dioscRi enables transferable prediction of clinical outcomes in multi-parameter cytometry data

Abstract

Multi-parameter cytometry technologies enable high-dimensional analysis of immune cell populations at single-cell resolution. Deep learning has been widely applied to these datasets, but existing methods often struggle with transferability across datasets due to technical variability, batch effects and identification of biologically relevant cell populations, limiting their utility in clinical research. We present dioscRi, a transferable deep learning framework that integrates a maximum mean discrepancy variational autoencoder for normalization and de-noising, enhancing cross-dataset compatibility. Changes in cell type proportions and marker expression are identified by structuring these features within biologically or empirically derived cell type hierarchies. These hierarchies are incorporated directly into an overlapping group LASSO model, improving the prediction of clinical outcomes. When applied to a coronary artery disease study, dioscRi recapitulated several known immune associations. Benchmarking across multiple datasets demonstrated dioscRi’s ability to transfer across cohorts with compatible marker panels and outperform existing methods on three of four datasets, establishing it as an interpretable tool for cytometry data analysis.

Cytometry-based prediction models often fail to generalise across patient cohorts. Here, the authors present dioscRi, a framework that enables interpretable prediction of clinical outcomes from immune cell populations across independent cohorts.

Introduction

Multi-parameter flow and mass cytometry technologies enable high-dimensional profiling of millions of cells in a single patient sample, a capability that can support a range of clinical applications, including diagnostics, prognostics, and therapeutic monitoring^1–3. These platforms can measure dozens of markers across millions of cells per sample, providing deep insights into immune states, tumor microenvironments, and signaling dynamics across diseases^4,5. However, translating these insights into clinical practice requires computational methods that can scale to large datasets, remain stable in the presence of technical variability, and generalize across cohorts without retraining. To be clinically actionable, such methods must also produce outputs that are interpretable to researchers and clinicians.

Deep learning encompasses a range of methods that have shown promise for cytometry analysis, owing to their ability to model the complexity of cellular systems. Architectures such as autoencoders (AE), variational autoencoders (VAE), and convolutional neural networks (CNN) are commonly used to extract meaningful representations from high-dimensional data while remaining relatively robust to noise and variability⁶. For example, MoE-Sim-VAE⁷ combines VAEs with Gaussian mixture models to enable unsupervised identification of cell types, implicitly separating biological variation from technical artifacts. DGCyTOF⁸ and DeepCNN⁹ apply autoencoders for supervised cell-type classification, with DeepCNN incorporating a calibration step to address batch effects across datasets. However, such calibration requires prior knowledge of batch structure, limiting its applicability in clinical settings where classification of a sample should be done independently of other samples. For patient classification, models such as CellCNN¹⁰ and DeepCNN use CNNs to learn complex relationships directly from single-cell measurements. Recognizing the importance of interpretability, DeepCNN also employs post-hoc permutation tests and decision trees to identify marker-level associations. Because these interpretability steps are applied after model training, they do not constrain the model during learning and may not reflect the true decision boundaries. As a result, existing approaches are not designed to generalize to unseen datasets and typically lack mechanisms for structured interpretation, making them difficult to apply confidently across cohorts or translate into clinical insight.

A longstanding solution to the challenges of interpretability and technical variability in cytometry has been the use of cell-type hierarchies^11,12. These hierarchies are typically constructed through manual gating, a conventional technique in fluorescence and mass cytometry that defines cell populations by drawing boundaries on a series of two-dimensional marker expression plots, which, taken together, encompass a complex phenotype including expression of multiple markers. Analysts begin by identifying broad immune categories, such as CD4 and CD8 T cells, NK cells, and B cells, and progressively refine these into more specific subsets such as regulatory versus conventional CD4 T cells, naïve versus memory conventional CD4 T cells, and so on. This hierarchical structure not only facilitates interpretation but also provides robustness to technical variation, as comparisons between child and parent populations help normalize differences across samples. Manual gating remains widely used because it reliably identifies cell subpopulations with known biological functions while effectively accounting for technical variability. However, it requires considerable expertise and is very time-consuming, especially when sample-to-sample variation from biological or technical sources requires that multiple gates be adjusted for each sample. Hence, the search for computational methods that can replicate and extend the complex cell subset structures within data derived from increasingly high-plex technologies^13,14. Recent approaches aim to connect cell populations to biological or clinical outcomes by leveraging hierarchical relationships in ways that preserve interpretability and improve scalability^15,16. Embedding hierarchical relationships into deep learning models could improve their resilience to technical variation while preserving biologically meaningful structure.

To address the need for models that are both interpretable and applicable to new samples without retraining the normalization component, we present dioscRi, a deep learning framework for predicting clinical outcomes from multi-parameter cytometry data. DioscRi combines a maximum mean discrepancy variational autoencoder (MMD-VAE) to learn a normalization scheme, with an overlapping group LASSO model that incorporates biologically or empirically derived cell-type hierarchies. This design enables the model to normalize and classify new samples without requiring batch labels, while producing structured outputs that reflect known immune relationships. We demonstrate the utility of dioscRi through its application to a coronary artery disease cohort, where it recapitulates a subset of previously published immune associations¹⁷. We further benchmark its performance across multiple public datasets, showing that dioscRi outperforms existing deep learning approaches on three of four datasets and performs comparably on the fourth. Together, these results establish dioscRi as a robust framework for clinically focused cytometry analysis.

Results

Introducing dioscRi

Here we present dioscRi, an interpretable framework for analyzing multi-parameter cytometry data. This framework integrates an MMD-VAE with a hierarchical group LASSO tailored for the prediction of clinical outcomes (Fig. 1). The input to the framework consists of raw gene or protein (markers) expression data of all samples, where rows represent individual cells and columns correspond to markers. The model output can include normalized expression data, unsupervised cell clusters, sample-level clinical outcome prediction, and the associations of cell-type proportions and/or markers with clinical outcomes. The MMD-VAE provides a transferable normalization scheme, ensuring it can be applied to unseen data. The overlapping group LASSO uses a biologically or empirically derived grouping of cell-type proportions and marker means per cell-type to accurately predict clinical outcomes and uncover associations. The model outputs a tree structure that can assess the associations of these proportions and marker means within cell-types with predictions. In addition, the framework allows for the inclusion of additional clinical information, such as age and gender, which can enhance the precision of clinical outcome predictions.

Fig. 1. Overview of dioscRi.

The framework uses (i) a robust maximum mean discrepancy variational autoencoder (MMD-VAE) to train a normalization filter scheme that can be applied to new data; (ii) linear discriminant analysis trained using either manually gated cell populations or unsupervised clustering to classify cell types; (iii) a hierarchical group least absolute shrinkage and selection operator (LASSO) that uses either a biologically defined or computationally derived tree to combine predicted cell type or cluster proportions and marker means in each sample to predict clinical outcomes; and (iv) a visualization of the associations between cluster or predicted cell type features and clinical predictions using heatmaps and the tree structures.

dioscRi effectively normalizes unseen samples

To assess the effectiveness of dioscRi’s normalization, we applied it to a mass cytometry study of coronary artery disease and evaluated its ability to reduce technical variability across samples. A discovery cohort of 111 samples was used to train the normalization Variational Autoencoder, which was then applied to a validation cohort of 58 samples. Once trained, the model was applied consistently across both cohorts to ensure a unified normalization strategy. The impact of normalization was first evaluated by examining co-expression patterns using biaxial scatter plots of CD3 and HLA-DR for samples 87 and 88 (Fig. 2A–D). In the raw data, broad and overlapping distributions indicated sample-to-sample inconsistency. Following normalization, these distributions became more compact and well separated, supporting improved consistency in population structure and enhancing comparability across samples for downstream analyses. Next, density plots of CD3 (Fig. 2E and F) revealed substantial improvement following normalization. In the raw data (Fig. 2E), distributions varied notably between samples, suggesting technical variability. After applying dioscRi (Fig. 2F), marker distributions became more closely aligned, reducing non-biological variation and preserving meaningful signal.

Fig. 2. Effect of dioscRi normalization on marker expression, cell type recovery and coronary artery disease (CAD) classification in the BioHEART-CT cohorts.

A–D Biaxial scatter and E, F density plots illustrate the effect of normalization on marker expression and gating consistency in samples 87 (validation cohort) and 88 (discovery). In the raw data (A, C, E), marker distributions are broad, clusters are poorly defined, and CD3 alignment is inconsistent, particularly in the validation cohort sample. After dioscRi normalization (B, D, F), distributions become tighter, cell populations separate more clearly, and CD3 expression aligns across samples. Red dashed lines indicate gating regions for specific subsets. G Coefficient of determination (R²) values from the relationship between predicted cell types and each marker for three normalization strategies and the raw data. H Coefficients of variation for each marker before and after normalization, expressed for each normalization method as a ratio to the raw, unnormalized data. n = 27 markers; the coefficient of variation for each marker was computed across the 58 biologically independent samples of the BioHEART-CT validation cohort, each from a different subject. Markers are the unit of study, the raw data provide the reference condition, and the three groups compared are dioscRi, cyCombine, and iMUBAC. Box plots show the center line as the median, the box bounds as the 25th and 75th percentiles, the whiskers as the most extreme values within 1.5 times the interquartile range, and the minima and maxima as the lowest and highest values within the whisker range; points beyond the whiskers are plotted individually. (I) Adjusted Rand index (ARI) evaluating whether clustering on the different normalized datasets recovers populations similar to the manually gated cell types. Area under the receiver operating characteristic curve (AUC) for models classifying CAD, showing improved patient classification in J models trained using cell types predicted from manual gating and K models trained with clustered cell populations. L Comparison of CAD classification reveals a higher AUC for dioscRi normalization compared with cyCombine or iMUBAC normalization. Source data for panels J–L are provided as a Source Data file.

dioscRi normalization reduces variance and improves cell annotation

We next assessed the effect of normalization across all markers using three quantitative metrics that complement the qualitative assessment of the biaxial scatter plots. We first computed the R² from linear models of marker expression against predicted cell type (Fig. 2G). DioscRi consistently produced the highest R² values relative to raw data and other methods, indicating effective noise reduction while preserving the biological signal. Next, comparisons of relative coefficient of variation (CV) (Fig. 2H) showed that dioscRi modestly reduced variability (median = 0.92) while maintaining a tight distribution across markers; cyCombine showed no change (median = 1.00), and iMUBAC achieved the largest reduction (median = 0.45), which, when interpreted in the context of the R² results, suggests this is driven by loss of meaningful biological heterogeneity. We then confirmed this balance at a more granular level by performing marker-level analysis of variability and signal intensity across samples (Supplementary Fig. 1). In the raw data, markers such as CD3, CD4, and HLA-DR exhibit higher variance, consistent with both biological reality and potential technical artifacts. In combination, these results indicate that dioscRi removes sufficient technical noise to stabilize CV without over-compressing true variation.

To assess the impact of normalization on cell type identification, we then performed clustering on the data that had been normalized by the different approaches. Models were also trained with either unnormalized or normalized data using the 11 manually gated cell populations in the discovery cohort to predict cell types in the discovery and validation cohorts. We assessed the concordance of the clustering and predicted cell types using the adjusted Rand index (ARI; Fig. 2I). DioscRi achieved an ARI of 0.83, outperforming cyCombine (ARI = 0.62), iMUBAC (ARI = 0.65), and the raw data (ARI = 0.62). The Rand index value of 0.83 for dioscRi indicates that the gated and clustered cell types would have biologically significant differences. However, in comparison to cyCombine and iMUBAC, the results underscore dioscRi’s capacity to preserve biologically meaningful structure even without explicit batch labels by producing clustering results that align with known cell-type assignments.

dioscRi normalization improves patient classification

To evaluate the impact of normalization on predictive performance, we applied dioscRi to a coronary artery disease dataset and compared outcomes using normalized and unnormalized data. Classifiers to predict coronary artery disease (CAD) were constructed that used predicted cell type proportions and marker means as input features, as well as the biologically-informed hierarchy of the manually gated cell populations. After training models on the discovery cohort, normalization led to substantial improvements in predictive performance, with AUC in the validation cohort improving from 0.68 to 0.79 (Fig. 2J). To demonstrate that the normalization performance benefits of dioscRi are not solely dependent on expertly derived hierarchical relationships, we then derived 11 cell populations from unsupervised clustering with FuseSOM¹⁸ and an empirical hierarchical tree as described in the methods. For this clustering-based model, normalization increased the AUC from 0.67 to 0.83 (Fig. 2K). Finally, we compared the benefit of dioscRi’s transferable normalization with the two established batch correction methods, cyCombine and iMUBAC. Both cyCombine and iMUBAC used information from the discovery and validation sets for normalization, while dioscRi only used the discovery data to train its normalization. Following normalization with each method, the resulting data were processed through the full dioscRi pipeline, and predictive performance was assessed using AUC scores for CAD status. Our dioscRi (AUC = 0.79) outperformed cyCombine (AUC = 0.63) and iMUBAC (AUC = 0.55) (Fig. 2L). In combination, these results underscore the strength of dioscRi’s normalization strategy, which outperforms alternatives in downstream subject-level classification without requiring explicit batch labels.

Cell type and marker associations provide complementary biological insights

To assess the interpretability of dioscRi, we first examined the coefficients from the overlapping group LASSO model using different feature sets of predicted cell types from models trained on expert gating. Fitting separate models using only proportions or only marker means provided complementary views: the proportions model identified differential abundance of B cells, pDCs, and CD4+ T cells (Fig. 3A), of which only pDCs were present differentially in the original gated data¹⁷ while marker means highlighted expression-level differences within other cell populations, including increased CD192 and CD194 in Tregs (Fig. 3B). However, the majority of the differences highlighted in the marker means analysis were not present in the original gated data (Supplementary Fig. 2A). For example, CD4+ Treg cells do not express CD14, and additionally CD141+ DCs do not express KLRG1. The combined model revealed some associations consistent with prior findings in coronary artery disease (CAD)¹⁷, with increases in CD4+ Treg expression of CD194, CD45RO and Ki67 (Fig. 3C). Conventional CD4+ T cells (Tconvs) were negatively associated with CAD through markers such as CD3, CD4, and CD127 (Fig. 3C). This reflects the previously documented changes in Tconv ¹⁷, since the expression of CD3, CD4 and CD127 differs significantly between naive T conv (reduced in CAD) and effector memory Tconv (increased in CAD). If Tconv had been divided into two populations (naïve and memory) in the original 11 cell-type designation, then additional cell type proportions rather than marker means would likely have been identified by the model. This example illustrates a complementary use of marker means: when biologically distinct sub-populations are collapsed in the chosen gating, phenotype-level signal can still surface at the population level, pointing the analyst to where the cell-type definitions could be refined.

Fig. 3. Overlapping group LASSO for the validation cohort identifies cell-type associations with CAD status for the predicted cell-type model trained using the discovery cohort.

A Each coefficient corresponds to a predicted cell type proportion: negative values (green) indicate higher abundance in CAD- samples, whereas positive values (purple) indicate enrichment in CAD+ samples. Marker-level coefficients are shown in (B) and combined cell type proportions and marker means in (C): negative values (blue) associate with CAD-, and positive values (red) associate with CAD+. Combining cell type proportions and marker means enhances patient classification for D models trained using predicted cell types and E models trained using cell-type clusters. Incorporating clinical covariates (age and sex) further improves prediction for F models trained using predicted cell types and G models trained using cell type clusters. Source data for panels D–G are provided as a Source Data file.

The prediction of CAD from models trained using manually gated cell types is likely limited by the ability of dioscRi to predict the hierarchically defined cell types. Heatmaps of the original gated cell types (Supplementary Fig. 3A) and the predicted cell types for the discovery and validation cohorts (Supplementary Fig. 3B and C) illustrate these differences. For example, in both the discovery and validation cohorts, expression of CD14, CD183, CD194, CD38, CD39, and CD56 in the three small DC populations is different from that in the manually gated data. The concordance between the size and marker expression of the original manual gated cell types and the predicted cell types is highest for the major cell types, such as CD4+ Tconv and CD8hi T cells. The initial choice of 11 gated cell types that were more evenly balanced in size would likely have improved the accuracy of the cell type prediction.

We next trained models that used unsupervised clustering in the discovery cohort to define cell types. When applied to the models built using the unsupervised clusters, dioscRi uncovered additional associations not evident from the annotation-guided analysis (Supplementary Fig. 4C). Cluster 7, which had high expression of CD19 and CD20 but also expressed CD1c, CD3, CD4, CD56, and KLRG1 and therefore may comprise a mixture of B cells, DCs, and multiple other cell types (Supplementary Fig. 4D) showed increased CD25 expression (Supplementary Fig. 4B) in CAD+ samples. It is not clear what this represents in biological terms, since cluster 5, whose pattern of marker expression corresponded much better with a pure population of B cells, did not show the same effect of CD25 expression on model performance. Clusters 1 and 4 were characterized by expression of CD56, a marker of NK cells, although their overall pattern of marker expression did not correspond with that of either gated or predicted NK cells (Supplementary Fig. 3). Cluster 4 contributes additional information from 3 markers, but the biological relevance is unclear, since NK cells do not express CD14, while expression of KLRG1 is generally low. Thus, while we will next demonstrate that the grouped LASSO approach provides excellent patient-level distinctions, it is difficult to interpret in biological terms when based on preliminary, uncurated, and unsupervised clustering.

Together, these results illustrate how cell type proportions capture shifts in the abundance of broadly defined populations, while marker means provide finer resolution by revealing differences in cell state within those populations. In this way, dioscRi offers a structured starting point for interpreting complex datasets, using coarse annotations to frame the analysis and marker-level features to uncover additional layers of phenotypic variation, with the understanding that some associations will require further validation.

Evaluating predictive features and immune associations

The dioscRi model that is used for patient classification integrates multiple immune features, including cell-type proportions and marker means per cell type, which are structured within a biologically or empirically informed hierarchy using a grouped LASSO model. To evaluate the contribution of these components to patient classification performance, we compared three versions of the model: one using only proportions, one using only marker means, and the full model combining both. For models trained using predicted cell types based on manual gates and a biologically derived tree structure from the training cohort, the full model combining proportions and marker means achieved the highest AUC of 0.79 for the validation cohort (Fig. 3D) in comparison to the proportions-only model (AUC of 0.69) and the means-only model (AUC of 0.73). As described in the “Methods” section, the validation cohort data were of poor quality, and the prediction of cell types without first applying time-gate filtering to remove significant machine artifacts likely compromised the estimation of proportions. Indeed, the marker means alone generated a higher AUC than proportions, and inclusion of both further improved model performance. In general, cell-type proportions from manually gated data have been more informative for understanding patient-level distinctions. Indeed, in the original published analysis of the BioHEART-CT study¹⁷, there were no significant differences in marker means between the corresponding gated populations in CAD+ and CAD− populations. These results suggest that proportions and marker means provide complementary information that is particularly useful in the analysis of datasets that contain substantial technical artifacts.

We next evaluated the patient classification performance for the models that used clustered cell types to annotate the cells. For these models, which also used an empirically derived tree structure, a different pattern emerged. The proportions-only model performed nearly as well as the full model, with AUCs of 0.78 and 0.83, respectively (Fig. 3E), while the means-only model had a smaller AUC of 0.72. This stronger performance of the proportions-only model suggests that the clustering process captures biologically meaningful variation in population abundances. These proportions may reflect underlying immune shifts more directly than marker intensities alone, particularly when clusters align well with disease-related immune states. Together, these findings indicate that combining proportions and marker means improves prediction under both annotation strategies.

We further evaluated the added value of dioscRi by comparing its performance to models using only clinical variables. In the original BioHEART-CT study that was designed to describe immune contributions to CAD, independent of known risk factors, SVM models based on cytometry data alone (AUC = 0.65) were inferior to those using only age and sex (AUC = 0.82). In contrast, the normalized clustering-based model in dioscRi is comparable to the age and sex model, achieving an AUC of 0.83 (Fig. 3E). Performance improved further when age and sex were included in conjunction with cytometry characteristics, reaching an AUC of 0.87 (Fig. 3G). This improvement demonstrates the added predictive value of properly normalized cytometry data, which is especially important for unsupervised clustering.

dioscRi outperforms state-of-the-art deep learning approaches

We next evaluated the performance of dioscRi in comparison to other deep learning methods, including CellCNN and DeepCNN. We performed this comparison across four cytometry datasets: the BioHEART-CT cohorts, the CMV study SDY519 dataset, the Breast Cancer dataset, and the COVID-19 dataset, each featuring distinct clinical outcomes. For BioHEART-CT, we followed the split previously described¹⁷ where the discovery cohort (111 samples) was used exclusively for training and the validation cohort (58 samples) exclusively for testing. For the CMV dataset, we also followed the split previously described⁹, using SDY519 for testing and the remaining studies for training. The COVID-19 and Breast Cancer datasets were randomly split 70–30 into training and testing sets. Across all datasets, we applied a consistent configuration for dioscRi using 11 clusters, while CellCNN and DeepCNN were used with their respective recommended settings. DioscRi outperformed both CellCNN and DeepCNN in all datasets except CMV SDY519 (Fig. 4). In the BioHEART-CT validation cohort, dioscRi achieved an AUC of 0.83 (95% CI: 0.70–0.93; Supplementary Fig. 9), well above CellCNN (0.60) and DeepCNN (0.62). For CMV, dioscRi achieved an AUC of 0.94 (95% CI: 0.86–0.99). Similarly, dioscRi performed strongly on the Breast Cancer dataset (AUC = 0.89 vs. 0.74 and 0.78) and the COVID-19 dataset (AUC = 0.98 vs. 0.70 and 0.57). The corresponding 95% bootstrap confidence intervals were 0.80–0.97 for Breast Cancer and 0.94–1.00 for COVID-19. For the CMV dataset, both dioscRi and DeepCNN achieved high AUCs (0.94), outperforming CellCNN (0.80). These results underscore dioscRi’s robustness and its transferability across cohorts that share a compatible marker panel, particularly in clinical settings where other models show reduced accuracy. In addition to high performance, dioscRi’s interpretable structure facilitates biological insight, a key advantage over more opaque deep learning approaches in which expression of cell markers cannot be readily mapped back to individual cells.

Fig. 4. Benchmarking of dioscRi against Deep CNN and Cell CNN across four cohorts.

DioscRi was applied to A the BioHEART-CT validation study and three external datasets, B COVID-19, C Breast Cancer, and D CMV, achieving the highest AUC on BioHEART-CT, COVID-19, and Breast Cancer, and an AUC comparable to Deep CNN on CMV. For the breast cancer and COVID-19 datasets, each cohort was split 70% for training and 30% for testing; for the CMV dataset, study SDY519 was reserved entirely for validation. Source data are provided as a Source Data file.

Factors affecting dioscRi performance

The modular nature of dioscRi offers flexibility across datasets, but its performance is shaped by several components of the pipeline. We systematically explored the main tunable hyperparameters (number of clusters, bottleneck layer size, and the number of reference samples K used for MMD-VAE normalization), assessed the robustness of the Frobenius-based reference-sample selection to batch structure, evaluated the pipeline’s sensitivity to the per-sample cell count, and ran a full component ablation to isolate the contribution of normalization, the hierarchical cell-type tree, the choice of model, and the feature set (Supplementary Figs. 5, 6, 7, 8, 10, 11).

The number of clusters was a strong driver of predictive performance. Across datasets such as BioHEART-CT, CMV, and Mathew et al., we observed that AUC generally peaks at intermediate values, typically around 11 or 12 clusters, before plateauing or slightly declining (Supplementary Fig. 5A). This suggests that dioscRi can capture underlying biological structure without overfitting when cluster number is chosen carefully, with cross-validation or heuristic-based approaches likely to help in new datasets.

The bottleneck layer size determined how much information was retained in the normalized representation, with moderate sizes (around 15) consistently yielding strong performance (Supplementary Fig. 5B). Overly small bottlenecks risked losing critical information, while larger ones tended to overfit. This was reflected in a slight upward trend in validation loss at larger sizes despite total loss remaining relatively stable (Supplementary Fig. 5C).

The MMD weight (MMD-λ) balances reconstruction fidelity against latent prior-matching. A training-only sensitivity analysis across all four datasets (Supplementary Fig. 12) shows that reconstruction loss on held-out reference samples is preserved for small MMD-λ and degrades as MMD-λ increases, while the prior-matching (MMD) term decreases monotonically; MMD-λ = 0.01 therefore provides effective regularization at negligible reconstruction cost and serves as a robust default across datasets.

The number of reference samples (K) used in MMD-VAE normalization had a clear optimum. AUC was maximal at K = 2 and did not improve with additional reference samples (Supplementary Fig. 5D). We further examined whether the Frobenius-based selection of these K samples is itself an artifact of batch structure. On both the BioHEART-CT discovery cohort and the multi-study CMV dataset, the most representative samples are drawn from multiple batches/studies rather than concentrated in one, and a permutational multivariate analysis of variance on the covariance distances attributes only 18.6% of covariance variation to batch on BioHEART-CT (42.6% on the more heterogeneous multi-study CMV dataset), with batch not significantly predicting which samples are ranked most representative (permutation p = 0.17; Supplementary Figs. 6 and 7). Downstream performance is moreover robust to the choice of reference samples. Retraining the full pipeline using the most representative, a random, and the least representative tiers yields indistinguishable predictive accuracy on BioHEART-CT, with a modest advantage for the most representative tier only on the more batch-heterogeneous CMV dataset (Supplementary Fig. 8). All three tiers also sit well above two reference baselines included in the same figure that demonstrate the normalization step is necessary, namely unnormalized data (AUC = 0.63) and basic per-sample z-scoring with no learned alignment across samples (AUC = 0.55), confirming that an MMD-VAE-based normalization is required for the reported predictive performance irrespective of which reference samples are used to train it.

We also assessed the sensitivity of the pipeline to the per-sample cell count used for clustering and feature computation. Retraining the full pipeline on BioHEART-CT with downsampling levels from 1000 to 50,000 cells per sample showed that predictive AUC and clustering ARI are stable across the entire range, with runtime and memory roughly flat up to ~10,000 cells and rising steeply beyond it without further gain (Supplementary Fig. 10). Major populations are recoverable from 1000 cells, while rarer populations (pDCs, CD1c+ DCs) require 3000–5000 cells, and CD141+ DCs (<0.05% frequency) are not recovered at any tested depth by naive unsupervised clustering.

To isolate the contribution of each pipeline component, we further evaluated all 24 combinations of MMD-VAE normalization (with/without), hierarchical cell-type tree (with/without), model (group lasso vs. elastic net) and feature set (proportions, means, or both) on the BioHEART-CT validation cohort (Supplementary Fig. 11). Normalization is the dominant factor whenever proportions are included; combining proportions and marker means outperforms either alone; the hierarchical tree has no measurable effect on predictions for this dataset; and group lasso and elastic net are equivalent in predictive performance (AUC 0.825 vs. 0.826), with group lasso retained for its added structured interpretability.

Together, these analyses show that dioscRi’s performance is most sensitive to the number of clusters and the bottleneck size; is robust to the choice and number of reference samples and to the per-sample cell count above a modest threshold; and is most strongly driven by the MMD-VAE normalization step when proportion-based features are used. This provides practical guidance for parameter tuning and supports the use of dioscRi in studies with limited sample sizes or modest per-sample cell counts. Because these analyses retrain the full pipeline across reference-sample and cell-count settings, they also characterize the variability not captured by the bootstrap confidence intervals, which are conditional on the fitted model. In practice, only the group-lasso penalty requires tuning, by BIC on the training data; the remaining settings are used at the defaults reported here. In particular, the number of clusters was fixed at 11 across all datasets, following the number of manually gated populations in BioHEART-CT, and was not tuned per dataset.

Discussion

To achieve interpretable and transferable insights from cytometry data, we created dioscRi, a framework combining deep learning normalization with biologically informed feature selection. DioscRi integrates an MMD-VAE for harmonizing multi-batch data and an overlapping group LASSO that leverages cell-type hierarchies to model immune variation. Applied to a coronary artery disease dataset, dioscRi achieved strong predictive performance and validated a subset of the known associations previously revealed by expert manual gating¹⁷. Across multiple datasets, dioscRi outperformed or matched existing deep learning approaches while preserving interpretability, demonstrating its value for both accurate prediction and biological discovery.

A major strength of dioscRi lies in its use of hierarchical group structures to model both cell type proportions and marker expression. This allows it to exploit the immune system organization to improve interpretability relative to models based on unsupervised clustering. Unlike black-box deep learning models, where feature contributions often require post-hoc explanation tools like SHAP, dioscRi yields coefficients that provide a structured starting point for investigating associations between immune features and clinical outcomes. However, the degree to which these coefficients are biologically interpretable depends on the quality of the upstream cell-type assignments. When cell types are defined by expert gating and correspond to well-characterized populations, the resulting associations can be mapped to known biology. When derived from unsupervised clustering, the biological meaning of individual coefficients may be less clear, as clusters can mix distinct populations or split biologically coherent ones. We therefore emphasize that dioscRi’s interpretability is primarily structural, in that coefficients are organized by cell type and marker within a hierarchy, rather than uniformly biological. While the incorporation of biologically defined hierarchies enhances interpretability by maintaining a degree of alignment with known cell-type relationships, our results also support previous findings¹⁵ showing that empirically derived hierarchies can be similarly effective. This balance between biological grounding and data-driven flexibility enables dioscRi to provide biological insights without compromising predictive performance.

Our analysis of the BioHEART-CT cohort demonstrates how dioscRi can enhance both prediction and biological insight from cytometry data. In the original BioHEART-CT study¹⁷, SVM models based solely on immune features showed modest predictive performance and did not improve upon simple clinical variables such as age and sex. Using overlapping group LASSO models, we observed more stable and accurate predictions, whether cell types were taken from the original manual gating or from unsupervised clustering. These improvements were supported by the framework’s normalization strategy and its structured modeling of cell-type proportions and marker expression. DioscRi reproduced some previously reported associations, such as decreased pDCs in CAD, and identified marker-level signals that had been obscured by the initial choice of cell types. This included features related to the previously known differences in the population sizes of naïve and memory CD4+ Tconvs. Importantly, integrating cytometry features with the clinical variables age and sex further improved predictive performance, illustrating the added value of immune profiling when supported by effective normalization and feature selection.

The complementary use of predicted cell-type proportions and marker means dioscRi captures immune variation at different resolutions. The proportions reflect differences in cell abundance, while the marker means capture changes in phenotype within the predicted cell types. In practice, the line between these two views is shaped by clustering resolution, where higher-resolution clustering can convert marker-level differences into abundance shifts by dividing broader populations into distinct subtypes. Interestingly, the proportion model performed particularly well with unsupervised clusters, likely because the clustering emphasized sample-level heterogeneity in dominant populations. These findings suggest that proportions can serve as strong predictors when clustering effectively captures meaningful immune variation. Although the combined model ultimately performs best, the strong performance of the predicted cell type model highlights its interpretability and value in workflows driven by data-derived cluster structures.

A deliberate design choice in dioscRi is the use of sample-level summary features (cell-type proportions and marker means) rather than single-cell representation learning at the patient level. The single-cell representation-learning baselines we evaluated, CellCNN and DeepCNN, underperformed dioscRi on the BioHEART-CT validation cohort (see the “Results” section). The summary-feature approach also aligns with how cytometry data are conventionally interpreted, enables structured regularization via hierarchical group LASSO, and yields coefficients that are directly interpretable in terms of specific cell populations and markers. Whether single-cell representation-learning approaches offer an advantage in regimes our cross-sectional benchmarks do not cover, such as longitudinal cytometry where the same individuals are sampled at multiple time points, is left to future work.

Despite its strengths, dioscRi has some limitations. While the VAE-based normalization reduces technical variability and enables transfer across batches, it does not explicitly model batch differences. In cases with substantial differences, such as shifts in instrument settings or sample handling, residual batch effects may persist and moderately influence downstream analyses. In addition, subtle panel variations, including different antibody clones or minor protocol changes, can introduce technical variability in marker expression that is difficult to distinguish from the biological signal. Although dioscRi’s latent representation is robust to many such differences, these sources of variation could impact performance when applying the model across datasets with substantial experimental heterogeneity. In practice, careful study design and harmonization of upstream protocols can mitigate many of these effects. Further, as a regression-based framework, the overlapping group LASSO underlying dioscRi extends naturally to multi-class (multinomial) and continuous (Gaussian) outcomes, and can accommodate co-morbidities through additional covariates, as already demonstrated by the age + sex inclusion. However, the current evaluation focuses on binary clinical outcomes. Performance under multi-class or continuous outcomes, co-morbidities, or complex clinical phenotypes with varying presentations at different disease stages has not been formally evaluated and represents an important direction for future work.

The transferability of dioscRi applies within matched or overlapping marker panels, where the MMD-VAE encoder and decoder normalize new samples in the same marker space without retraining, and the downstream components are fitted on the training data and applied within that same panel. Extending dioscRi to operate across heterogeneous panels is an exciting next step, building on recent advances in adjacent single-cell modalities where methods such as scArches¹⁹ learn shared latent representations that can be transferred and incrementally refined across datasets with different feature sets. Analogous marker-agnostic or shared-latent architectures applied to cytometry would let dioscRi integrate and predict across panels that overlap only partially, substantially broadening its reach across the rapidly expanding catalog of multi-parameter cytometry assays.

In summary, dioscRi offers a robust and interpretable framework for extracting biologically relevant insights from cytometry data and applying them to clinically important prediction tasks. By integrating transferable deep learning-based normalization with structured modeling of cell type features, it overcomes key limitations of existing approaches and improves both performance and interpretability. Across diverse datasets, dioscRi demonstrates consistent predictive strength. These capabilities position dioscRi as a practical tool for advancing immune profiling in translational research and clinical applications.

Methods

This research complies with all relevant ethical regulations; this study is a computational reanalysis of previously published, publicly available datasets, and ethical approval and informed consent were obtained by the original investigators as reported in the corresponding publications^17,20–23.

Overview of dioscRi

dioscRi consists of various steps: (i) transferable data normalization and denoising using a maximum mean discrepancy variational autoencoder (MMD-VAE), (ii) generation of cell types through unsupervised clustering, (iii) extraction of cell type and sample-level features, (iv) hierarchical grouping of cell type-level features, (v) clinical outcome prediction using overlap group lasso, and (vi) model visualization and interpretation.

Data pre-processing

For all datasets, we applied an arcsinh transform (y = arcsinh(x/5)). Following the transformation, each data set was sub-sampled to 10,000 cells from each sample and combined for analysis.

Before training the MMD-VAE model, a Min–Max scaling to [0, 1] was applied.

$$y=\frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)}$$ 1

Min–max scaling is commonly employed to improve the training efficiency and convergence of deep learning models.

Maximum mean discrepancy variation autoencoder architecture

The dioscRi architecture integrates an encoder and decoder within a variational autoencoder (VAE) framework. It is enhanced by the maximum mean discrepancy (MMD) to better match the latent space distribution to the prior, improving adaptability to non-Gaussian data distributions^24,25. The maximum mean discrepancy (MMD) penalty was implemented to enforce distributional similarity between the encoded latent variables and a prior distribution, following the method described in ref. ²⁶. An unbiased U-statistic estimator of the MMD was computed using an inverse multi-quadratic (IMQ) kernel. We summed IMQ kernels computed at varying scales to capture multi-scale relationships. The encoder compresses the input data into a lower-dimensional representation, thereby retaining only the most pertinent information by filtering out noise. The decoder then reconstructs this encoded data to its original form by minimizing the difference between the original input and the decoded output.

The model can be formulated as

$$z=f_{\mathrm{encoder}}\left(x\right)$$ 2

$$x^{\prime }=f_{\mathrm{decoder}}\left(z\right)$$ 3

$${\mathcal{L}}_{\mathrm{BCE}}=-\mathbb{E}\left[x\log \left(x^{\prime }\right)+\left(1-x\right)\log \left(1-x^{\prime }\right)\right]$$ 4

$${\mathcal{L}}_{\mathrm{MMD}}={\mathbb{E}}_{z,z^{\prime }~q_z}\left[k\left(z,z^{\prime }\right)\right]+{\mathbb{E}}_{z,z^{\prime }~p_z}\left[k\left(z,z^{\prime }\right)\right]-2{\mathbb{E}}_{z~q_z,z^{\prime }~p_z}\left[k\left(z,z^{\prime }\right)\right],$$ 5

$${\mathcal{L}}_{\mathrm{VAE}}={\mathcal{L}}_{\mathrm{BCE}}+\lambda \cdot {\mathcal{L}}_{\mathrm{MMD}}$$ 6

Here, f_encoder is the encoder function that compresses the input data x into a lower-dimensional latent representation z. The decoder function f_decoder reconstructs the input data from z back to its original form $x\prime$. The binary cross-entropy loss ${\mathcal{L}}_{\mathrm{BCE}}$ measures the difference between the input and the reconstructed output. For the MMD loss, k( ⋅ , ⋅ ) is the IMQ kernel. This penalty encourages the encoded latent representation to align with the desired prior distribution while allowing flexibility for modeling complex data distributions. The overall loss ${\mathcal{L}}_{\mathrm{VAE}}$ combines the binary cross-entropy loss ${\mathcal{L}}_{\mathrm{BCE}}$ with the MMD regularization term ${\mathcal{L}}_{\mathrm{MMD}}$, balanced by a hyper-parameter λ that controls the influence of the MMD term.

MMD-VAE implementation

The model architecture includes an encoder that reduces the dimensionality of the input data through dense layers with ReLU and GELU activations, followed by a decoder that reconstructs the input using sigmoid activation. The training objective optimizes a total loss, combining binary cross-entropy for reconstruction and maximum mean discrepancy (MMD) to align the latent space distribution with a standard normal prior.

The MMD-VAE model is implemented in R using the Keras library²⁷, with RMSprop²⁸ as the optimizer and a default learning rate of 0.001. The MMD-VAE is trained for 100 epochs with a batch size of 32, and the latent dimension is conservatively set to balance information retention. The model is validated using a separate dataset during training. The MMD-VAE comprises three hidden layers whose sizes are set deterministically from the number of input markers D as (D−4, D−8, D−11). Table 1 lists the hyperparameters used for each dataset.

Table 1.

Hyperparameters used for each dataset when fitting dioscRi MMD-VAE normalization and overlap-group lasso

Name	Layer sizes	Number of clusters	Cells sampled	MMD-λ	Batch size	Input size	α	λ
Wagner et al - Breast Cancer	26-22-19	11	10,000	0.01	32	30	0.8	0.0500950095009501
CMV - Study SDY519	19-15-12	11	10,000	0.01	32	23	0.4	0.057030703070307
Bioheart-CT Discovery	23-19-16	11	10,000	0.01	32	27	1.0	0.0516151615161516
Mathew et al - COVID-19	12-8-5	11	10,000	0.01	32	16	0.6	0.0535153515351535

All hyperparameters are set using the training data only; the validation and external test cohorts are reserved exclusively for final evaluation. The MMD weight (MMD-λ) is a fixed light-regularization constant (MMD-λ = 0.01) applied uniformly across datasets, and the MMD-VAE is monitored during training on a held-out subset of reference samples. The number of clusters is fixed a priori to the number of manually gated populations in BioHEART-CT, and the batch size, number of epochs, and cells sampled per sample are fixed defaults. The overlapping group-lasso penalty parameters α and λ are the only hyperparameters formally optimized, and are selected on the training data by minimizing the Bayesian information criterion.

Selection of a representative sample for MMD-VAE training

To reduce the total size of the training data while retaining core biological information, we select representative samples for training the MMD-VAE model. We ensure the model captures key biological variations without redundancy by selecting samples with the most representative characteristics. To achieve this, we use the method described in ref. ²⁹, where each sample is evaluated based on its average covariance Frobenius norm between it and other samples.

The intuition behind this criterion is that the sample-level covariance matrix encodes the joint marker structure of a sample, and a sample whose covariance lies near the centroid of all sample covariances (i.e., has the smallest average pairwise Frobenius distance) is one whose joint marker structure is broadly typical of the cohort. Training the MMD-VAE to reconstruct such a sample, therefore, biases the latent space toward representations that generalize across the cohort rather than capturing the idiosyncratic structure of outlier samples or single batches. This is consistent with the principle of second-order distribution matching used in domain adaptation, where aligning covariance statistics has been shown to be an effective surrogate for full distribution matching³⁰.

The Frobenius norm quantifies the difference between the covariance matrices of pairs of samples, effectively measuring how similar or different each sample is to others. Formally, the Frobenius norm between samples i and j is defined as

$${\mathrm{FN}}_{ij}={∥{\mathrm{Cov}}_i-{\mathrm{Cov}}_j∥}_F=\sqrt{{\sum }_{k=1}^m{\sum }_{l=1}^m{\left[{\left({\mathrm{Cov}}_i-{\mathrm{Cov}}_j\right)}_{kl}\right]}^2},$$ 7

where Cov_i and Cov_j are the covariance matrices for samples i and j, respectively, m is the number of markers, and ${\left(\cdot \right)}_{kl}$ denotes the $\left(k,l\right)$ entry of a matrix. The stable sample is selected by returning the sample that satisfies:

$$\arg {\min }_i\left(\frac{1}{n}{\sum }_{j=1}^n\left({\text{FN}}_{ij}\right)\right)$$ 8

Using this criterion, we select the top K samples with the minimum average Frobenius norms as the set of reference samples. K was set to 2 in all main analyses.

Unsupervised clustering and cell type classification

To generate cell types from the normalized training data, we use the FuseSOM algorithm¹⁸, which combines multiple similarity metrics through multiview ensemble learning and hierarchical clustering to create cell types.

For the testing data, cell types were generated by training a classifier on the cell types identified in the training data. A linear discriminant analysis (LDA) model was used for all cell type classifications. The embeddings from the VAE model served as features for unsupervised clustering and manual gating. Cell type classification models were fit using the Caret R package.

The cell-typing step of dioscRi is modular: any cell-typing strategy that produces per-cell labels can be plugged in here, whether expert manual gates, unsupervised clustering, or reference-based annotation. For illustration we use the FuseSOM package together with a simple LDA classifier to assign labels on new samples, and where expert manual gates are available (e.g., from the original BioHEART-CT study) we use those directly.

Feature engineering

After cell type identification, we generate a set of features, including the logit of the cluster proportions in each sample and the mean marker expression per cell type in each sample. The logit proportion for a cluster j in sample i is defined as

$${\pi }_{ij}=\log \left(\frac{p_{ij}}{1-p_{ij}}\right)=\log \left(\frac{\frac{c_{ij}}{k}}{1-\frac{c_{ij}}{k}}\right),$$ 9

where n is the number of samples, k is the number of cells in each sample, m is the number of clusters, and c_ij is the count of cells in sample i for i = 1, 2, …, n that belong to cluster j for j = 1, 2, …. , m.

The mean marker expression per cell type across all samples is defined as

$${\overline{x}}_{ij}=\frac{1}{n_{ij}}{\sum }_{k=1}^{n_{ij}}{x}_{ijk},$$ 10

where x_ijk is the marker expression value of the kth cell belonging to cell-type j and sample i. The value $n_{ij}$ is the total number of cells belonging to cell-type j in sample i.

Incorporating empirically or biologically informed grouping structure

Before classifying samples, we integrate cell type hierarchies into the proportions, using either biologically informed or empirically derived groupings. To empirically derive groupings, we construct a hierarchical clustering tree based on Pearson correlation distance and Ward linkage, generating all possible groupings by cutting the tree at various heights. Alternatively, biologically derived hierarchies are predefined using prior knowledge of cell-type relationships by respective domain experts. These hierarchies, whether empirically or biologically informed, are combined with the marker means for each cell type to create the final set of features used for classification.

Group logistic Lasso with overlapping groups

The Lasso (least absolute shrinkage and selection operator) is a regularization method used for feature selection and model regularization³¹. Given a matrix X of predictors with n samples and p features, and a response vector y, the Lasso minimizes the following objective function:

$$Q\left(\beta \mid X,y\right)=\frac{1}{2n}{\sum }_{i=1}^n{\left(y_i-X_i\beta \right)}^2+\lambda {\sum }_{j=1}^p{∥{\beta }_j∥}_1$$ 11

where β represents the coefficient vector, ${\sum }_{j=1}^p{∥{\beta }_j∥}_1$ is the ℓ₁-norm penalty, and λ is a regularization parameter that controls the level of sparsity in the solution. By penalizing the absolute values of the coefficients, Lasso encourages sparsity, driving some coefficients to zero and excluding irrelevant features. This approach is particularly effective in high-dimensional settings, where it reduces overfitting and yields interpretable models.

The Group Lasso extends this framework by promoting sparsity at the group level, enabling feature selection based on predefined groups of predictors³². Assume that the predictors are partitioned into J non-overlapping groups G₁, G₂, …, G_J, with each group containing a subset of features. The Group Lasso objective is

$$Q\left(\beta \mid X,y\right)=\frac{1}{2n}{\sum }_{i=1}^n{\left(y_i-X_i\beta \right)}^2+\lambda {\sum }_{j=1}^J\sqrt{K^j}{∥{\beta }_{G_j}∥}_2$$ 12

where ${∥{\beta }_{G_j}∥}_2$ denotes the ℓ₂-norm of the coefficients within group G_j, and K^j is the number of elements in group j.

We adapt the Group Lasso with a logistic loss function for binary outcomes to handle classification. Given a binary response vector y with entries y_i ∈ {0, 1}, the logistic loss is defined as

$$\mathcal{L}\left(\beta \mid X,y\right)=-\frac{1}{n}{\sum }_{i=1}^n\left[y_i\log \left(p_i\right)+\left(1-y_i\right)\log \left(1-p_i\right)\right]$$ 13

where

$$p_i=\frac{1}{1+\exp \left(-X_i\beta \right)}$$ 14

represents the probability of y_i = 1 given X_i. The Group Lasso with logistic loss thus becomes:

$$Q\left(\beta \mid X,y\right)=-\frac{1}{n}{\sum }_{i=1}^n\left[y_i\log \left(p_i\right)+\left(1-y_i\right)\log \left(1-p_i\right)\right]+\lambda {\sum }_{j=1}^J\sqrt{K^j}{∥{\beta }_{G_j}∥}_2$$ 15

This formulation encourages sparsity in the group structure, making it well-suited for binary classification tasks involving predefined groups of features.

The Overlapping Group Lasso further generalizes this approach by allowing predictors to belong to multiple groups³³. This extension is particularly useful in cases where variables naturally fall into overlapping categories, such as immune cell features that are shared across related cell types within a hierarchy. To accommodate overlapping groups, we decompose the coefficient vector β into group-specific latent vectors γ_j, each associated with a group G_j. The design matrix X is then expanded into $\tilde{X}$ by duplicating columns for variables that appear in multiple groups, so that each group’s coefficients can be estimated independently. The objective for the Overlapping Group Lasso is

$$Q\left(\gamma \mid \tilde{X},y\right)=L\left(y,\tilde{X}\gamma \right)+\lambda {\sum }_{j=1}^J\sqrt{K^j}{∥{\gamma }_j∥}_2$$ 16

where $L\left(y,\tilde{X}\gamma \right)$ denotes the logistic loss, γ = (γ₁, …, γ_J) represents the concatenated vector of group-specific coefficients, and ∣∣γ_j∣∣₂ applies the ℓ₂-norm penalty within each group. The overlapping group structure enables shared predictors to have multiple group-specific effects, enhancing model flexibility. After fitting the model, the original coefficients β are reconstructed as

$$\beta ={\sum }_{j=1}^J{\gamma }_j$$ 17

To further regularize overlapping cases, we add a ridge penalty term, defining the final objective as

$$Q\left(\gamma \mid \tilde{X},y\right)=L\left(y,\tilde{X}\gamma \right)+{\lambda }_1{\sum }_{j=1}^J\sqrt{K^j}{∥{\gamma }_j∥}_2+\frac{{\lambda }_2}{2}{∥\beta ∥}_2^2$$ 18

where λ₁ = αλ and λ₂ = (1 − α)λ. To select the optimal α and λ, we fit models over a grid of ten α values, {0.1, 0.2, …, 1.0}, and a fixed grid of 10,000 λ values evenly spaced over [0.05, 1]. For every (α, λ) pair, we compute the Bayesian information criterion, $\mathrm{BIC}=\mathrm{deviance}+\mathrm{df}\cdot \log n$. For each α, the optimal λ is selected via the elbow method applied to its BIC sequence, and the final (α, λ) is the combination minimizing BIC across all α values. The selected dioscRi hyperparameters for each dataset are provided in Table 1. We implemented the overlapping group lasso using the grpregOverlap R package³³, modified to support logistic loss for binary classification. This approach provides a flexible and interpretable solution for feature selection when variables belong to multiple groups.

In the overlapping group lasso model, each coefficient vector γ_j represents the contribution of a specific group G_j to the overall model, allowing for a unique interpretation of overlapping features. Since $\beta ={\sum }_{j=1}^J{\gamma }_j$, the original coefficient vector β is constructed by summing the group-specific latent effects γ_j. This decomposition means that each predictor can be influenced by multiple groups simultaneously, with each γ_j capturing the effect of group G_j independently. Consequently, the magnitude and direction of γ_j can be interpreted as the specific impact of group G_j on the predictor, highlighting how particular groupings contribute to the model’s predictive power. When interpreting the model, non-zero γ_j values indicate that the group G_j plays a significant role, while the summed γ coefficients in β provide the overall effect of each predictor. This approach improves interpretability by revealing which features are selected and which groups drive their inclusion, offering insights into overlapping pathways or shared structures across groups.

Implementation of other approaches for benchmarking comparisons

Training of cyCombine and iMUBAC

We benchmarked the normalization performance of dioscRi against other normalization methods that explicitly consider batch information. These methods included cyCombine³⁴ and iMUBAC³⁵. For analysis, both cyCombine and iMUBAC were run with default parameters.

Training of CellCNN and DeepCNN

The predictive performance of dioscRi was benchmarked against CellCNN¹⁰ and DeepCNN⁹. Since both methods work directly on FCS files, after splitting each dataset into training and testing, the resulting cells were written back to FCS files for predictive modeling using these methods.

CellCNN was run using the hyperparameters recommended in its original publication (filter number = 5, l₂ coefficient = 1e−4). The Adam optimizer was used for loss optimization, with a learning rate of 0.01 over 500 epochs. After training, the model was evaluated using the testing set.

DeepCNN was run using the hyperparameters recommended in its original publication (number of filters = 3, dropout rate equals 0.2). The Adam optimizer was used for loss optimization, with a learning rate of 0.01 over 500 epochs. After training, the model was evaluated using the testing set.

Datasets

We employed a variety of datasets across different technologies and diseases to evaluate the performance of dioscRi. Brief descriptions of each dataset are provided below, with more details outlined in Table 2.

Table 2.

Benchmark datasets

Name	Technology	Description	No. of samples	Clinical outcome
Wagner et al.—Breast Cancer	CyTOF	Predicting tumor in breast cancer samples	194	Tumor/non-tumor
CMV—Study SDY519	CyTOF	Predicting positive vs. negative CMV titer results in influenza samples	472	Positive/negative
Bioheart-CT Discovery	Mass cytometry	Discovery study for BioHEART-CT	111	Gensini_bin = 1/0
Bioheart-CT Validation	Mass cytometry	Validation study for BioHEART-CT	58	Gensini_bin = 1/0
Mathew et al.—COVID-19	Flow cytometry	Profile of CD8+ non-naive T cells to distinguish recovered from COVID-19 vs. healthy	214	COVID-19 recovered/healthy

Five published datasets were used to benchmark and compare the performance of dioscRi to other deep-learning models. The name refers to each dataset throughout the manuscript.

BioHEART-CT

To evaluate the dioscRi framework, we used mass cytometry data generated from the BioHEART-CT study^17,20. This prospective longitudinal cohort study to identify immune cell populations associated with coronary artery disease (CAD). Cell profiling was performed using mass cytometry with 41 tagged antibodies, enabling a detailed analysis of cellular characteristics. The major clinical outcome used in the previously published analysis¹⁷ was the Gensini score, a continuous measure of coronary artery disease (CAD) severity. For prediction tasks, we binarized this outcome: samples with a Gensini score of 0 were labeled as CAD−, and those with a score > 0 were labeled as CAD+. The original study¹⁷ included discovery and validation cohorts that underwent expert manual gating to identify more than 100 subpopulations. The discovery cohort of 111 samples across 6 batches comprising a total of 41,350,484 cells was then analyzed using a logistic regression model to identify 18 subpopulations whose size was identified as most different between CAD+ and CAD−. The sizes of the same 18 subpopulations were manually gated in the validation cohort, consisting of 58 samples across 3 batches. Importantly, the validation cohort was acquired using a mass cytometry platform that introduced significant time-dependent changes in multiple signals³⁶, so the expert gating included preliminary time gates to exclude artifactual data. For the current study, the discovery cohort was downsampled to include 1,106,443 cells. The validation cohort was downsampled to 580,000 cells without applying any time gates. For the discovery cohort, 11 manually gated cell populations, identified as detailed in Supplementary Fig. 4 of the original manually gated analysis¹⁷ were used as the basis for cell type prediction. The 11 populations were conventional CD4+ T cells (Tconv, mean 38%), CD4+ regulatory T cells (Treg, mean 1.9%), CD8+ T cells (CD8hi mean 17%), CD8lo-neg T cells (CD8lo mean 6.2%), B cells (mean 10%), NK cells (mean 10.2%), CD14+ monocytes (mean 12%), CD16+ monocytes (mean 2.1%), plasmacytoid dendritic cells (pDC mean 0.36%), CD1c+ dendritic cells (CD1c DC mean 0.65%), and CD141+ dendritic cells (CD141 DC mean 0.044%). Of the 41 metal-tagged antibodies used to assess marker expression in the original mass cytometry data, 27 were included as markers in the discovery and validation datasets.

Breast cancer tumor

This dataset comprises 144 human breast tumor samples and 50 non-tumor tissue samples measured with mass cytometry²¹. The primary objective was to characterize the features of breast cancer ecosystems and their correlations with clinical data. The study evaluated 73 proteins in approximately 26 million cells, utilizing tumor-focused and immune cell-centric antibody panels. This extensive profiling helps understand the protein expression patterns and cellular interactions within breast cancer and adjacent non-cancerous tissues. We analyzed 194 samples, predicting each as tumor vs. non-tumor. The dataset was split into 70% training and 30% testing.

Cytomegalovirus

This dataset encompasses data from nine human immunology studies (SDY112, SDY113, SDY305, SDY311, SDY315, SDY472, SDY478, SDY515, and SDY519) featuring 472 peripheral blood mononuclear cell (PBMC) samples from 313 subjects across the nine studies^23,37–39. Diagnosing latent Cytomegalovirus (CMV) is challenging since it is usually asymptomatic with limited observable effects on the immune system. These studies aimed to identify immune responses to latent CMV infections. Our analysis of this dataset followed that used in the DeepCNN manuscript⁹, which used studies SDY519 and SDY515 as testing and validating datasets, and the rest for model training. We held out study SDY519 (60 samples) as the test set. Study SDY515 (79 samples) was not used, as our evaluation did not require a separate validation set in addition to the held-out test study, leaving 333 samples from the remaining seven studies for model training. Clinical response is individuals having a positive or negative cytomegalovirus (CMV) infection. A set of 23 markers common to all nine studies was used.

COVID-19 PBMC CD8+ non-naive T cells

This dataset is derived from a high-dimensional flow mass cytometry analysis of COVID-19 samples, alongside comparisons with recovered individuals and healthy controls²². The study involved an integrated analysis of approximately 200 immune features. These immunological data were combined with about 50 clinical features, providing a comprehensive overview that facilitates understanding of the relationships between the immunology of SARS-CoV-2 infection and various clinical outcomes, including disease patterns, severity, and progression. We restricted analysis to CD8+ non-naive T cells and used COVID-19 recovery as the clinical response of interest. The dataset was randomly split into 70% training and 30% testing sets.

Benchmarking protocol

All approaches were evaluated using the same train/test splits on each dataset, the same marker subsets, and the same preprocessed data for each dataset. The BioHEART-CT and CMV study both had batch information that could be used for train/test splits. For BioHEART-CT, the discovery cohort from the original manuscript served as the training set and the validation cohort as the test set. For CMV, following the original manuscript, batch SDY519 was held out for testing. For the Breast Cancer and COVID-19 datasets, which did not have annotated batch information, a fixed 70/30 random split was used for each dataset.

To quantify uncertainty in the reported AUC values, 95% confidence intervals were computed using a stratified bootstrap of the held-out test predictions. For each dataset, we drew 2000 bootstrap samples from the fixed test/validation predictions, resampling positive and negative classes independently (stratified bootstrap) to preserve the test-set class balance in every resample. For each resample, we recomputed the AUC; the reported interval is the 2.5th–97.5th percentile of the resulting 2000 AUC values. These intervals quantify the sampling uncertainty of the AUC conditional on the fitted model and the fixed held-out predictions; by construction, they do not capture variability from the stochastic pipeline components (MMD-VAE initialization and clustering), the per-sample cell count, or the train/test partition. The pipeline components and per-sample cell count are characterized separately in the sensitivity analyses reported in the Results.

Model evaluation

Various evaluation metrics were used to evaluate components of dioscRi.

Evaluation of normalization by clustering

After data normalization, the adjusted Rand index (ARI) was used to assess clustering performance. The ARI metric assesses the agreement between two clusterings, taking a value of 1 for perfect agreement and 0 for agreement no better than expected by chance. All values reported here lie between 0 and 1. The ARI score can be defined as:

$$\mathrm{ARI}=\frac{{\sum }_{ij}\left(\begin{array}{c}\hfill n_{ij}\hfill \\ \hfill 2\hfill \end{array}\right)-\left[{\sum }_i\left(\begin{array}{c}\hfill a_i\hfill \\ \hfill 2\hfill \end{array}\right){\sum }_j\left(\begin{array}{c}\hfill b_j\hfill \\ \hfill 2\hfill \end{array}\right)\right]/\left(\begin{array}{c}\hfill n\hfill \\ \hfill 2\hfill \end{array}\right)}{\frac{1}{2}\left[{\sum }_i\left(\begin{array}{c}\hfill a_i\hfill \\ \hfill 2\hfill \end{array}\right)+{\sum }_j\left(\begin{array}{c}\hfill b_j\hfill \\ \hfill 2\hfill \end{array}\right)\right]-\left[{\sum }_i\left(\begin{array}{c}\hfill a_i\hfill \\ \hfill 2\hfill \end{array}\right){\sum }_j\left(\begin{array}{c}\hfill b_j\hfill \\ \hfill 2\hfill \end{array}\right)\right]/\left(\begin{array}{c}\hfill n\hfill \\ \hfill 2\hfill \end{array}\right)}$$ 19

where n_ij is the number of pairs of elements that are in the same set in both clusterings, a_i is the total number of pairs in the same set for the first clustering, b_j is the total number of pairs in the same set for the second clustering, and n is the total number of elements.

Evaluation of outcome prediction

The ROC-AUC metric was used to assess the performance of dioscRi and competing methods for predicting clinical outcomes. The metrics are defined using the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values.

The ROC-AUC measures the area under the receiver operating characteristic (ROC) curve, which plots the true positive rate (TPR) against the false positive rate (FPR) at various threshold settings. The AUC score is a single scalar that summarizes the performance across all possible classification thresholds, with 1 indicating perfect classification and a value of 0.5 indicating no better than chance. The formula for TPR and FPR are:

$$\text{TPR}=\frac{\text{TP}}{\text{TP}+\text{FN}}$$ 20

$$\text{FPR}=\frac{\text{FP}}{\text{TN}+\text{FP}}$$ 21

Statistics and reproducibility

All analyses were done in R version 4.5.0. No statistical method was used to predetermine sample size; each dataset was analyzed at the size reported in its originating publication. Cells were downsampled to 10,000 per sample before model fitting. For the CMV dataset, study SDY519 was held out for testing, and study SDY515 was not used, following the benchmarking protocol of the original publication. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

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

Author contributions

E.P. and E.W. conceived and designed the study. E.P. and E.W. led the method development and guided the evaluation data analysis with input from H.M. and J.Y. B.F. generated the original gated discovery and validation cohorts that were then curated by E.W., who implemented all data analytics and developed the corresponding R code. E.P. and E.W. drafted the initial manuscript with writing contributions from S.R., and all authors (E.W., S.R., G.F., J.Y., B.F., H.M., and E.P.) wrote, edited, and reviewed the text. B.F. extensively edited the manuscript and figures to provide additional clarity related to the immunological aspects of the study. All authors approved the final version.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

The following sources of funding for each author are gratefully acknowledged: the AIRinnoHK program of the Innovation and Technology Commission of Hong Kong to J.Y. and E.P. Australian Research Council Discovery Early Career Researcher Award (DE200100944) funded by the Australian Government to E.P.; and the University of Sydney Postgraduate Excellence Award for E.W. The funding sources had no impact on the study design, in the collection, analysis, and interpretation of data, the writing of the manuscript, and in the decision to submit the manuscript for publication.

Data availability

No new data were generated in this study; all datasets analyzed were previously published. The cytomegalovirus data used in this study are available in the ImmPort database under accession codes SDY112, SDY113, SDY305, SDY311, SDY315, SDY472, SDY478, SDY515, and SDY519 (https://www.immport.org). The breast cancer data used in this study are available from Wagner et al.²¹, the COVID-19 data from Mathew et al.²², and the BioHEART-CT data from Kott et al.^17,20, in each case as described in the originating publication. The processed data used to generate the results reported here are available in the Zenodo repository under https://doi.org/10.5281/zenodo.15694581, openly available under a CC-BY-4.0 licence. Source data are provided with this paper.

Code availability

The dioscRi R package is available on Github and is available under the GPL-3 license. All code for analysis can be found at Github.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77142-7.