2. i Network of scDNS-predicted disease genes and NCATS drugs with potential antiviral activity. Nodes represent viral components, drugs, and disease genes. Edges represent drug–target or virus–host interactions. Box plots (b, c, f) show the median, IQR, and whiskers extending to 1.5 × IQR. Source data are provided as a Source Data file.

At a finer resolution, scDNS revealed heterogeneous effects of SARS-CoV-2 on AT2 cells, characterized by a functional gradient ranging from low to high perturbation scores (Fig. 5d). Cells with high perturbation scores exhibited significant upregulation of 616 genes that can be considered as a transcriptional signature of severely infected cells (Fig. 5e). These signatures were enriched in pathways critical for SARS-CoV-2 survival and replication, including antigen presentation, protein translation, viral replication, and life cycle regulation (Supplementary Fig. 5c). Among these, we identified elevated expression of secretory leukoprotease inhibitor (SLPI) in cells with high perturbation scores (Fig. 5e). Spatial transcriptomics⁴⁵ revealed that AT2 cells with these signatures were primarily located in bronchial-proximal regions (Fig. 5f), highlighting this niche as a key site of SARS-CoV-2 infection. This spatial-functional relationship was underscored by increased SLPI expression, an anti-protease predominantly produced in the airway⁴⁶ (Fig. 5g). We further applied scDNS to immune cells, revealing pronounced heterogeneity in response to SARS-CoV-2 infection (Supplementary Fig. 5d). Signature genes from cells with high perturbation scores were mapped onto bulk PBMC RNA-seq data from 274 COVID-19 patients⁴⁷, showing that signatures of CD4⁺ T cells, CD8⁺ T cells, cycling NK/T cells, and dendritic cells increased markedly with disease severity—more prominently than genes identified by direct case–control comparisons (Supplementary Fig. 5d–h). Functional enrichment analysis further demonstrated that these upregulated signatures were significantly enriched in inflammatory pathways, COVID-19 infection–related signaling cascades, and previously reported severity-associated gene programs (Supplementary Fig. 5i–k, Supplementary Data 1, and Supplementary Note 3). These signatures increased markedly with disease severity, establishing a strong link between transcriptional dysfunction of these signatures and COVID-19 progression.

To assess the therapeutic potential of scDNS-prioritized genes, we evaluated the antiviral activity of drugs and compounds targeting them. For this analysis, we utilized antiviral efficacy profiles of ~10,000 approved drugs and compounds from the NCATS COVID-19 portal and linked each compound to its targets using DrugBank annotations. Across NCATS assays, only 40% of drugs showed antiviral activity, and this proportion increased modestly to 45% when we restricted to virus-interacting host proteins (Fig. 5h). In contrast, prioritizing targets by scDNS based on high perturbation scores with increasingly stringent thresholds substantially improved success rates, reaching 56% among the top 50 genes and 62% when further limited to those host genes directly interacting with SARS-CoV-2. Furthermore, projection of scDNS-nominated genes onto drug–target interaction networks identified 76 genes linked to 257 candidate compounds, all of which exhibited anti–SARS-CoV-2 activity in at least two NCATS assays (Fig. 5i). These compounds included modulators of viral replication and host immune signaling pathways, providing a diverse repertoire of translationally promising therapeutic options. Collectively, these findings highlight scDNS as an effective framework for prioritizing therapeutic candidates in cohort dataset analysis, while reducing the risk of failure.

scDNS identifies TIMM44 as a regulator sensitizing pancreatic cancer cells to gemcitabine

To demonstrate the ability of scDNS to facilitate novel biological discovery, we applied it to a scRNA-seq dataset of 1749 in vitro PANC-1 pancreatic cancer cells subjected to gemcitabine treatment or vehicle control³⁰ (Fig. 6a). scDNS identified 2669 genes implicated in gemcitabine’s mechanism of action (MoA). The top-enriched pathway was the DNA damage response, consistent with gemcitabine’s well-established function in inhibiting DNA replication and inducing DNA damage through incorporation into DNA and depletion of deoxyribonucleotide pools (Fig. 6b). Notably, scDNS successfully prioritized the known gemcitabine targets RRM1 and RRM2 as the 44th and 27th most perturbed genes, respectively. In addition, 448 highly ranked genes were linked to cell-cycle regulation and DNA repair, 23 of which have been experimentally confirmed to contribute to gemcitabine’s MoA (Fig. 6c). For example, CHK1, a central mediator of gemcitabine-induced DNA damage signaling⁴⁸, showed minimal differential expression (log₂FC = 0.14, FDR = 1, Wilcoxon test) yet a strong perturbation score (Z-score = 6.3, FDR = 2.1 × 10⁻⁸), highlighting scDNS’s ability to capture functional disruptions beyond expression changes. Beyond canonical DNA damage pathways, scDNS also revealed non-canonical effects of gemcitabine on mitochondrial function (Fig. 6d). Among these, SIRT1 (DEA: log₂FC = 0.18, FDR = 0.06, Wilcoxon test; scDNS: Z-score = 6.3, FDR = 2.5 × 10⁻¹⁴), an NAD⁺-dependent deacetylase that promotes mitochondrial biogenesis and redox homeostasis, was markedly perturbed and has been reported to mediate gemcitabine’s MoA⁴⁹.

Fig. 6. scDNS identifies TIMM44 as a non-canonical target and delineates heterogeneous drug responses to gemcitabine in pancreatic cancer.

a UMAP projection of PANC-1 cells showing clustering by treatment: vehicle control (green) and gemcitabine (purple). b Gene set enrichment analysis (GSEA) of high-perturbation genes showing enrichment in DNA repair, cell cycle, and mitochondrial pathways; the bottom panel shows perturbation score distribution. c, d Gene networks of highly perturbed genes for gemcitabine’s canonical MoA (DNA damage, cell cycle, (c)) and noncanonical MoA (mitochondrial dysfunction, (d)). e UMAP projection of PANC-1 cells clustered by cell-level scDNS perturbation score, revealing six cellular subpopulations (clusters 0–6) with distinct responses to gemcitabine. Violin plots show the sc-Zscore distribution per cluster (n per cell cluster shown in Source Data). f Hazard ratios (HRs) estimated from Cox regression using enrichment scores of the transcriptional signatures of each cell population, based on the clinical information from the TCGA PAAD cohort (n = 178). Data are presented as mean ± 95% confidence intervals (CIs). Statistical significance was assessed using two-sided Wald tests. g Heatmap showing CancerSEA gene set enrichment scores across 6 tumor cell subpopulations. h Heatmap of gene perturbation scores across treatment-enriched clusters (C1, C3, and C6), showing enrichment of highly perturbed genes in cluster C3 associated with cell cycle and mitochondrial functional pathways. i Volcano plot comparing scDNS perturbation scores and differential expression (log₂FC) for all genes, with mitochondrial function–related genes highlighted. The UMAP (right) shows TIMM44 perturbation concentrated in C3. j–m Experimental validation: j gemcitabine + MitoBloCK-10 cytotoxicity (n = 3 independent experiments), k PCA of bulk RNA-seq data, l heatmap of differentially expressed genes under combination treatment, and m bar plot of up-/downregulated gene enrichment. Significance in (h and m) was assessed using a hypergeometric test with Benjamini–Hochberg adjustment. Box plot (e) shows the median, IQR, and whiskers extending to 1.5 × IQR. Source data are provided in the Source Data file.

Based on cell-level perturbation scores, we identified six cellular subpopulations exhibiting distinct responses to gemcitabine, with cluster 3 showing the strongest perturbation (Fig. 6e). Projection of each subpopulation’s transcriptional signature onto TCGA pancreatic adenocarcinoma (PAAD) patient survival data revealed significantly different survival outcomes, where the signatures of cluster 3 are associated with the poorest prognosis (hazard ratio = 1.76, p = 3.3 × 10⁻⁵, two-sided Wald test) (Fig. 6f). In contrast, clusters defined solely by gene expression showed weaker prognostic stratification (Supplementary Fig. 6a, b). Cluster 3 was characterized by metastatic potential, impaired apoptotic regulation, and enhanced invasiveness (Fig. 6g). Genes with the highest perturbation scores in cluster 3 were enriched for cell-cycle and mitochondrial functions (Fig. 6h). Among them, SDHA (succinate dehydrogenase complex flavoprotein subunit A) was the most differentially perturbed gene (Fig. 6i). SDHA deficiency leads to succinate accumulation, which has been implicated in drug resistance through ubiquitin-cullin pathway dysregulation in several cancers, including acute myeloid leukemia⁵⁰. The second-most perturbed gene, TIMM44, a core component of the presequence translocase-associated motor (PAM) complex of the mitochondrial inner membrane, exhibited no significant change in expression (log₂FC = 0.33, FDR = 1, Wilcoxon test) but a marked perturbation in network activity (Z-score = 9.0, FDR = 1.1 × 10⁻¹⁵) (Fig. 6i). Western blot analyses confirmed stable TIMM44 expression before and after gemcitabine treatment (Supplementary Fig. 6c). Complementary proteomics/phosphoproteomics further showed stable total TIMM44 protein but a significant increase in S180 phosphorylation after treatment (FC = 2.6, p = 0.006, two-sided t-test) (Supplementary Fig. 6d). S180 maps to a functionally important domain involved in TIMM44’s association with the translocon and recruitment of the mtHsp70⁵¹, and is supported as a bona fide regulatory site by recurrent detection in PhosphoSitePlus and strong evolutionary conservation (Supplementary Fig. 6e–g). These orthogonal data are consistent with a potential activity or interaction-state shift of TIMM44, providing supportive evidence for its prioritization by scDNS even in the absence of differential expression. Independent bulk RNA-seq datasets from TCGA and GTEx showed abnormally elevated TIMM44 expression in tumor versus normal tissues, indicating its potential as a therapeutic target (Supplementary Fig. 6h). Similarly, an independent single-cell dataset⁵² confirmed TIMM44 upregulation specifically in tumor cells (Supplementary Fig. 6i), supporting its potential as a tumor-selective target with low off-target risk.

To experimentally validate TIMM44-mediated sensitization to gemcitabine, we performed in vitro combination assays using gemcitabine and the mitochondrial import inhibitor MitoBloCK-10 (MB-10). MB-10 potentiated the cytotoxic effect of gemcitabine (Fig. 6j and Supplementary Fig. 6j). At low gemcitabine concentrations (≤8 μM), the combination of MB-10 and gemcitabine exhibited a synergistic effect with an average Chou–Talalay combination index (CI) of 0.44. Transcriptomic profiling revealed that gemcitabine induced major transcriptional changes, further amplified in the gemcitabine + MB-10 combination (Fig. 6k). DEA showed that downregulated genes were enriched for cell-cycle and DNA repair pathways. Notably, several of these genes—including SIRT1⁴⁹, DUSP1⁵³, EP300⁵⁴, and MCL1⁵⁵—have been previously implicated in gemcitabine resistance (Fig. 6l, m). Collectively, these results demonstrate that scDNS effectively delineates heterogeneous drug responses, identifies non-canonical perturbation mechanisms, and uncovers novel sensitivity-associated genes such as TIMM44, highlighting its potential as a therapeutic target for improving gemcitabine efficacy.

Discussion

In this study, we present scDNS, a computational framework designed to quantify reconfiguration of GINs and evaluate gene-level functional perturbations and their cellular consequences, representing a substantial advance in single-cell network analysis. Conceptually, scDNS advances beyond conventional expression-centric analyses by treating regulatory perturbation as a network divergence problem, enabling systematic detection of functional dysregulation that may occur independently of large transcriptional changes. The first major innovation of scDNS lies in its prior-aware and context-adaptive network reconstruction module, which leverages a GAT–based deep learning model to refine prior biological networks into context-specific GINs. The reconstructed networks faithfully represent diverse cell type- and state-specific gene–gene interactions, enabling the systematic assessment of network reconfiguration induced by experimental perturbations across different cell populations. This balance between biological prior knowledge and data-driven refinement allows scDNS to remain sensitive to context-specific rewiring while avoiding the instability and overfitting risks commonly associated with fully de novo network inference in sparse single-cell settings. The second innovation is the introduction of joint and conditional probability distributions to complementarily quantify changes in gene–gene dependencies between contrasting conditions, where joint densities provide a stable global characterization of network reconfiguration, whereas conditional densities are more sensitive to localized rewiring events confined to small cellular subpopulations. By integrating joint and conditional perspectives, scDNS provides a unified measure that is robust across perturbation scales, from widespread pathway activation to partial-penetrance rewiring confined to specific subpopulations.

In benchmarking analyses, scDNS consistently demonstrated strong sensitivity across diverse perturbation regimes, including expression-only changes, rewiring-only scenarios, and combined effects. Stress tests involving moderate removal and corruption of interaction edges further demonstrate that scDNS can reliably recover perturbation signals under realistic prior-network incompleteness and uncertainty. Furthermore, scDNS’s network-based perturbation index detects gene-level functional alterations and the corresponding effects on cell populations, which are unable to be detected by traditional differential expression (DE) or regulon-based methods. DE approaches (e.g., Wilcoxon, DESeq2⁸) inherently ignore molecular network information, whereas activity inference tools (e.g., VIPER¹², SCENIC¹³) rely on static regulon definitions that fail to capture dynamic rewiring of regulator–target relationships under perturbation. By directly quantifying shifts in dependency structures—the edges of the network—scDNS bridges this gap, capturing dynamic remodeling of regulatory interactions and revealing multi-layered functional alterations spanning both transcriptional and non-transcriptional dimensions.

scDNS is computationally scalable. Once embeddings are generated, runtime and memory are driven primarily by the number of genes and gene-graph density rather than cell count, enabling practical analysis of large single-cell datasets (Supplementary Table 1 and Supplementary Fig. 7a, b). Moreover, because scDNS quantifies perturbations through coherent shifts in gene–gene dependency structure and network-level divergence, it is intrinsically less susceptible to technical confounders that predominantly perturb absolute expression statistics, such as library-size scaling and stochastic measurement noise (Supplementary Note 3). Consistently, donor-aware analyses in independent IFN-β–treated lupus PBMC and COVID-19 lung cohorts showed that cohort-level perturbation patterns were preserved across donors and were not explained by common technical covariates (e.g., total UMI counts and detected genes), while the cell-resolved scDNS scores still exposed biologically meaningful inter-patient heterogeneity—such as an immune-activated outlying control donor and severity-associated variation in the fraction of high-response cells validated in an independent bulk RNA-seq cohort (Supplementary Note 3).

Despite its advantages, scDNS has certain limitations. First, it currently operates under a two-condition paradigm (control versus perturbation). Extending it to multifactorial designs—such as time-series, dose–response, or combinatorial perturbations—would require generalizing the divergence framework into higher-dimensional or tensor-based forms. Second, while scDNS incorporates curated interaction priors, incomplete annotation remains a challenge for under-studied genes and pathways. Nonetheless, degree-normalized statistics and robustness analyses indicate that scDNS tolerates substantial prior incompleteness, suggesting that meaningful rewiring signals can still be recovered even when interaction knowledge is sparse. Future integration of multi-omics priors, such as chromatin accessibility or protein interaction dynamics, may further mitigate this limitation. Third, while scDNS detects the presence of functional perturbations, it does not yet localize their specific molecular layer; integration with complementary omics datasets could further enhance mechanistic interpretability.

Beyond methodological advances, the true strength of scDNS lies in the biological insights it enables. In IRF4^p.T95R-associated immunodeficiency, scDNS not only identified IRF4 as the most perturbed regulator but also delineated its pleiotropic effects on B-cell maturation and plasma-cell differentiation—consistent with prior biochemical evidence showing that the IRF4^p.T95R mutation alters DNA-binding specificity and reshapes regulatory network topology. Notably, scDNS uncovered previously unrecognized perturbations within NK/NKT lineages, suggesting a functional shift from short-lived effector phenotypes toward memory-precursor states. In IFN-β stimulated PBMCs from lupus patients, scDNS prioritized IFN-β receptors, JAK1, and WTAP as key upstream regulators in the NF-κB signaling cascade, despite minimal transcriptional changes. This observation indicates that IFN-β primarily induces non-transcriptional functional perturbations of upstream regulators, such as JAK1 autophosphorylation³⁹ and WTAP-mediated m⁶A modification⁴⁰—consistent with recent reports. Cell-type-specific analyses further revealed that IFN-β exerts the strongest perturbations in CD4⁺ T cells and monocytes, inducing tolerance in activated T-cell subsets and heterogeneous activation in monocytes, consistent with multi-omics evidence highlighting cell-state-dependent interferon signaling in systemic lupus erythematosus⁵⁶. Importantly, scDNS achieved superior enrichment of interferon-related pathways compared to DEA or regulon-based methods, underscoring its power to dissect complex multi-layered signaling cascades. In the COVID-19 lung single-nucleus RNA-seq dataset, scDNS identified widespread functional perturbations in host–virus interaction networks, most prominently in alveolar type I and II (AT1/AT2) cells, where reconfigured networks were linked to profibrotic remodeling and viral replication programs. Within AT2 subpopulations, scDNS uncovered spatially restricted responses, characterized by strong SLPI induction within bronchial-proximal niches under severe infection. Notably, perturbation ranking based on scDNS achieved a 62% success rate in predicting antiviral efficacy, significantly outperforming random and virus-interactor baselines. In gemcitabine-treated pancreatic cancer cells, scDNS identified genes involved in DNA damage response and mitochondrial function as key perturbation targets and uncovered heterogeneous drug responses among cells. A subpopulation exhibiting high malignancy and proliferation potential showed a network signature strongly associated with poor prognosis. Within this population, TIMM44 displayed pronounced functional perturbation despite stable expression levels, suggesting that subcellular relocalization or altered regulatory context—rather than expression alone—drives mitochondrial sensitivity to chemotherapeutic stress. Targeting TIMM44 in combination with gemcitabine yielded significant synergistic effects, highlighting its potential as a therapeutic co-target. Across these diverse contexts, scDNS consistently links gene-level rewiring to cell-state–resolved consequences, enabling mechanistic interpretation of how regulatory perturbations propagate through cellular populations.

More broadly, scDNS enables quantitative measurement of functional network remodeling at single-cell resolution, opening new avenues for mechanistic and translational research. Modeling divergence dynamics under incremental perturbations may facilitate the development of AI-driven perturbation simulators capable of predicting network responses in silico prior to experimentation. Integrating scDNS-derived divergence maps with spatial transcriptomics, chromatin accessibility, or proteomic data could reveal how cellular reprogramming is modulated by the microenvironment, spatial architecture, and epigenetic state. In summary, scDNS bridges the gap between descriptive transcriptomic profiling and quantitative assessment of functional network perturbation. By moving beyond expression-level differences to capture the dynamic remodeling of GINs, scDNS unveils hidden layers of regulatory dysregulation and establishes a robust, generalizable framework for mechanistic dissection and therapeutic target discovery across diverse biological systems.

Methods

Context-adaptive GINs reconstruction

A semi-supervised analysis framework based on GATs was designed to identify context-adaptive regulatory networks by integrating single-cell transcriptomic data with a comprehensive prior knowledge network. The overall analysis pipeline was comprised of two steps:

(1) Prior knowledge network construction: A comprehensive biological network encompassing gene regulatory interactions (GRIs), PPIs, and signaling pathways was first constructed. The GRIs were sourced from the TRRUST database¹⁹ and the R Dorothea package (v1.7.4)²⁰. PPIs were obtained from the IntAct²¹ and STRING databases²². Signaling pathways were collected from the KEGG²³ and SuperPathway²⁴ databases. In total, 332,740 unique interactions involving 21,006 genes were collected.

(2) Pre-set high-confidence positive and negative edges: Standard Seurat preprocessing was applied to the scRNA-seq data to identify major cell types. Positive edges were defined as those connecting highly expressed and biologically relevant gene pairs within each cell type. Specifically, the COSG⁵⁷ method was used to identify the top 100 marker genes for each cell type. In addition, housekeeping genes with an expression frequency greater than 75% and average expression above 1 were incorporated. These genes, along with their interactions in the prior knowledge network, were classified as positive interactions for each cell type. Negative edges were defined as (i) gene pairs exhibiting a geometric mean expression of zero across all cell types, and (ii) those randomly sampled pairs lacking prior biological evidence of interaction.

Node and edge encoding

Node embeddings for the prior knowledge network were obtained by applying the Node2Vec algorithm to map each vertex $i$ to a 512-dimensional vector $\overrightarrow{{\mathbf{h}}_i}$. This process captured both local neighborhood structure and global topological properties. For a candidate edge between genes $G_i$ and $G_j$ in cell population $C_k$, the edge feature vector was constructed by integrating: (i) cell-specific co-expression $f_{g_i,g_{j,C_k}}^{\exp }$, defined as the geometric mean expression of the gene pair within $C_k$; (ii) global co-expression statistics $f_{g_1,g_2}^{\mathrm{MI}}$ and $f_{g_1,g_2}^{\mathrm{DREMI}}$, computed as MI and density-resampled mutual information between genes across the whole dataset, log-transformed and sigmoid-normalized; and (iii) contextual features ($f_{g_1,g_2}^{\mathrm{context}}$), a one-hot encoding of biological context (e.g., cell identity and treatment). The final edge feature vector is therefore

$$\overrightarrow{{\mathbf{h}}_{G_i,G_{j,}{C}_k}}=\left[f_{G_i,G_{j,}{C}_k}^{\exp },f_{G_i,G_j}^{\mathrm{DREMI}},f_{G_i,G_j}^{\mathrm{MI}},f_{G_i,G_j}^{\mathrm{context}}\right]$$ 1

Graph attention mechanism for context-adaptive regulatory network prediction

A GATv2-based graph attention encoder was employed to couple with a joint contrastive learning module. This module incorporates graph-level representation learning using DGI and an attention score-based edge supervision mechanism.

GAT encoder module

For a cell-type-specific gene network $G=\left(V,E\right)$, where $V$ is the node set and $E$ is the edge set representing potential interactions $e_{i,j,C_k}$ between genes $g_i$ and$g_j$ in cell type $C_k$, these edges were employed for joint encoding. Given the node input feature $\overrightarrow{{\mathbf{h}}_i}$ and edge feature $\overrightarrow{{\mathbf{h}}_{G_i,G_{j,}{C}_k}}$, the features are linearly transformed using the learnable weight matrices $W_v\in {\mathbb{R}}^{∣V∣\times F^{\prime }}$ and $W_e\in {\mathbb{R}}^{∣E_{all}∣\times F^{\prime }}$, where $F\prime$ is the embedding dimension, and $∣E_{all}∣$ is the total number of edges across all cell type networks. Then, the attention score for the message propagation from node $j$ to node $i$ in cell type $C_k$ was calculated as:

$${\alpha }_{i,j,C_k}=\mathrm{Tanh}\left({\overrightarrow{a}}^T\left[W_v\overrightarrow{h_i}+W_v\overrightarrow{h_j}+W_e\overrightarrow{h_{i,j,C_k}}\right]\right)$$ 2

where $\overrightarrow{a}$ is a learnable attention vector, and $\mathrm{Tanh}$ is the hyperbolic tangent activation function. This score reflects the importance of the neighbor $j$ and its connecting edge to the representation of node $i$. A Softmax normalization is applied over all incoming edges to node $i$ to obtain the normalized attention weight

$${\alpha }_{i,j,C_k}^{\prime }={\mathrm{Softmax}}_{j\in N_i}\left({\alpha }_{i,j,C_k}\right)$$ 3

This normalized score is used for weighted aggregation during message passing and serves as a proxy for the edge existence probability in the next supervised task.

Joint contrastive learning module

The model’s GAT encoder was jointly optimized using both unsupervised contrastive learning and supervised edge classification. DGI is utilized to learn structure-aware node embeddings $Z=f_{\mathrm{GATv}2}\left(V,E\right)$. This is achieved by maximizing the MI between the node representations $Z$ and the graph-level summary vector $s=\mathrm{sigmoild}\left(\mathrm{mean}\left(Z\right)\right)$, while minimizing the MI between the corrupted graph’s node representations $\tilde{Z}$ (generated by randomly shuffling input node features) and the summary $s$. The DGI loss is defined as:

$${\mathcal{L}}_{DGI}=-E_{Z,\tilde{Z}}\left[\log \sigma \left(z^{T}s\right)+\log \left(1-\sigma \left({\tilde{z}}^{T}s\right)\right)\right]$$ 4

This loss ensures the model can distinguish the true graph structure from structural noise.

Additionally, the normalized attention weights serve as the prediction $\widehat{y}={\alpha \prime }_{i,j,C_k}$ for the ground-truth labeled edges $y$ (1 for positive interactions and 0 for negative ones). Focal Loss is employed to mitigate the severe positive/negative sample imbalance:

$${\mathcal{L}}_{\mathrm{edge}\mathrm{classification}}=-\alpha {\left(1-\widehat{\mathrm{y}}\right)}^{\gamma }\log \left(\widehat{y}\right)-\left(1-\alpha \right)\left(1-\mathrm{y}\right){\widehat{y}}^{\gamma }\log \left(1-\widehat{y}\right)$$ 5

Hyperparameters α and γ were set to 0.75 and 1.5, respectively, ensuring that the model focused on hard-to-classify samples. The final overall training objective is the weighted sum of the supervised and self-supervised losses:

$${\mathcal{L}}_{\mathrm{total}}={\mathcal{L}}_{\mathrm{edge}\mathrm{classification}}+\lambda \cdot {\mathcal{L}}_{\mathrm{DGI}}$$ 6

The weight coefficient $\lambda$ was set to 0.1 to balance the two objectives.

Training and edge prediction

Prior networks for each cell type across different conditions were integrated into a multiplex graph as the GAT input, allowing each gene pair to have multiple edges representing diverse cellular contexts. This design enables the model to simultaneously learn interaction patterns across these contexts. To enhance the model robustness and reproducibility, we independently trained the model across 10 random seeds. For each cell type, the attention scores for each edge across the 10 training runs were aggregated, and their Rank Product (RP) was then computed. The final ranking of gene interactions for each cell type was derived by independently sorting the edges based on their RP. To ensure high-confidence gene interactions, the minimum rank across all cell types was employed as the final evaluation criterion. Gene interactions that consistently ranked in the lowest 20% across all cell types were considered to have a low likelihood of occurring in the single-cell data system.

Estimate the perturbation scores for genes and cells

scDNS characterized the dynamic relationships of gene pairs in the network (G ↔ G_i) using joint and conditional gene expression probability densities. Each method had distinct advantages in capturing interactions within both high-density and low-density cell populations, allowing for a comprehensive analysis. The distribution densities were estimated based on a k-nearest neighbor (KNN) approach on a two-dimensional grid, following the widely used procedure of van Dijk et al.²⁸. A small random expression value was incorporated to improve robustness against extreme distributions, such as the zero-inflated distribution frequently observed in single-cell data.

Jensen–Shannon divergence (JSD) was utilized to quantify changes in probability densities across conditions. It provided valuable insights into both local and global information divergences between two probability distributions, providing a robust metric for assessing gene interaction dysregulation. To emphasize biologically relevant interactions, point-wise JSD was weighted by gene expression levels. The statistical significance of JSD was determined using a null distribution generated from 2 × 10⁴ JSD values derived from randomized gene pairs not included in known interaction networks, with additional randomization of gene expression within each condition to account for background variability.

Finally, gene prioritization was performed at both global and single-cell levels. The significance of JSDs associated with a given gene was aggregated while accounting for network degree to avoid bias. Z-scores from joint and conditional probability densities were then combined to obtain an overall gene-level Z-score. Finally, single-cell Z-scores (sc-Zscores) were computed to quantify each gene’s effector cells based on weighted point-wise JSD. A detailed description of the algorithm was provided in the Supplemental Information.

Simulation setup for benchmarking datasets

Three simulated datasets were generated from iAT2, H1-hESCs, and PANC-1 cells, with cells randomly assigned to control or in silico–perturbed groups. Perturbations were applied to a subset of 100 target genes under three biologically motivated regimes: (i) expression-only changes (DGE), (ii) rewiring-only changes affecting regulatory relationships or expression patterns (DGR), and (iii) coupled expression and rewiring changes (DGE + DGR). For rewiring changes, the relationships between each target gene and its partner genes were systematically modified to simulate context-dependent regulatory changes, including activation/inactivation and switches between positive and negative regulation. Only the top 60% of partner genes, ranked by their average Spearman correlation with other partners, were selected for rewiring. Perturbations were constrained within local neighborhoods and functional modules to preserve structured and biologically plausible alterations, enabling benchmarking of scDNS under realistic regulatory complexity. Expression-level perturbations were simulated by multiplying the expression of each target gene in perturbed cells by a factor of two. Combined perturbations were generated by applying both expression-level and rewiring changes simultaneously.

To systematically assess the effect of perturbation strength, perturbed cells were defined as hybrids between control cells and fully perturbed cells. We introduced a perturbation strength parameter, α (ranging from 0 to 1), to control the similarity between perturbed and control expression profiles. The resulting expression profile of each perturbed cell was computed as:

$$E_{\alpha }=\left(1-\alpha \right)\cdot E_{\mathrm{before}\mathrm{perturbation}}+\alpha \cdot E_{\mathrm{after}\mathrm{perturbation}}$$ 7

where $E_{\mathrm{before}\mathrm{perturbation}}$ and $E_{\mathrm{after}\mathrm{perturbation}}$ represent the expression profiles of the control and perturbed cells, respectively. Perturbation strength α was varied to create a gradient of transcriptional differences, simulating conditions ranging from subtle to strong perturbations. We additionally varied the fraction of cells subjected to perturbation from 10% to 100% to assess scDNS’s sensitivity across different levels of heterogeneity.

To further evaluate scDNS’s ability to detect subpopulation-specific perturbations, we first performed clustering of the single-cell data to define independent cell subpopulations. For three selected subpopulations, we then applied DGR, DGE, and combined perturbations to 30 highly expressed marker genes per subpopulation, simulating biologically relevant perturbation scenarios across distinct cell clusters.

Comparison of perturbation detection methods

Gene-level perturbation analysis was performed using waddR (v1.15.0)¹⁰, memento (v0.1.1)¹¹ in two modes (memento_de and memento_dv), and Seurat’s FindMarkers function (v4.3.0)⁵⁸, including the following statistical frameworks: Wilcoxon rank-sum test, bimod likelihood-ratio test, ROC-based marker identification, Student’s t-test, negative binomial and Poisson generalized linear models, logistic regression, and MAST hurdle model. Cell-level perturbation detection was conducted using MELD (v1.0.2)¹⁶, Milo (v2.5.1)¹⁴, Augur (v1.0.3)¹⁷, and scDist (v1.1.4)¹⁵.

scRNA-seq data pre-processing

The Seurat R package (v4.3.0)⁵⁸ was utilized to preprocess the scRNA-seq data. Following standard quality control procedures in single-cell data analysis, low-expression genes were filtered out and then normalized using the NormalizeData function with default parameters. Highly variable genes were identified using the FindVariableFeatures function with the default “vst” method. Subsequently, the principal component analysis (PCA) was applied to the scaled gene expression of highly variable genes. The datasets exhibiting batch-related differences were integrated using the Harmony R package (v0.1.0)⁵⁹ with default parameters. The first 30 PCs were used to construct a shared nearest neighbor graph for cells and to identify cell clusters using the FindNeighbors and FindClusters functions, respectively. The clustering results were visualized using UMAP by the RunUMAP function. Cell annotation was performed using marker genes derived either from the original dataset or from a curated reference. Marker genes for each cluster were identified with the FindAllMarkers function using the “wilcox” method.

Single-cell protein activity inference

Single-cell transcriptomic data were first clustered to define coherent cell subpopulations. For each cluster, pseudo-bulk gene expression profiles were generated by averaging the expression of constituent single cells. Gene regulatory networks were reconstructed from these pseudo-bulk profiles using ARACNe⁶⁰ for identifying the downstream targets (regulons) of each protein. Single-cell protein activity was subsequently inferred using the VIPER R package⁶¹. Proteins with small regulon size (<20 target genes) were excluded. Differences in protein activity between groups were assessed using the Wilcoxon rank-sum test.

Analysis of the expression levels of IRF4 target genes

The IRF4^p.T95R-specific regulatory network was constructed using differential ChIP-seq peaks and expressed genes between IRF4^p.T95R and wild-type samples (data available in ref. ³). The target genes of IRF4^p.T95R were defined as those exhibiting an absolute log₂ fold change greater than 1 in both gene expression levels and normalized peak counts. We classified the target genes into positive and negative regulators based on their upregulation and downregulation in expression levels. The regulatory network of IRF4^p.T95R was then employed to infer its specific abnormal activity in each cell through enriched regulon analysis implemented by the viper function in the viper R package (v1.28.0)⁶¹. IRF4 transcriptional activity was inferred using the pySCENIC Python package (v0.12.1)¹³. Pseudotime and stemness analyses of IRF4 p.T95R B cells were performed using Monocle3 (v1.3.1)⁶² and CytoTRACE (v0.3.3)⁶³ with default parameters.

Survival analysis

Cluster-specific marker genes (|log₂FC| > 0.25, adjusted p < 0.01) were identified from gemcitabine-treated single-cell data using the FindAllMarkers function in R Seurat package. To evaluate clinical relevance, we retrieved TCGA pancreatic adenocarcinoma (TCGA-PAAD) raw count matrices and associated clinical metadata through the R TCGAbiolinks package (v2.35.3)⁶⁴. Each cluster’s highly expressed genes were treated as a transcriptional signature and projected onto bulk tumor profiles by computing sample-level signature enrichment/activity scores using the VIPER algorithm (R viper package (v1.32.0)⁶¹), producing one quantitative score per signature per sample. Associations between signature scores and patient outcome were assessed using univariate Cox proportional-hazards regression with the R survival package (v3.5-3)⁶⁵; signature score was modeled as a continuous covariate and results were reported as hazard ratios (HRs) with 95% confidence intervals.

Cell culture and drug treatment

Human pancreatic adenocarcinoma cell line, PANC-1 cells (CL-0184, Wuhan Pricella Biotechnology Co., Ltd.), was cultured in DMEM supplemented with 10% fetal bovine serum and 1% penicillin-streptomycin at 37 °C in a humidified incubator with 5% CO₂. Cells were seeded in 96-well plates (3 × 10³ cells/well) and allowed to adhere overnight. Then, cells were treated with MitoBloCK-10 (HY-115467, MedChemExpress) and gemcitabine (HY-17026, MedChemExpress), separately or in combination, for 72 h. Cell viability was assessed using the CCK-8 assay (C0037, Beyotime Biotechnology). Briefly, 10 μL of CCK-8 reagent was added to each well and incubated for 2 h at 37 °C. Absorbance was measured at 450 nm using a microplate reader. Background absorbance from blank wells was subtracted. Experiments were performed with three independent biological replicates.

Drug interactions were evaluated by calculating the CI using the Chou–Talalay method (https://www.combosyn.com/). CI values were calculated using Compusyn software based on dose–response curves generated from multiple drug concentrations. CI < 1 indicates synergy, CI = 1 indicates an additive effect, and CI > 1 indicates antagonism.

For Western blot analysis, cells were treated with gemcitabine at indicated concentrations. Cells were lysed in RIPA buffer supplemented with protease inhibitors. Protein concentration was determined using a BCA protein assay. Equal amounts of protein (20 μg per lane) were separated by 10% SDS-PAGE and transferred onto PVDF membranes. Membranes were blocked with 5% non-fat milk for 1 h at room temperature and incubated overnight at 4 °C with anti-TIMM44 antibody (A21141, Abclonal; 1:1000 dilution). After washing, membranes were incubated with HRP-conjugated secondary antibody (1:5000 dilution) for 1 h at room temperature. Signals were visualized using enhanced chemiluminescence (ECL). Protein levels were normalized to GAPDH (A10868, Abclonal).

RNA sequencing for pancreatic cancer cells

For transcriptomic analysis, cells were treated with MB10 (0.64 μM), gemcitabine (0.50 μM), their combination, or DMSO as a vehicle control. Total RNA was extracted from PANC-1 cells using the Total RNA Extractor (Trizol) (B511311, Sangon Biotech) following the manufacturer’s protocol. Libraries were prepared using the Illumina TruSeq RNA Sample Preparation Kit, following standard protocols. The constructed libraries were sequenced on an Illumina NovaSeq 6000 platform. Raw sequencing data were processed using FastQC for quality control and trimmed with Fastp to remove low-quality reads and adaptors. The resulting clean reads were aligned to the human genome (GRCh38) using the STAR aligner. Gene expression profiles were quantified with feature Counts. DEA was performed using DESeq2 (v1.34.0)⁸ with default parameters, and genes with absolute log₂ fold change > 1 and adjusted p < 0.01 were defined as differentially expressed.

Phosphoproteomic profiling by LC–MS/MS

PANC-1 cells were seeded 24 h before treatment and, at 60–70% confluency, treated with equal volumes of DMSO (control) or 1 μM gemcitabine for 36 h. Cells were washed with ice-cold PBS, harvested by scraping and snap-frozen in liquid nitrogen. Frozen cell pellets were lysed in RIPA buffer (P0013B, Beyotime) supplemented with 1 mM PMSF (ST507, Beyotime) and 1× phosphatase inhibitor cocktail A (P1082, Beyotime). Lysates were sonicated on ice and centrifuged at 15,793 × g for 10 min at 4 °C. Supernatants were collected as total protein extracts, and protein concentration was determined using a BCA assay (23225, Thermo Scientific). Equal amounts of protein were processed using the SP3 (single-pot solid-phase-enhanced sample preparation) method, including acetonitrile-induced binding to magnetic beads, ethanol and acetonitrile washes, reduction with dithiothreitol, alkylation with chloroacetamide, and digestion with Trypsin/LysC. Phosphopeptides were enriched using an IMAC phosphopeptide enrichment kit (A32992, Thermo Scientific) according to the manufacturer’s instructions. Enriched peptides were desalted using SOLA™ SPE 96-well plates and dried in a vacuum concentrator prior to LC–MS/MS analysis. Peptides were separated on a Vanquish Neo UHPLC system coupled to an Orbitrap Astral mass spectrometer (Thermo Fisher Scientific). Samples were loaded onto a trap column and separated on a PepMap™ Neo analytical column (15 cm × 150 μm, 2 μm). Mobile phase A consisted of 0.1% formic acid in water and mobile phase B consisted of 0.1% formic acid in 80% acetonitrile.

Reporting summary

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

Author contributions

Z.X. conceived and supervised the study. C.H. implemented the algorithm and performed data analyses with substantial assistance from Y.L., B.F., Y.M., and Z.Z.; Y.L. developed the GAT-based deep learning module under the supervision of Y.X.; J.Z. and Z.L. carried out the experimental validation under the supervision of Q.X.; Q.X. contributed to the revision of the manuscript and provided funding and materials for the experiments. C.H. and Z.X. wrote and revised the manuscript, and all authors reviewed and approved the final version.

Peer review

Peer review information

Nature Communications thanks Longda Jiang and Ruidong Xue for their contribution to the peer review of this work. A peer review file is available.

Data availability

scRNA-seq data from cells carrying triple mutations (TERT^C228T, TP53^G743A, and PDGFRA^D842V)³¹ were obtained from the NCBI Gene Expression Omnibus (GEO) under accession ID GSE229901. scRNA-seq data of DSP-knockout cells²⁹ were obtained from GEO under accession GSE189068. scRNA-seq data of IRF4^p.T95R B cells and PBMCs³ were retrieved from GEO under accessions GSE215936 and GSE215938, respectively. PBMC scRNA-seq following IFN-β treatment³⁸ were accessed from NCBI GEO database accession GSE96583. COVID-19 single-nucleus RNA-seq (snRNA-seq) datasets⁴² were accessed from GEO under accession GSE171524, and PBMC bulk RNA-seq datasets⁴⁷ were accessed under GSE217948. COVID-19 spatial transcriptomics data⁴⁵ were obtained from the crost database⁶⁶ (accession ID VISDS000628). scRNA-seq data of PANC-1 cell lines treated with gemcitabine³⁰ were obtained from GEO under accession GSE186960. Bulk RNA-seq data of PANC-1 cells treated with gemcitabine, MitoBlock-10, or their combination, as well as the phosphoproteomics data of gemcitabine-treated PANC-1 cells, have been deposited in the Genome Sequence Archive under project PRJCA055925. An example dataset has been deposited in Zenodo and is available at 10.5281/zenodo.18767663⁶⁷. Source data are provided with this paper.

Code availability

scDNS is implemented as an R package and is available on GitHub (https://github.com/xiaolab-xjtu/scDNS), under the CC BY 4.0 license. The specific version of the code associated with this publication is archived in Zenodo and is accessible via 10.5281/zenodo.18760951⁶⁸.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-71507-8.