2.1. scRADAR establishes a high-performance benchmark across diverse single-cell drug-response phenotype cohorts

We systematically evaluated scRADAR on nine held-out scRNA-seq drug-response phenotype cohorts using a unified benchmark that included five representative bulk-to-single transfer-learning methods (SCAD [30], scATD [31], scDEAL [23], SSDA4Drug [29], and scAdaDrug [24]), a scGEN-adapted single-cell perturbation-oriented baseline, and four conventional supervised classifiers trained on the same target-domain single-cell splits (LR [44], RF [45], MLP [46], and XGBoost [47]). The nine baseline methods shown in Fig 2 include the five transfer-learning baselines and four target-domain supervised classifiers, whereas the scGEN-adapted comparator is reported separately in S5 Table. Across this unified evaluation framework, scRADAR showed the strongest overall performance profile, achieving cohort-averaged AUROC, AUPRC, Accuracy, Precision, Recall, and F1-score values of 0.967, 0.964, 0.953, 0.954, 0.958, and 0.956, respectively. These results suggest that the performance of scRADAR reflects not only access to labeled target-domain cells but also the contribution of its mechanism-guided representation learning and prototype-routing design.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. scRADAR achieves superior predictive performance and stability across nine held-out scRNA-seq drug-response phenotype cohorts.

(A–C) Bar plots comparing the mean AUROC (A), AUPRC (B), and F1-score (C) of scRADAR with nine baseline methods, including representative transfer-learning and target-domain supervised comparators, across the nine held-out test cohorts. Error bars denote 95% t-intervals derived from cross-validation-based models, indicating robustness to training data variations. (D) Cohort-averaged performance heat map summarizing AUROC, AUPRC, Accuracy, Precision, Recall, and F1-score across all methods; warmer colors indicate better performance.

https://doi.org/10.1371/journal.pcbi.1014392.g002

We first assessed global discrimination using AUROC (Fig 2A). Among the transfer-learning baselines, scDEAL provided the strongest average ranking performance with a mean AUROC of 0.830, whereas MLP was the most competitive supervised comparator with a mean AUROC of 0.920. Against these reference points, scRADAR improved the mean AUROC by 0.137 and 0.047, corresponding to relative gains of 16.5% and 5.1%, respectively. Notably, scRADAR maintained uniformly high cohort-level discrimination, with all nine AUROC values exceeding 0.90 and spanning a narrow high-performance range from 0.905 to 0.989. The magnitude of the improvement was especially evident in several representative cohorts. In GSE111014 (Ibrutinib), AUROC increased from 0.921 for MLP to 0.987 for scRADAR, an absolute gain of 0.066 (+7.2%). Likewise, in GSE131984 (Paclitaxel), scRADAR reached an AUROC of 0.978, outperforming MLP by 0.073 (+8.1%). GSE140440 (Docetaxel) represents a particularly challenging cohort because it is the smallest dataset in the benchmark after preprocessing (n_total = 284) and contains two prostate cancer cell-line backgrounds, DU145 and PC3. In this setting, the held-out evaluation is more sensitive to sampling variation, class-boundary instability, and cell-line-specific transcriptional differences. Even under this small-sample and mixed-background setting, scRADAR sustained a strong margin, yielding an AUROC of 0.983 versus 0.944 for MLP.

We next turned to AUPRC and F1-score, which provide a more operational view of classification quality by jointly reflecting ranking behavior and thresholded decision performance (Fig 2B and 2C). Here, the advantage of scRADAR became even more pronounced. Its mean AUPRC reached 0.964, compared with 0.811 for the strongest transfer-learning baseline and 0.909 for the strongest supervised baseline, corresponding to absolute gains of 0.153 and 0.055. Expressed in relative terms, these improvements amount to 18.9% over scDEAL and 6.1% over MLP. An analogous pattern was observed for F1-score: scRADAR achieved a mean F1-score of 0.956, exceeding SSDA4Drug (0.851) by 0.105 (+12.3%) and MLP (0.900) by 0.056 (+6.2%). Importantly, these gains were not confined to isolated datasets. scRADAR produced consistently elevated cohort-level AUPRC values, all above 0.90, ranging from 0.908 to 0.993, while F1-scores remained uniformly high across the benchmark, ranging from 0.905 to 0.986. In GSE131984 (Paclitaxel), for example, AUPRC increased from 0.893 for MLP to 0.981 for scRADAR, a gain of 0.088 (+9.9%), while F1-score rose from 0.885 to 0.971, a gain of 0.086 (+9.7%). A similarly clear separation was observed in GSE111014 (Ibrutinib), where scRADAR improved AUPRC from 0.908 to 0.982 and F1-score from 0.899 to 0.968 relative to MLP.

The cohort-averaged heat map in Fig 2D further highlights the coordinated nature of these gains. Rather than improving one metric at the expense of another, scRADAR occupied the leading position across all six summary measures. Relative to the best overall supervised baseline, MLP, scRADAR increased Accuracy from 0.900 to 0.953, Precision from 0.907 to 0.954, Recall from 0.894 to 0.958, and F1-score from 0.900 to 0.956. These correspond to absolute improvements of 0.053, 0.047, 0.064, and 0.056, respectively, or relative gains of 5.9%, 5.2%, 7.2%, and 6.2%. Particularly noteworthy is the increase in Recall, suggesting that scRADAR more effectively preserves Sensitive/Resistant label separation at the final decision threshold while maintaining strong precision. This balanced improvement profile is difficult to obtain with conventional classifiers and is consistent with the view that biologically grounded pathway encoding and sparse prototype routing yield a more robust decision geometry than generic supervised mapping.

An additional point worth emphasizing is that the supervised baselines were themselves highly competitive. MLP reached mean AUROC and AUPRC values of 0.920 and 0.909, while XGBoost achieved 0.908 and 0.898, underscoring the stringency of the benchmark. Against this stronger comparator tier, scRADAR nevertheless maintained a consistent performance margin across the summary metrics. Taken together, these results position scRADAR as a robust and competitive framework within the unified benchmark and support the view that its performance is linked to the combination of biologically grounded pathway encoding, mechanism-aware conditioning, and prototype-based decision structure.

We further included a scGEN-adapted baseline to compare scRADAR with a representative single-cell perturbation-oriented model. Under the same target-domain split and threshold-selection protocol, the scGEN-adapted baseline achieved cohort-averaged AUROC, AUPRC, and F1-score values of 0.884, 0.867, and 0.852, respectively (S5 Table). These results indicate that a single-cell perturbation-style latent representation is informative for this task, but its performance remains below both the strongest target-domain supervised comparator and full scRADAR. Thus, the performance advantage of scRADAR cannot be attributed simply to using single-cell data, but is more consistent with the combined contribution of pathway-level encoding, mechanism-aware drug conditioning, and sparse prototype routing.

2.2. Systematic ablation disentangles component-wise contributions and validates architectural choices

To rigorously attribute performance gains to specific design choices, we conducted a comprehensive component-wise ablation study, with detailed performance metrics summarized in Table 1. Results are reported as mean values with 95% t-intervals to quantify both the magnitude of predictive power and the stability of model training.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Ablation study of scRADAR components and input feature configurations. Data are reported as mean ±95% t-intervals derived from five fold-trained models (K = 5). The best performance in each metric is highlighted in bold. Combined indicates the full feature set (ssGSEA + PROGENy for pathways; Static + Dynamic for drugs). Abbreviations: FiLM, feature-wise linear modulation; Proto., prototype-based routing; Top-k, sparse routing using the k most similar prototypes. The “MLP head” variant replaces the prototype routing mechanism with a standard classifier.

https://doi.org/10.1371/journal.pcbi.1014392.t001

We first examined the synergy of multi-view biological representations by isolating the metabolic (ssGSEA) and signaling (PROGENy) inputs. The integration of both views proved essential for robust discrimination, as evidenced by the performance deficits observed in single-view models. Specifically, relying solely on ssGSEA or PROGENy led to absolute AUROC decreases of 2.4% and 3.0%, respectively, relative to the full model’s score of 0.967. A similar trend was observed in AUPRC, where the full model (0.964) outperformed the single-view variants by approximately 3.0% to 3.8%. Notably, as detailed in the stability analysis, the uncertainty intervals for single-view models widened by over 3-fold (e.g., AUROC error margin increasing from ±0.5% to ±1.7%), indicating that dual-view integration not only enhances predictive accuracy but also stabilizes model convergence against training data variations. This confirms that metabolic state and signaling potential provide non-redundant information: one captures the energetic baseline, while the other reflects dynamic responsiveness. This interpretation was further reinforced by a direct cross-view correlation analysis. Specifically, we quantified pairwise Pearson correlations between ssGSEA-derived metabolic features and PROGENy-derived signaling features within each cohort and observed only modest overall coupling across datasets (S1 Fig). Across the nine cohorts, mean absolute cross-view correlation ranged from 0.052 to 0.178, whereas median absolute correlation ranged from 0.036 to 0.139, arguing against the interpretation that the two pathway views are simple re-encodings of the same signal. In a representative cohort, GSE152469, the compact cross-view matrix revealed a sparse pattern dominated by weak correlations, with only a limited subset of pathway pairs showing stronger local correspondence. Together, these results indicate that the two views are partially connected yet globally complementary.

We also assessed whether the main performance profile was sensitive to the feature-normalization strategy. Across the four pathway-feature normalization settings, full scRADAR remained within a narrow performance range (S2 Table). The default z-score setting achieved AUROC, AUPRC, and F1-score values of 0.967, 0.964, and 0.956, respectively, while the alternative normalization settings produced comparable values. These results indicate that the predictive conclusions were not driven by a single normalization choice.

To address the dependence of pathway-based modeling on the selected gene-set collection, we further performed pathway-view and pathway-set sensitivity analyses. The ssGSEA-only and PROGENy-only variants remained informative but were lower than the full dual-view representation, supporting the use of both pathway activity and signaling perturbation views (S3 Table). Replacing Reactome with Hallmark produced comparable performance within the same range, and the expanded Reactome + Hallmark + PROGENy representation yielded comparable ranking performance. These results indicate that the main conclusions were not driven by a single pathway collection, although disease- or lineage-specific gene sets may still improve interpretive resolution in specific cohorts.

Because B-cell malignancy pathways may be particularly relevant to the Ibrutinib cohort GSE111014, we additionally evaluated B-cell-specific signatures, including CD40 signaling, NF-κB signaling, B-cell receptor signaling, BTK downstream signaling, B-cell activation, apoptosis, antigen presentation, and proliferation-related programs. Adding these signatures produced comparable performance with small numerical changes in AUPRC and F1-score (S4 Table). The expanded B-cell setting retained the same B-cell signatures and further incorporated Hallmark features, again yielding performance within the same range. These results suggest that B-cell-specific signatures provide complementary context information, while the default Reactome + PROGENy representation already captures the major response-associated signal in this cohort.

We next examined the contribution of the hybrid drug fingerprint, which combines pharmacological priors with perturbation-derived context. Using static fingerprints alone already yielded a strong baseline, with AUROC and F1-score values of 0.958 and 0.944, respectively. Incorporating dynamic LINCS L1000 signatures further improved performance across all three summary metrics, raising AUPRC to 0.964 and F1-score to 0.956. Notably, the dynamic-only variant remained weaker than the static-only variant, trailing by 1.0% in AUROC and 1.2% in F1-score. This pattern suggests that perturbation-derived transcriptional information contributes useful context, but is most effective when anchored by canonical target and pathway annotations. Taken together, these comparisons support the value of integrating mechanism-based priors with data-driven perturbational signals rather than relying on either source alone.

Architectural ablations further clarified which components were most responsible for the performance of the full model. Among these variants, replacing prototype routing with a standard MLP head produced the largest degradation, with AUROC falling from 0.967 to 0.928, AUPRC from 0.964 to 0.915, and F1-score from 0.956 to 0.907. The magnitude of this drop indicates that the prototype-based decision structure is not merely an interpretability add-on, but a central component for representing heterogeneous response modes within and across cohorts. By contrast, removing FiLM modulation yielded a smaller yet consistent decline across metrics, including a 0.6% reduction in F1-score, indicating that FiLM-based drug-dependent pathway reweighting contributes complementary value beyond the pathway representation itself. Viewed together, these results support a division of labor within the architecture: FiLM provides drug-dependent pathway reweighting, whereas prototype routing organizes this conditioned information into interpretable response archetypes.

Finally, the comparison between sparse and dense routing highlights the value of explicit decision sparsity. Relative to dense routing, top-k sparse routing produced a modest but consistent gain in predictive performance while yielding a clearer improvement in robustness. In particular, the AUPRC uncertainty interval was reduced by 50%, from ±0.006 to ±0.003, indicating that sparse routing dampens prototype-level noise and stabilizes optimization under training variability. Importantly, this gain in stability was achieved without sacrificing predictive accuracy, suggesting that sparsity improves not only interpretability but also the reliability of the learned decision geometry.

Because the component-wise ablations above were performed under the primary within-cohort held-out-cell protocol, we further assessed whether the mechanism-aware design retained value when the test drug itself was unseen during model development. We performed an additional controlled within-study unseen-drug transfer analysis using the three drug-specific subsets of GSE131984. In each split, two drug-specific subsets were used for training and the remaining drug was held out for testing, with no cells from the held-out drug used for training, validation, threshold selection, or hyperparameter selection. This setting controls for study, platform, and cell-line background while testing transfer across transcriptionally distinct drug responses.

Compared with the cell-only MLP, drug-aware models improved macro-average performance, indicating that mechanism-aware drug information provided transferable signal across held-out drug responses (S6 Table). Full scRADAR achieved the highest macro-average AUROC, AUPRC, and F1-score, although drug-specific profiles varied across held-out drugs. These results support the use of FiLM-conditioned pathway reweighting and sparse prototype routing as a compact mechanism-aware interaction design, without implying that the model captures all higher-order context-specific drug-cell interactions.

2.3. Hybrid drug fingerprints show consistency with canonical mechanisms and highlight candidate context-specific response-associated features

To examine whether the learned drug representations were consistent with pharmacological knowledge, we analyzed post hoc drug-fingerprint score profiles using the procedure described in Section 4.11.1, with the results shown in Fig 3. By comparing static prior-derived mechanism scores with dynamic perturbation-derived pathway scores, we examined whether the diagnostic score profiles were consistent with known mechanisms of action.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Hybrid drug fingerprints show canonical and context-specific score patterns.

(A–D) Drug-fingerprint score profiles for four representative drugs. The x-axis shows the z-scored hybrid drug-fingerprint score. Blue points denote static prior-derived mechanism scores, whereas red points denote dynamic perturbation-derived pathway scores. Static and dynamic scores were standardized separately within each drug-specific profile. These scores are intended as post hoc diagnostic summaries rather than direct measurements of pathway activation.

https://doi.org/10.1371/journal.pcbi.1014392.g003

For highly specific targeted therapies, the model showed score patterns consistent with known mechanisms, illustrated in Fig 3A–C. In the case of Palbociclib (a CDK4/6 inhibitor), Fig 3A shows that the model assigned top-ranking weights to Cdk4, Cdk6, and Cell Cycle programs in both static and dynamic views, supporting the interpretation that cell-cycle-related features were top-ranked in the Palbociclib diagnostic score profile. Similarly, the analyses for microtubule-stabilizing agents, such as Docetaxel (Fig 3B) and Paclitaxel (Fig 3C), consistently prioritized Microtubule Stabilizer and Mitosis signatures. This concordance between prior knowledge and dynamic perturbation-derived score profiles supports the interpretation that scRADAR uses pharmacologically relevant features in these representative cases.

Beyond recapitulating known pharmacological associations, the dynamic component highlighted candidate non-canonical associations that were not represented by static priors alone, as exemplified by the Cisplatin cohort in Fig 3D. While the static prior correctly highlighted DNA Replication and DNA Crosslinker terms, the data-driven dynamic signature unexpectedly prioritized Hormone Ligand Binding as a top-ranked feature, shifting it from a negative static score to a highly positive dynamic weight. We interpret this pattern as a candidate context-specific resistance-associated signal, whereby hormone-related signaling may mark a cohort-specific transcriptional association with Cisplatin response labels in this retrospective dataset. Although this interpretation remains hypothesis-generating, it is consistent with independent studies reporting associations between steroid hormone signaling and platinum-response phenotypes in reproductive cancers [48–51]. We therefore interpret these dynamic signatures as model-derived diagnostic signals that may prioritize candidate context-specific hypotheses, rather than as direct evidence of pathway activation, inhibition, or causal resistance mechanisms.

2.4. Response landscapes reveal intratumoral heterogeneity and motivate calibration beyond a single operating point

A critical challenge in precision oncology is resolving clinically relevant intratumoral heterogeneity (ITH), where resistance-associated subpopulations can influence cohort-level response patterns. To assess scRADAR’s resolution, we projected single-cell predictions onto low-dimensional manifolds and examined how fractions of cells assigned to the Sensitive label vary across transcriptional neighborhoods, as illustrated in Fig 4.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Intratumoral heterogeneity motivates calibration beyond a single operating point.

(A–C) Left: UMAP embeddings of tumor cells colored by transcriptional clusters. Right: Bar plots of the fraction of cells assigned to the Sensitive label within each cluster, revealing pronounced within-cohort variability. (D–E) Threshold calibration for Palbociclib and Erlotinib. The vertical grey line denotes the global threshold selected on the validation set. Red points show post hoc cluster-wise F1-optimizing thresholds computed using cluster-specific response-associated labels after model evaluation; these thresholds were used only for descriptive calibration analysis. The difference between global and cluster-wise thresholds summarizes local score-label calibration heterogeneity rather than a prospective threshold-adaptation procedure.

https://doi.org/10.1371/journal.pcbi.1014392.g004

Transcriptional clustering delineates discrete phenotypic neighborhoods, within which scRADAR reveals sharply divergent response profiles (Fig 4A–C). In the Cisplatin cohort, visualized in Fig 4A, specific subpopulations, including Clusters 3, 4, 8, and 9, exhibited near-uniform enrichment of the Sensitive label, with observed Sensitive-label fractions approaching 1.0. In contrast, distinct groups such as Clusters 5 and 6 show near-zero Sensitive-label fraction, indicating near-uniform enrichment of the Resistant label, with additional neighborhoods occupying intermediate states. A similar structure is observed in Erlotinib-treated cohorts (Fig 4B and 4C), where a neighborhood enriched for the Sensitive label, specifically Cluster 0 with a Sensitive-label fraction of approximately 0.85, coexists with a Resistant-enriched neighborhood, Cluster 1, with Sensitive-label fractions near zero in both datasets. These patterns are consistent with non-random organization of response-associated labels across transcriptional neighborhoods, although the UMAP visualization itself should be interpreted as descriptive rather than as direct evidence of causal biological structure. To statistically support this descriptive observation, we performed cluster-level response-label enrichment tests using label permutation while preserving cluster sizes and the cohort-level label composition (S7 Table). In the Cisplatin cohort, multiple clusters showed significant enrichment for either Sensitive- or Resistant-labeled cells after FDR correction, including Sensitive-enriched Clusters 2, 3, 4, 7, 8, and 9 and Resistant-enriched Clusters 0, 1, 5, and 6. In the two Erlotinib cohorts, the largest transcriptional neighborhoods showed reproducible label-enrichment patterns, with Cluster 0 enriched for Sensitive-labeled cells and Cluster 1 enriched for Resistant-labeled cells in both datasets. Smaller clusters were interpreted cautiously because their enrichment estimates were less stable.

This pronounced heterogeneity challenges the assumption that a single operating point is sufficient across all subpopulations, as analyzed in Fig 4D and 4E. Using validation-guided calibration, we selected global thresholds (vertical grey lines), specifically 0.24 for Palbociclib and 0.42 for Erlotinib, to optimize cohort-level performance. While these global operating points promote generalizability, post hoc cluster-wise analyses, computed as described in Section 4.11.2 and shown as red points, revealed substantial deviations in local score-label calibration patterns. Fig 4D demonstrates that in the Palbociclib cohort, Cluster 0 favors a markedly higher threshold compared to the global operating point, whereas the majority of other clusters favor lower thresholds. We further characterized Palbociclib Cluster 0 to clarify why this neighborhood showed a markedly higher post hoc threshold. Cluster 0 was Resistant-labeled-cell-enriched in the held-out analysis, containing 16 Sensitive-labeled and 86 Resistant-labeled cells (S8 Table and S2 Fig). Despite this imbalance, the model showed strong local score-label separation, with a local AUROC of 0.948 and an AUPRC of 0.703, which was substantially above the local Sensitive-label prevalence baseline of 0.157. Sensitive-labeled cells in this cluster received high predicted Sensitive probabilities, whereas Resistant-labeled cells received low predicted Sensitive probabilities, resulting in a median probability difference of 0.930. The top model-derived pathway deviations were dominated by cell-cycle and mitotic programs, consistent with a cell-cycle-related transcriptional state in this Palbociclib context. Therefore, the high cluster-wise threshold should be interpreted as a post hoc summary of local score-label calibration within a Resistant-labeled-cell-enriched transcriptional neighborhood, rather than as an independently validated biological cutoff. In the Erlotinib case (Fig 4E), cluster-wise optima span both above and below the global cutoff: some clusters favor substantially higher decision thresholds, whereas some lower-threshold neighborhoods lie below the global threshold. This illustrates that a single global decision boundary may summarize cohort-level performance while masking local score-label calibration differences across transcriptional neighborhoods.

The divergence between global and cluster-specific optima highlights a limitation of one-size-fits-all predictors for precision medicine. A fixed cohort-level threshold may achieve strong average metrics yet underperform in identifying clinically relevant resistance-associated neighborhoods concentrated in specific regions [52]. By explicitly mapping subpopulation-aware decision behavior, scRADAR provides a descriptive framework for interpreting response predictions in the presence of ITH and for prioritizing transcriptional neighborhoods for follow-up experimental or clinical annotation. These cluster-wise threshold differences are intended as a post hoc characterization of local decision geometry rather than a threshold adaptation procedure for unlabeled future samples.

2.5. Pathway-level attribution highlights candidate signaling and metabolic features associated with Resistant-labeled states

Beyond predictive accuracy, scRADAR provides pathway-level attributions that translate cell-wise response scores into interpretable molecular programs. By aggregating attributions across cells, we derived two complementary insights: cluster-resolved pathway activity landscapes that expose neighborhood-specific programs, as shown in Fig 5A and 5C, and global contrasts between Resistant- and Sensitive-labeled cells that prioritize putative resistance-associated pathways, summarized in Fig 5B and 5D.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Pathway landscape analysis identifies candidate resistance-associated pathway programs.

(A, C) Bubble plots showing cluster-wise mean pathway activity (z-scored) for Erlotinib (A) and JQ1 (C), where color indicates pathway activity (Z-score) and circle size denotes effect size (|β|). (B, D) Stacked bars summarizing top-ranked pathway effect scores contrasting cells assigned to Resistant (red) versus Sensitive (blue) labels. Note the enrichment of TGF-β-related attribution among Erlotinib-associated Resistant-labeled cells (B) and pyruvate-metabolism-related attribution among JQ1 Resistant-labeled cells (D).

https://doi.org/10.1371/journal.pcbi.1014392.g005

In the EGFR-TKI cohort (Erlotinib), the contrast between Resistant- and Sensitive-labeled cells showed elevated TGF-β-related attribution patterns and reduced attribution to canonical EGFR-related signaling features in the Resistant-labeled group (Fig 5B). Importantly, this attribution pattern was accompanied by higher TGF-β-related signals, most notably TGF-β signaling, consistent with previously reported mesenchymal-like and stress-adaptive programs associated with TKI failure, although the present analysis does not establish a causal transition [53,54]. Furthermore, the cluster-level landscape visualized in Fig 5A reveals substantial spatial heterogeneity across neighborhoods. Specific patterns emerge in proliferation-related programs (e.g., M-phase/mitotic checkpoint) and extracellular-matrix–associated processes, suggesting that proliferation-related and microenvironment-associated features vary across the manifold rather than forming a single uniform Resistant-labeled pattern.

Similarly, in the BET-inhibitor cohort (JQ1), cytoskeletal and metabolic signatures emerged as prominent features associated with Resistant-labeled cells. As illustrated in Fig 5D, the top response-associated pathway features characterizing Resistant-labeled cells include Rho-family GTPase cycles (RHOD/RHOC) and pyruvate metabolism, suggesting cytoskeletal plasticity and energetic metabolism as candidate axes for follow-up perturbation-based validation [55]. The cluster-resolved landscape further emphasizes that these programs are not ubiquitously active but rather heterogeneous across neighborhoods. Fig 5C shows that while some clusters exhibit broad attenuation of multiple pathways, others display selective activation of specific modules—such as SREBF-related regulation and vascular interaction signatures—indicating multiple response-associated transcriptional patterns rather than a monolithic Resistant-labeled phenotype.

Collectively, these pathway maps should be interpreted as hypothesis-generating model-derived associations rather than direct experimental evidence of causal mechanisms. They provide a basis for prioritizing candidate response-associated programs for future perturbation-based validation [56]. Accordingly, terms such as “TGF-β-associated” or “metabolic response-associated” are used here to describe retrospective model-derived associations with response-associated labels, not experimentally verified mechanisms of resistance.

2.6. Sparse prototype routing exposes archetypal decision modes and preserves interpretability in complex microenvironments

To move beyond the “black-box” perception of deep learning, we dissected the internal decision structure of scRADAR through prototype analysis, as visualized in Fig 6. By examining the learnable connections between latent prototypes and biological pathways, we show that scRADAR organizes high-dimensional transcriptomic variability into a compact set of latent response archetypes that can be characterized by pathway-attribution patterns. The prototype–pathway weight maps provide a concise lookup table linking each latent prototype to characteristic pathway signatures.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Sparse prototype routing yields interpretable latent response archetypes.

(A,B) Prototype–pathway weight heatmaps (black dots: P < 0.05). (C,D) UMAP maps of per-cell activation for representative prototypes that were post hoc enriched for Sensitive- or Resistant-labeled cells. (E–G) Predicted probability of the Sensitive response phenotype projected onto UMAP embeddings across a low-purity microenvironmental sample (E) and malignant-dominant cohorts (F,G), revealing response-probability gradients across transcriptional neighborhoods.

https://doi.org/10.1371/journal.pcbi.1014392.g006

In the Erlotinib cohort, Fig 6B highlights that the learned prototypes align with distinct regulatory axes. Specifically, Prototype P1 (~10% of cells) showed elevated attribution to TGFBR3-mediated TGF-β-related features () [54,57]. Beyond this axis, scRADAR identifies separable biological programs: Prototype P8 (~9%) is significantly associated with ABC transporter lipid homeostasis () [58], suggesting a distinct metabolic/efflux-related signature, while Prototype P5 (~9%) shows significant attribution to SCF–SKP2 degradation of p27/p21 (), consistent with cell-cycle checkpoint–linked regulation [59]. Turning to the Ibrutinib cohort (Fig 6A), the most prevalent prototype P4 (~14%) shows positive attribution to the pancreatic acinar lineage program () and carries prominent weights on the RAC3 GTPase cycle and MET–PTK2 signaling. Notably, another frequent prototype, P11 (~11%), is associated with aberrant PI3K signaling (), indicating a separable PI3K-centered module [60]. Overall, these attribution patterns support the interpretation that the learned prototypes are associated with coherent pathway signatures rather than functioning solely as arbitrary latent routing anchors.

Having established these pathway-defined archetypes, we next projected prototype activations onto the UMAP manifold to examine whether individual prototypes were locally enriched in response-associated transcriptional neighborhoods (Fig 6C and 6D). In the Palbociclib dataset, P0, which was post hoc enriched among Sensitive-labeled cells, showed higher activation in regions with elevated predicted probabilities of the Sensitive response phenotype, whereas P7, which was enriched among Resistant-labeled cells, was activated in regions dominated by Resistant-labeled cells (Fig 6C). In the JQ1 dataset, a similar complementary pattern was observed between P7 and P12: P7 was preferentially activated in regions enriched for Sensitive-labeled cells, whereas P12 was preferentially activated in regions enriched for Resistant-labeled cells (Fig 6D). These observations indicate that the prototype labels are post hoc interpretations based on routing enrichment and associated response labels, rather than predefined biological classes. The resulting topology supports the view that sparse routing organizes cells into multiple latent response archetypes instead of relying on a single global decision rule.

Finally, we further examined whether this structured interpretability remained evident under the confounding complexity of tumor microenvironments. Visualized in Fig 6E–G, scRADAR maintained informative response-probability patterns in low-purity, microenvironment-rich settings. In GSE111014 (Fig 6E), the model preserved heterogeneous gradients in the predicted probability of the Sensitive response phenotype within the malignant compartment while exhibiting coherent, cell-type-resolved patterns across immune and stromal populations. Similar malignant-resolved response-probability gradients were evident in the Erlotinib and Docetaxel cohorts (Fig 6F and 6G). Collectively, these observations suggest that prototype-based routing provides a structured post hoc summary of response-associated pathway patterns in a manner that remains informative under complex microenvironmental admixture.

To further test whether prototype activations corresponded to reproducible response-associated structures rather than arbitrary routing anchors, we quantified response-label and transcriptional-cluster enrichment among the top-routed cells for representative prototypes (S9 Table). Across multiple drug cohorts, selected prototypes showed significant post hoc enrichment for Sensitive- or Resistant-labeled cells and were also enriched within specific transcriptional clusters after FDR correction. For example, Palbociclib P4 was enriched among Resistant-labeled cells and Cluster 0, whereas Palbociclib P6 was enriched among Sensitive-labeled cells and Cluster 2. Similar label- and cluster-enrichment patterns were observed in JQ1, Erlotinib, Ibrutinib, and Cisplatin cohorts. These results support the interpretation that prototypes function as learned latent response archetypes anchored in both pathway-attribution patterns and local transcriptional neighborhoods, while their biological labels remain post hoc summaries rather than predefined classes.

4.1. Overview of the scRADAR framework

The scRADAR framework integrates biologically grounded modules into a unified end-to-end pipeline to predict drug-response phenotypes from single-cell transcriptomes (Fig 1). The workflow begins with data acquisition and preprocessing as outlined in Fig 1A, where raw scRNA-seq matrices undergo rigorous quality control and normalization. To mitigate the technical variability inherent in gene-level features, Fig 1B illustrates the construction of a dual-view cellular representation that aggregates metabolic pathway activities—estimated via ssGSEA using GSVA [62] —together with signaling perturbation responses inferred by PROGENy [61]. Parallel to cellular encoding, we synthesize a mechanism-aware drug fingerprint (Fig 1C) by integrating static target annotations with dynamic perturbational signatures. These representations are then coupled via Feature-wise Linear Modulation (FiLM), depicted in Fig 1D, which reweights pathway embeddings in a drug-dependent manner according to the mechanism-aware fingerprint. Finally, as demonstrated in Fig 1E, predictions are generated through interpretable prototype routing, where a query vector maps each cell–drug pair to a sparse set of latent prototypes, decomposing predictions into transparent, biologically meaningful archetypes.

4.2. Problem formulation

We formulate single-cell drug-response phenotype prediction as a binary classification problem. Let denote a dataset of single cells. For each cell i, represents the gene expression profile across G genes, and denotes the mechanism-aware feature vector of the applied drug. The label indicates the cohort-harmonized binary response phenotype with denoting the Sensitive label and denoting the Resistant label. Our objective is to learn a parameterized function that estimates the predicted probability of the Sensitive response phenotype: . Because the included GEO cohorts use heterogeneous annotation schemes, Sensitive and Resistant denote cohort-harmonized response-associated labels; in longitudinal, treatment-stage, or subset-constructed cohorts, they should be interpreted as operational binary labels for benchmarking rather than direct single-cell functional viability measurements.

4.3. Data acquisition and preprocessing

4.3.1. Data acquisition.

Publicly available scRNA-seq drug-response cohorts were collected from the NCBI Gene Expression Omnibus (GEO) repository [63–65]. We curated nine independent cohorts for single-cell drug-response phenotype prediction spanning five cancer types and seven drugs (Table 2), corresponding to GSE111014 [66], GSE117872 [67], GSE149214 [68], GSE149383 [68], GSE152469 [69], GSE140440 [70], and three drug-specific subsets derived from GSE131984 [71]. Cohorts were included based on the following criteria: (1) human cancer-derived single-cell transcriptomes; (2) exposure to a defined small-molecule perturbation; (3) availability of gene-by-cell count matrices with accompanying cell-level annotations; and (4) clearly defined or reproducibly harmonizable binary response-associated labels. Because these GEO cohorts originated from independent studies with heterogeneous annotation schemes, binary response labels were harmonized on a dataset-specific basis using the original study metadata, sample annotations, preprocessing-defined comparison states, and cohort-specific response definitions reported in the corresponding source publications. Study-defined binary groups were used directly when available; otherwise, Sensitive and Resistant denote harmonized operational labels used for binary benchmarking, consistent with prior single-cell drug-response modeling studies that adopted comparable binary response settings for retrospective scRNA-seq cohorts [23,29]. Cohort-level sources of label assignment, supporting references, and detailed harmonization rules are summarized in S1 Table, and the corresponding analysis-ready cohort statistics are reported in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Overview of single-cell drug-response phenotype datasets used in this study. Summary of the scRNA-seq drug-response phenotype cohorts used in this study, including GEO accession, cancer type, drug, and sample source. n_total denotes the number of quality-controlled cells retained for analysis, whereas n_sensitive and n_resistant denote the numbers of cells assigned to the Sensitive and Resistant response-associated labels, respectively, after dataset-specific label harmonization. Cohort-level sources of label assignment and detailed harmonization rules are provided in S1 Table. “Patient-derived” denotes primary clinical samples, whereas “Cell line” denotes in vitro perturbation experiments using the indicated cell line(s). For GSE131984, three drugs were profiled in the same SUM159 model, and counts are reported separately for each drug-specific comparison.

https://doi.org/10.1371/journal.pcbi.1014392.t002

4.3.2. Data preprocessing.

We applied a standardized quality control (QC) and normalization pipeline to remove low-quality cells and uninformative genes. For each dataset, mitochondrial genes were identified based on gene symbols, and per-cell QC metrics were computed, including the number of detected genes (), total UMI counts (), and the fraction of mitochondrial transcripts () [72] within the scverse ecosystem [73]. Cells were retained only if they satisfied , and . Genes expressed in fewer than 3 cells were removed to reduce noise from extremely sparse features.

Let denote the post-QC raw count matrix, where the entry represents the raw count of gene j in cell i. We then performed library-size normalization followed by log-transformation to stabilize variance. Specifically, counts in each cell i were normalized to a fixed library size of and transformed as

(1)

yielding the normalized expression vector used for downstream feature construction and model training [74].

Downstream machine-learning utilities for evaluation and model selection were implemented using scikit-learn [75].

4.4. Dual-view pathway representation

Given the high sparsity (often exceeding 90% zeros) and high dimensionality of scRNA-seq data, directly using gene-level expression as model input can be unstable and sensitive to technical noise [76]. To obtain a more robust and biologically meaningful description of the cellular state, we construct a dual-view pathway representation that jointly captures metabolic activities and signaling pathway responses for each cell.

4.4.1. Metabolic view (ssGSEA).

For the metabolic view, we quantify pathway activity using single-sample Gene Set Enrichment Analysis (ssGSEA) based on curated gene sets from the Reactome database [77]. Reactome was selected as the default ssGSEA gene-set collection because it provides manually curated and hierarchically organized pathways with broad coverage across cellular processes, which facilitates consistent feature construction and cross-cohort comparison. Nevertheless, a fixed pathway collection cannot capture every disease- or lineage-specific program. Therefore, Reactome-derived ssGSEA scores should be interpreted as a standardized pathway basis rather than an exhaustive representation of all potentially relevant biology. To evaluate whether the main conclusions depended on this pathway choice, we performed pathway-set sensitivity analyses using an alternative Hallmark-based representation, an expanded Reactome + Hallmark representation, and B-cell-specific signatures for GSE111014. Let be the gene set corresponding to the m-th metabolic pathway. To quantify the activity of this pathway in a specific cell i, ssGSEA transforms the cell’s gene expression vector into a rank-ordered list. It then calculates an enrichment score that reflects the relative overrepresentation of genes in among highly expressed genes in cell i [36,78]. Evaluating ssGSEA across all selected metabolic pathways yields a metabolic-view embedding:

(2)

where denotes the number of metabolic pathways, and denotes the transpose operator, so that is a column vector.

4.4.2. Signaling view (PROGENy).

To capture the cell’s dynamic potential to respond to drug perturbations, we construct a complementary signaling view using the PROGENy model for pathway activity inference [61].

PROGENy leverages large-scale perturbation transcriptomics to derive a pathway footprint weight matrix .

Given the gene expression vector , we denote the inferred activity of signaling pathway p in cell i by computed as a weighted linear combination:

(3)

where denotes the expression of gene g in cell i, and is the specific weight linking gene g to signaling pathway p derived from the PROGENy model. Here, the collection of all weights forms the matrix , where is the number of pathways. In practice, we use the official human PROGENy weights and retain the top 500 genes with the largest absolute weights for each pathway.

Stacking activities across all signaling pathways yields the signaling-view embedding:

(4)

4.4.3. Fusion into a unified pathway embedding.

Finally, we fuse the metabolic and signaling views into a unified pathway-level representation for each cell. For each train–validation–test split, pathway-score standardization was fit using only the training cells. The resulting training-set means and standard deviations were then applied unchanged to the corresponding validation and held-out test cells. This fold-specific standardization procedure was used for scRADAR and all baselines requiring scaled features, thereby preventing validation or test-set information from entering feature normalization. We denote this operation by , applied independently to and . The resulting cell-level pathway embedding is obtained by concatenation:

(5)

where denotes the total dimensionality of the dual-view pathway feature vector. This pathway-level embedding serves as the input representation for subsequent mechanism-guided modulation and prototype-based prediction in scRADAR. To assess whether the predictive conclusions were sensitive to pathway-feature normalization, we further performed a normalization sensitivity analysis after pathway-score construction. The default setting used training-cell-fitted z-score standardization. Three alternative scaling strategies were evaluated under the same model architecture and data-splitting protocol: robust median/IQR scaling, quantile-normal transformation, and min–max scaling to [0,1]. This analysis was used only to assess preprocessing robustness and was not used for model selection.

4.5. Mechanism-aware drug fingerprint

We explicitly encode pharmacological prior knowledge through a mechanism-aware drug fingerprint vector . This vector integrates two complementary sources of information: (1) a static fingerprint , constructed from curated drug–target and pathway annotations provided in the GDSC metadata; and (2) a dynamic fingerprint , derived from LINCS L1000 small-molecule perturbation signatures and projected into the signaling pathway space via PROGENy. Here, and denote the dimensionalities of the static and dynamic fingerprints, respectively.

To improve reproducibility, drug identifiers were harmonized before constructing these fingerprints. Drug names from the single-cell cohorts were standardized by converting them to lowercase, unifying spacing and punctuation variants, and removing non-informative salt-form descriptors or vendor-specific suffixes only when the underlying active compound identity was unchanged. The standardized names were then matched to curated GDSC drug annotations to obtain . For the dynamic component, the same standardized names were matched to LINCS L1000 perturbagen names and aliases. When a direct name match was not available, aliases were manually checked, and only unambiguous matches referring to the same active compound were retained. When multiple eligible LINCS signatures were available for the same compound, small-molecule perturbation signatures with mappable gene identifiers were summarized by an unweighted mean at the compound level before projection into the PROGENy signaling pathway space to obtain . Before fusion, and were independently normalized so that their scales were comparable. When one component was unavailable after matching, the available component was retained, and the missing component was represented by a zero vector in the normalized feature space so that all drugs shared the same final fingerprint dimensionality. To construct the final hybrid representation, we employ a weighted concatenation strategy:

(6)

where and are scalar hyperparameters controlling the relative scaling of static and dynamic features, respectively, and Concat(⋅) denotes vector concatenation. Unless otherwise specified, we set in our experiments, thereby preserving the full dimensionality of both information sources while allowing for differential re-weighting if needed.

4.6. Feature-wise linear modulation (FiLM)

To model the conditional influence of drug mechanisms on cellular state, scRADAR incorporates a Feature-wise Linear Modulation (FiLM) layer [35]. Given the hybrid drug fingerprint for the drug applied to cell i, a multilayer perceptron generates channel-wise affine parameters:

(7)

where and correspond to the scaling and bias coefficients specific to sample i, respectively. In other words, the MLP maps the drug-fingerprint space to the fixed dimensionality of the cellular pathway embedding.

These parameters dynamically recalibrate the cellular embedding via an affine transformation:

(8)

where denotes element-wise multiplication. Intuitively, FiLM acts as a drug-dependent pathway reweighting step. The cellular pathway embedding first describes which metabolic and signaling programs are active in a given cell. The drug fingerprint then generates scaling and shifting coefficients that increase or decrease the contribution of these pathway features according to the drug’s curated targets and perturbation-derived signatures. Therefore, the same cellular pathway state can be interpreted differently under different drug mechanisms before being routed to the prototype layer. In this study, FiLM is used as a mechanism-level conditioning layer rather than as a fully cell-adaptive interaction module. This design captures drug-dependent reweighting of cell-state features, but it does not model all higher-order context-specific drug–cell interactions. The conditioned representation is then passed to the prototype routing module for phenotype prediction.

4.7. Interpretable prototype routing

To decompose black-box predictions into a set of interpretable patterns, scRADAR employs a sparse prototype routing layer. We introduce a learnable prototype bank:

(9)

where consists of latent prototype vectors, and denotes the j-th prototype representing a distinct molecular pattern. is the dimensionality of the FiLM-conditioned cell–drug embedding.

4.7.1. Top-k sparse routing.

For each cell–drug pair, the FiLM-conditioned pathway representation is first mapped to a query vector:

(10)

where denotes the query vector for sample i (representing the conditioned state of the cell), and and are the learnable projection matrix and bias, respectively.

We then compute the similarity between the query and each prototype . To ensure that routing is based on pattern alignment rather than magnitude, we employ cosine similarity:

(11)

where denotes the cosine similarity between cell i’ s query and prototype j and denotes the L₂-norm. To encourage sparse and decisive routing, we retain only the top- most similar prototypes for each cell and denote their index set by , where is a sparsity hyperparameter and L is the total number of prototypes. Routing weights are then obtained by applying a temperature-scaled softmax restricted to :

(12)

Here, is a temperature hyperparameter controlling the sharpness of the routing distribution, and indexes prototypes within the selected set. By construction, the routing weights satisfy [79].

4.7.2. Prototype-based prediction head.

Each prototype is associated with a learnable scalar response logit . The final predicted response probability for cell i is then computed as a weighted sum of prototype-specific logits followed by a sigmoid:

(13)

where denotes the sigmoid function. This design ensures that each prediction can be traced back to a small subset of highly activated prototypes, providing an interpretable decomposition of the model’s decision in terms of latent response archetypes.

4.8. Loss function and optimization

We train scRADAR end-to-end by minimizing a composite objective that balances predictive accuracy and prototype diversity:

(14)

where denotes the prediction loss, represents the prototype diversity regularization, and is a scalar hyperparameter controlling the trade-off between the two terms.

Model parameters are optimized using the Adam optimizer [80]. Key optimization settings, such as the learning rate and batch size, are treated as tunable hyperparameters.

The prototypes are not predefined Sensitive or Resistant classes. They are randomly initialized learnable latent anchors optimized jointly with the prediction loss and the prototype-diversity regularization. Their biological interpretation is assigned only after training. Specifically, a prototype can be described as Sensitive-enriched or Resistant-enriched only when cells with high routing weights to that prototype show post hoc enrichment of the corresponding cohort-harmonized response-associated label. Thus, prototype labels should be interpreted as summaries of learned latent response archetypes rather than as predefined biological categories.

4.8.1. Sample-weighted binary cross-entropy.

Let denote the logit output of the prototype-based prediction head for cell i, and let be the corresponding ground-truth response-associated label. The per-sample binary cross-entropy loss is:

(15)

where is the sigmoid function. To handle label imbalance and incorporate sample-level importance, we apply a non-negative weight to each sample (provided as batch_weights in the implementation) and minimize the weighted average loss [81]:

(16)

In practice, is implemented using a numerically stable binary cross-entropy function with reduction disabling, allowing for element-wise weighting by sample weights .

4.8.2. Prototype diversity regularization.

To prevent mode collapse and encourage a diverse prototype bank, we penalize the squared cosine similarity between different prototypes. Let denote the matrix whose rows are prototype vectors . We first apply row-wise L₂-norm to obtain , and form the Gram matrix , where equals the cosine similarity between prototypes and . The diversity loss is defined as:

(17)

which corresponds to the mean squared cosine similarity across all pairs of distinct prototypes. Minimizing discourages highly correlated prototypes and thus promotes a more expressive and non-redundant prototype set.

4.9. Experimental design and evaluation protocol

4.9.2. Cross-validation and robustness assessment.

All hyperparameter tuning and model selection were conducted exclusively within the training split to prevent data leakage. We employed a group-aware K-fold cross-validation strategy (default K = 5) on the training partition, ensuring that all cells originating from the same patient or biological sample were confined to a single fold. In cases where group numbers were insufficient for five folds, we adaptively reduced K while preserving group integrity; if no grouping variable was available, we utilized stratified K-fold cross-validation to maintain class balance.

To quantify model robustness against training data variations, we retained the models generated from the K folds (with hyperparameters fixed) and evaluated each model independently on the fixed held-out test set. Performance metrics are reported as the mean and 95% t-interval derived from these K inferences, reflecting training-induced variability (e.g., from data splitting and initialization). This interval is calculated as:

(18)

where s is the standard deviation of the metric across the K runs/folds and K is the number of folds.

4.9.3. Threshold selection.

As described in the loss function section, scRADAR is trained to minimize a weighted composite objective. After training, the raw logit outputs are converted to probabilities via the sigmoid function. To handle potential class imbalance, we optimize the decision threshold adaptively for each cross-validation fold rather than using a default of 0.5.

Specifically, for each fold f, we perform a one-dimensional grid search over the candidate thresholds and select the threshold that maximizes the F1-score on the corresponding validation subset:

(19)

where denotes the ground-truth labels and denotes the predicted probabilities for the f-th validation fold, and is the indicator function. The optimal threshold is then applied to the predictions of the f-th model on the held-out test set. For prospective inference on unlabeled future samples, however, the decision threshold should be pre-specified from development data rather than re-optimized on the incoming sample. In practice, a stable choice is to use the median of the fold-wise optimal thresholds estimated during model development.

4.10. Baseline implementation and benchmarking protocol

To provide a rigorous and comparable benchmark, we evaluated scRADAR against three categories of comparison methods within a unified target-domain evaluation framework. The first category comprised five representative bulk-to-single transfer-learning or domain-adaptation baselines, namely SCAD [30], scATD [31], scDEAL [23], SSDA4Drug [29], and scAdaDrug [24]. Following their original formulations, these methods used bulk RNA-seq data as the source domain and scRNA-seq cohorts as the target domain, with input features aligned by the intersection of shared genes. Source-domain composition followed each method’s official setting: SCAD, SSDA4Drug, and scAdaDrug were trained using GDSC bulk RNA-seq only, whereas scDEAL and scATD additionally incorporated CCLE bulk profiles (i.e., a GDSC+CCLE integrated source), as supported by their original implementations.

The second category was a scGEN-adapted baseline implemented as a single-cell perturbation-oriented comparator. Because the curated cohorts in this study provide cohort-harmonized binary response-associated labels rather than paired perturbation trajectories for every drug condition, scGEN was adapted as a scGEN-style variational latent representation learner using the same cellular pathway representation. The learned latent representation was then used for downstream binary response-phenotype classification under the same target-domain data-splitting and threshold-selection protocol. Mechanism-aware drug fingerprints were not added to this baseline because they are not native to the scGEN-style latent representation setting.

The third category comprised conventional supervised classifiers trained directly on the target single-cell data, including logistic regression (LR) [44], multilayer perceptron (MLP) [46], and XGBoost [47]. These models were trained using the same target-domain training data as scRADAR, without any bulk pretraining, domain adaptation, or external source-domain supervision.

For each target cohort, we first created a group-aware 80/20 holdout split, strictly separating patients, donors, samples, or batches whenever such metadata were available, to define a fixed independent test set. We then performed 5-fold group-aware cross-validation on the remaining 80% training portion for hyperparameter tuning and model selection. In datasets with insufficient numbers of biological groups, we reduced the number of folds while preserving group integrity; when no grouping variable was available, we used stratified K-fold splitting.

To ensure comparability, all methods—transfer-learning baselines, the scGEN-adapted baseline, supervised baselines, and scRADAR—were evaluated under the same target-domain splits, preprocessing pipeline, and decision-threshold protocol. Feature standardization (z-score normalization), when required by a given model, was fit strictly on the training data available within each fold and then applied to the corresponding validation fold and held-out test set, thereby preventing information leakage. For transfer-learning methods, preprocessing was performed using the source-domain data together with the target training fold in a manner consistent with their original design; for target-only supervised models and scRADAR, preprocessing was fit using the target training fold only. Predicted probabilities were converted to binary labels using the same adaptive thresholding procedure for all methods: the decision threshold was selected by grid search to maximize validation F1-score within each fold and then fixed for evaluation on the held-out test set. Final performance is reported as the mean and 95% t-interval across cross-validation-derived runs.

This benchmark design helps disentangle improvements associated with target-domain supervision, single-cell perturbation-style latent representation learning, and scRADAR’s mechanism-guided architecture and prototype-routing design.

4.11. Post hoc interpretability analyses

4.11.1. Model-derived drug-fingerprint score analysis.

After model training, we performed a post hoc analysis of drug-fingerprint score profiles to compare static prior-derived mechanism scores with dynamic perturbation-derived pathway scores for representative drug cohorts. This analysis was used only for interpretation and visualization and was not used for model training, model selection, threshold optimization, or hyperparameter tuning.

For each representative drug cohort, the exported drug-mechanism profile contained two signed score sources. The static score for mechanism or pathway feature m is denoted by , and the dynamic score for the same feature is denoted by . Static scores were derived from curated mechanism annotations, whereas dynamic scores were derived from perturbation-associated pathway profiles represented in the model diagnostic output. Static and dynamic entries referring to the same mechanism or pathway label were mapped to a shared display label. Entries labelled as “Other” or uninformative miscellaneous categories were excluded from visualization. When a feature was present in only one source, its score in the other source was set to zero for visualization.

To focus the plot on the most informative mechanism or pathway features, features were ranked by the larger absolute magnitude of the two score sources:

(20)

The top-ranked features were retained as the displayed feature set for drug cohort d, denoted by . Because static and dynamic scores originate from different information streams and may have different numerical scales, they were standardized separately within each drug-specific displayed profile. For the static source, we computed

(21)

where and are the mean and standard deviation of the static scores across , and is a small constant for numerical stability. The dynamic scores were standardized analogously:

(22)

The paired values and were then visualized in Fig 3 as static prior-derived and dynamic perturbation-derived drug-fingerprint scores, respectively.

These scores should be interpreted as post hoc diagnostic summaries of the hybrid drug-fingerprint module rather than as independent experimental measurements of pathway activation or causal feature attributions. Static scores summarize curated prior mechanism information, whereas dynamic scores summarize perturbation-derived pathway information represented in the model diagnostic output. Therefore, deviations between static and dynamic scores, such as the Hormone Ligand Binding signal in the Cisplatin cohort, were interpreted as candidate context-specific associations only when they were biologically plausible and supported by external literature. Conversely, negative or discordant scores, including RTK signaling or DNA-crosslinker-related features, were treated as directional differences within the diagnostic profile and interpreted cautiously rather than as direct evidence of pathway inhibition or activation.

4.11.2. Post hoc cluster-wise threshold analysis.

Cluster-wise threshold analysis was performed only as a post hoc characterization of local score-label relationships across transcriptional neighborhoods. It was not used for model training, model selection, global threshold selection, or prospective prediction. The model first produced a continuous predicted probability of the Sensitive response phenotype for each cell i. A global decision threshold was selected on the validation set according to the procedure described in Section 4.9.3 and was then fixed for held-out test evaluation.

To examine whether different transcriptional neighborhoods exhibited different local calibration patterns, cells were assigned to the transcriptional clusters visualized in Fig 4. Let c denote a cluster and let denote the set of cells assigned to that cluster. For each cluster, we searched the same candidate-threshold grid used in the global thresholding procedure: .

The post hoc cluster-wise threshold was defined as a threshold that maximized the F1-score within that cluster:

(23)

where denotes the cohort-harmonized response-associated label, with corresponding to the Sensitive label and corresponding to the Resistant label, denotes the predicted probability of the Sensitive response phenotype, and is the indicator function. Clusters with small cell numbers were interpreted cautiously because their F1-optimizing thresholds can be unstable.

Because was estimated using cluster-specific labels after model evaluation, it should not be interpreted as a deployable decision rule for unlabeled future samples. Instead, it summarizes how the relationship between predicted probabilities and response-associated labels varies across transcriptional neighborhoods. A cluster-wise threshold higher than the global threshold indicates that cells in that neighborhood require a higher predicted Sensitive probability to achieve the best local F1-score, whereas a lower threshold indicates a different local score-label calibration pattern. These thresholds were therefore used to visualize heterogeneity in local decision geometry, not to claim independently validated biological subtypes or to adapt predictions in a prospective setting.