acceptodds
Under review as a conference paper at ICLR 2027

Drug Response Prediction: Shifting from Cell Lines to Patients through Pathway-Structured Interaction Learning

Abstract

Drug response models can achieve strong global correlations while remaining less accurate at ranking patients for the same drug. This discrepancy matters when transferring knowledge from cell-line screens to patient cohorts with lim- ited response labels. We propose EviDR, an evidence-guided framework that combines large-scale PRISM supervision with limited BeatAML ex vivo response labels. EviDR maps both domains into a shared Reactome pathway space and in- tegrates molecular drug representations with cross-fitted drug–pathway Evidence derived from source response associations. Reliability-weighted Evidence rele- vance guides a drug-conditioned pathway gate, while signed Evidence interacts with sample pathway states to construct patient-specific representations. Under patient-held-out five-fold evaluation, EviDR achieves the highest mean Macro Spearman of 0.329 among the evaluated methods, compared with 0.314 for Trans- DRP and 0.313 for MAC-CDR. It also achieves 1.790× enrichment of experimen- tally sensitive samples relative to group-specific random selection. Comparisons with random patient–drug pair splitting highlight the importance of evaluating un- seen patients, while pathway analyses identify recurring source associations and show prediction sensitivity to Evidence identity and direction. These findings sup- port pathway-structured supervised transfer for within-drug response ranking and provide an explicit connection between source-derived associations and patient- specific pathway representations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cross-domain learning, transfer learning, drug response prediction, pathway-guided learning, patient stratification, precision oncology

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xG7yOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes EviDR, a framework for predicting drug responses that generalizes from cancer cell lines to individual patients. The method leverages a shared Reactome pathway space, cross-fitted drug–pathway evidence, and drug-conditional interaction representations to enable patient-specific drug response prediction. In experiments using five-fold cross-validation with patients completely held out, EviDR achieved a Macro Spearman coefficient of 0.329, outperforming existing comparison methods.

Critical strengths
  1. The research question has practical significance. This paper focuses on cross-domain drug response prediction from cell lines to patients and emphasizes the ranking of responses among different patients to the same drug, which has clearer practical value compared to merely optimizing global correlation metrics.

  2. The evaluation protocol is rigorous. The authors employed patient-level partitioning to prevent cross-contamination of patient information between training and testing, and used the drug-specific Macro Spearman correlation as the primary metric. Additionally, they demonstrated the importance of the evaluation protocol through randomized paired experiments.

  3. The experimental analysis is thorough. The paper provides multiple baseline comparisons, quintile results, Bootstrap uncertainty analysis, source-domain supervised experiments, and Evidence ablation; overall, the experimental design is comprehensive.

Critical weaknesses

W1. The actual contribution of the core Evidence mechanism is relatively limited. The main innovation of this paper is the use of reliability-weighted, signed drug–pathway Evidence to guide patient-specific interactions. However, as shown in Appendix Table 10, after removing reliability, correlation, or sign information, the Macro Spearman coefficient remains at 0.327–0.328, which is very close to the 0.329 of the full model. In particular, even after randomly shuffling the pathway mappings of the Evidence, the performance still reaches 0.328. This leads me to question whether the biological associations contained in Evidence truly contribute to a meaningful predictive gain.

W2. The performance advantage over existing methods is not very pronounced; could this be due to variations in random seed fluctuations?

EviDR’s Macro Spearman coefficient is 0.329, while TransDRP and MAC-CDR are 0.314 and 0.313, respectively, representing a relatively limited improvement. EviDR outperformed MAC-CDR in only three out of five test setups, and the 95% confidence interval for the performance difference obtained via patient clustering bootstrap is [−0.012, 0.039], which includes zero. Therefore, the current evidence is insufficient to demonstrate that this method has a consistent performance advantage across different patient populations.

W3. The contribution of cross-domain transfer to the performance gains has not yet been fully distinguished from that of the Evidence mechanism.

Appendix Table 10 shows that training using only the target domain alone achieved 0.319; adding PRISM Evidence without joint training on the source domain only improved this to 0.321, while the full model reached 0.329. This suggests that the performance improvement may stem more from the additional source-domain supervision than from the proposed Evidence interaction mechanism. Furthermore, the source-domain data scale experiment in Figure 3 was conducted on only one test fold, which is insufficient to demonstrate the stability of this pattern across patient subdivisions.

Questions for authors

1.Why does the model’s performance show almost no decline after randomly shuffling the mappings between Evidence and pathways? Can the independent contributions of its biological structure and symbolic information to the predictions be further demonstrated?

2.Could you include additional experiments with multiple random seeds and paired statistical analyses against the strongest baseline to further validate the stability of the performance improvement at the patient level?

3.Could you conduct large-scale experiments on source-domain supervision across all five test folds and include a supervised matching baseline that uses the same model capacity but does not incorporate Evidence?

If the author addresses the concerns I raised, I will be willing to update my score.

6 comments
AuthorsOpenReview ↗

Response to Reviewer xG7y (Part 1/3): Evidence Permutation Nulls & Constrained Readout Probes (W1, Q1)

Thank you very much for recognizing the clinical value of our within-drug patient ranking setup, our strict patient-held-out evaluation protocol, and our experimental design, as well as for your willingness to update your score once your questions are addressed. We present our responses in three parts covering: (1) Evidence structure and sign isolation, (2) 3-seed stability and paired statistical comparisons, and (3) 5-fold source-supervision scaling and capacity-matched comparisons (marked in blue in the revised PDF).

Comment 1 (W1 & Q1): Why does the model's performance show almost no decline after randomly shuffling the mappings between Evidence and pathways [in Appendix Table 14b, previously Table 10b]? Can the independent contributions of its biological structure and symbolic information to the predictions be further demonstrated?

Response: We thank the reviewer for this insightful question. We clarify why retraining from scratch compensates for degraded Evidence in Appendix Table 14b (revised Appendix A.7), and provide three experiments isolating the contributions of Evidence's pathway alignment and directional sign:

  1. Why Retraining from Scratch Absorbs Perturbed Evidence: Note that shuffling drug-specific Evidence pathways while keeping patient pathway states xx fixed does not commute with elementwise modulation (x⊙(Pdsd)≠Pd(x⊙sd)x \odot (P_d s_d) \neq P_d(x \odot s_d)), so dense layers cannot simply invert the permutation. Instead, when retrained end-to-end from scratch over PRISM and BeatAML, the full architecture adapts around degraded Evidence via complementary prior and bypass branches: in Eq. (9), the gate gd=σ(Wgzd+bg+pd+ηrrd)g_d = \sigma(W_g z_d + b_g + p_d + \eta_r r_d) combines curated Reactome drug-target priors (pdp_d) and learned molecular gate projections (WgzdW_g z_d) alongside empirical Evidence (rdr_d), while Eq. (12) retains the unmodulated sample and drug representations (zx∥zdz_x \parallel z_d), allowing trainable interaction and fusion layers to rely on alternative signals and down-weight noisy Evidence channels.

  2. Isolating Biological Pathway Alignment and Sign Without Retraining Compensation: To test Evidence information directly without bypass fallback or end-to-end retraining adaptation, we conducted three complementary evaluations (Appendix A.6, Table 13 & Figure 4):

  • Test A (100x Structured PRISM Response-Permutation Null): Recomputing cross-fitted Evidence across 100 permutations of PRISM responses and quality weights within each drug ×\times lineage stratum (preserving exact 5-fold splits and missingness), real PRISM data recovers 3.8x more high-reliability associations (545 vs. 142.6, p=0.0099p = 0.0099) and significantly higher cross-fold overlap among nonzero supported pathways across Top-5/10/20 cutoffs (all p=0.0099p = 0.0099).
  • Test B (Constrained No-Bypass & Frozen-Projection Readout Probes): Removing the unmodulated zxz_x bypass, removing Wgzd+bg+pdW_g z_d + b_g + p_d, and eliminating σ(0)=0.5\sigma(0)=0.5 leakage in a No-Bypass Neural Probe across the 64 source-supported test drugs, real signed Evidence achieves a 4.2x higher coverage-adjusted ranking score (0.1094 +/- 0.0113, active on 64/64 drugs) than the Response-Permuted Null (0.0261 +/- 0.0132, where 40/64 drugs collapse to constants). Furthermore, applying a fixed 32-dim random projection of ux,du_{x,d} concatenated with drug features before Ridge readout (without retraining fintf_{\text{int}}) yields a consistent ordering: Response-Permuted Null (0.0116) < Unsigned (0.0287) < Pathway-Shuffled (0.0357) < Real Signed Evidence (0.0431) (+20.7% over pathway-shuffled and +50.2% over unsigned controls).
  • Test C (Fixed-Model Inference Perturbations): Holding the trained predictor fixed (ηr=2.0241,ηs=1.0100\eta_r = 2.0241, \eta_s = 1.0100), replacing or sign-reversing the ~8.1 / 1,000 active Evidence pathways per drug shifts predictions by 0.731 and 1.006 original BeatAML AUC units on high-quality drugs and flips 1.30%–1.39% of Top-20% prioritized sensitive patients.
Diagnostic Experiment / ConditionMetric / ObservationResponse-Permuted NullUnsigned EvidencePathway-ShuffledReal Signed Evidence (Ours)
100x PRISM Permutation TestRetained Pairs (R≥0.5R \ge 0.5)142.6 pairs——545 pairs (p=0.0099p = 0.0099)
100x PRISM Permutation TestSupported Top-5 / 10 / 20 Overlap19.1% / 18.4% / 16.0%——28.1% / 30.8% / 29.4% (p=0.0099p = 0.0099)
No-Bypass Neural Probe (5-Fold)Coverage-Adjusted Spearman0.0261 +/- 0.0132 (24/64)——0.1094 +/- 0.0113 (64/64, 4.2x)
Frozen-Projection Readout ProbeCoverage-Adjusted Spearman0.0116 +/- 0.01070.0287 +/- 0.01630.0357 +/- 0.01580.0431 +/- 0.0129 (+20.7% / +50.2%)
AuthorsOpenReview ↗

Response to Reviewer xG7y (Part 2/3): 3-Seed x 5-Fold Stability & Paired Statistical Analyses (W2, Q2)

Comment 2 (W2 & Q2): Could you include additional experiments with multiple random seeds and paired statistical analyses against the strongest baseline to further validate the stability of the performance improvement at the patient level?

Response: We thank the reviewer for this suggestion. We performed paired statistical comparisons across the held-out test folds/drugs and evaluated 3 predetermined random seeds (2026, 2027, and 2028) across all 5 outer folds (15 runs per method; Appendix A.4.4, Table 12) against both the overall runner-up TransDRP (0.3139) and MAC-CDR (0.3126):

  1. Paired Statistical Comparisons Against the Strongest Baselines: In Table 2 and Appendix Table 8, the marginal fold-to-fold standard deviation (~0.04) largely reflects shared cohort variation across the 5 patient partitions (for instance, Fold 5 yields the highest Macro Spearman across all 8 models in Table 8). Evaluating paired differences on the exact same held-out test folds accounts for this shared inter-fold variation:
  • EviDR vs. TransDRP (2nd best overall): EviDR outperforms TransDRP in all 5 out of 5 outer folds (+0.013, +0.003, +0.023, +0.014, +0.023), achieving a paired mean gain of +0.0152 (two-sided paired t-test p=0.0164p = 0.0164). EviDR likewise wins 5/5 folds against CaDRReS, CODE-AE, WISER, Drug2TME, DeepSADR, LinClass-7, Elastic Net, and all BeatAML-only per-drug baselines, and 4/5 folds against AITL.
  • EviDR vs. MAC-CDR (3rd best overall): MAC-CDR scores higher in Folds 4–5 (0.345 and 0.404) but drops sharply in Folds 1–3 (e.g., 0.241 vs. EviDR's 0.293 in Fold 3, yielding the largest cross-fold SD of 0.0631). Consequently, while the pooled patient-clustered bootstrap 95% CI (+0.013, [-0.012, +0.039]) includes zero due to this cross-fold variance, EviDR achieves a +0.0165 mean fold gain, outperforms MAC-CDR on 60.6% of the 94 evaluable drugs (paired drug-level bootstrap 95% CI: [+0.002, +0.023]), and ranks 1st across mean NDCG@5/10/20%.
  1. Multi-Seed Robustness Across 15 Runs (3 Seeds x 5 Folds): As shown in the table below, EviDR achieves both the highest overall Macro Spearman (0.3214) and the lowest cross-seed standard deviation (0.0073), outperforming TransDRP on every seed and leading MAC-CDR by +0.0125 overall.
MethodSeed 2026 (5-Fold)Seed 2027 (5-Fold)Seed 2028 (5-Fold)3-Seed Mean +/- Seed SD (15 Runs)Fold Wins (Seed 2026)
TransDRP [AAAI 2025]0.3139 +/- 0.03730.3051 +/- 0.05150.2959 +/- 0.04830.3050 +/- 0.00900 / 5 vs. EviDR (p=0.0164p = 0.0164)
MAC-CDR [PR 2026]0.3126 +/- 0.06310.3164 +/- 0.04790.2976 +/- 0.05710.3089 +/- 0.00992 / 5 vs. EviDR (60.6% drugs win)
EviDR (Ours)0.3291 +/- 0.04000.3146 +/- 0.04110.3205 +/- 0.04580.3214 +/- 0.00735/5 vs. TransDRP; 3/5 vs. MAC-CDR
AuthorsOpenReview ↗

Response to Reviewer xG7y (Part 3/3): 5-Fold Source Scaling, Capacity-Matched Baseline & Clarity (W3, Q3)

Comment 3 (W3 & Q3): Could you conduct large-scale experiments on source-domain supervision across all five test folds and include a supervised matching baseline that uses the same model capacity but does not incorporate Evidence?

Response: We have completed both experiments across all five outer test folds (Appendix A.7.1, Table 15 and Figure 6):

  1. Five-Fold Source-Domain Supervision Scaling: Holding target labels, Evidence, architecture, and validation criteria fixed, we varied the fraction of PRISM observations in the direct source loss across all 5 folds. Full (100%) PRISM supervision achieves the best Macro Spearman (0.3291), best Macro Pearson (0.3200), and lowest five-fold RMSE (46.422, reduced from 46.878 at 0% direct PRISM loss; Figure 6b). Intermediate fractions (50%–75%) exhibit a mild non-monotonic transition due to source–target gradient competition under partial source coverage: at 25%, optimization remains anchored by BeatAML target supervision, whereas at 50%–75% the source loss weight is large enough to steer shared representations while subsampling still fragments cross-lineage diversity and valid within-batch drug pairs (Ps,d\mathcal{P}_{s,d}, Eq. 14), a partial-coverage effect that is resolved at 100% source supervision.

  2. Capacity-Matched Supervised Baseline Without Pathway Gate & Evidence (766,988 vs. 766,989 Parameters): We constructed a Dual-Tower Concatenation MLP (y^x,d=MLP(zx∥zd)\hat{y}_{x,d} = \text{MLP}(z_x \parallel z_d)) that removes the pathway gate and Evidence interaction branch while widening the Fusion MLP to match EviDR within 1 parameter (766,988 vs. 766,989), trained under identical 100% PRISM + BeatAML objectives. Together, Table 15 supports three distinct controlled comparisons:

  • Direct Source-Supervision Gain (Fixed EviDR + Evidence): Increasing direct PRISM supervision from 0% to 100% lifts Macro Spearman from 0.3214 to 0.3291 (+0.0077) and lowers five-fold RMSE from 46.878 to 46.422.
  • Pathway-Gating & Evidence Interaction Gain (Matched Capacity & 100% Supervision): Compared against the 766,988-parameter Concatenation MLP (which still retains EviDR's Reactome pathway space, biological regularizers, and drug-balanced pairwise ranking loss, already beating all 13 external baselines), EviDR's joint pathway-gating and Evidence interaction design improves Macro Spearman from 0.3239 to 0.3291 (+0.0052), winning in 4 out of 5 outer folds (Fold 1: +0.0296, Fold 3: +0.0085, Fold 4: +0.0058, Fold 5: +0.0033) and improving Macro Pearson (0.3200 vs. 0.3160).
  • Overall Transfer Gain: Relative to Target-Only EviDR (0.3194), Full EviDR achieves a +0.0097 Macro Spearman gain, confirming that joint pathway gating and Evidence modulation consistently improve within-drug patient stratification alongside pathway-level biological traceability.
Model / Supervision Setting (All 5 Folds)ParametersMacro Spearman (Up)Macro Pearson (Up)Outer-Fold Wins
Target-Only EviDR (BeatAML only, no PRISM)766,9870.3194 +/- 0.03700.3118 +/- 0.0432—
EviDR Source-Scaling: 0% Direct PRISM Loss (+Evidence)766,9890.3214 +/- 0.04110.3125 +/- 0.0445—
EviDR Source-Scaling: 25% Direct PRISM Supervision766,9890.3231 +/- 0.03370.3156 +/- 0.0437—
EviDR Source-Scaling: 50% Direct PRISM Supervision766,9890.3186 +/- 0.03940.3090 +/- 0.0426—
EviDR Source-Scaling: 75% Direct PRISM Supervision766,9890.3151 +/- 0.03990.3067 +/- 0.0497—
Same-Capacity Concatenation MLP (100% PRISM, w/o Gate/Evidence)766,9880.3239 +/- 0.03610.3160 +/- 0.03711 / 5 vs. EviDR
Full EviDR (100% PRISM + Pathway Gate & Evidence)766,9890.3291 +/- 0.04000.3200 +/- 0.04464 / 5 vs. MLP

Manuscript Readability Revisions: To enhance clarity, we structured Sections 3.1–3.3 into an explicit three-step walkthrough detailing how cross-fitted Evidence (rd,sdr_d, s_d) modulates patient pathway states (xx), added a master notation table in Appendix A.1 (Table 4), clarified the fixed-Evidence protocol in Figure 3, and integrated the structured permutation nulls (Figure 4) as well as the 5-fold scaling curves and capacity-matched comparisons (Appendix A.7.1, Table 15 and Figure 6). We respectfully hope these findings merit an upward revision of your score.

Reviewer_kLTROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors tackle the problem of drug response prediction in patients, using cross-domain data from cell lines and integrating drug pathway evidence. The approach EviDR, combines drug features, pathway evidence and reactome representations to obtain drug responses on the BeatAML public dataset. The key objective is to rank patients on the same drug.

Critical strengths
  1. The authors have tackled a relevant problem, which is an ongoing area of research.

  2. Apart from comparisons against other drug response prediction approaches, the authors have also presented the cytarabine case study, to aid in biological interpretation of the results.

  3. The authors have reported mean and standard deviation given the size of datasets they are working with.

Critical weaknesses

There are several unanswered questions (please refer to the section on Questions for details), but the key weakness is related to the benchmarking against recent SOTA methods and also generalization to non-AML datasets. In addition, in table 2, for most metrics, the standard deviation of EviDR overlaps and is close to the next best method.

Questions for authors
  1. The authors have shown their approach on AML, but does this generalize to others like the TCGA-LAML cohort?

  2. Does this approach also generalize to non-AML datasets?

  3. How does EviDR compare against other approaches that also deal with cell line to patient drug response prediction for AML, like Zeng et al[1]?

  4. The evaluation metric is the Pearson and Spearman correlation, however the p-values are not reported. Please provide this to address the issue of significance, given the small dataset size.

  5. How does EviDR compare against general cell line to patient drug response prediction methods like THERAPI[2], GANDALF[3], Drug2TME[4], DeepSADR[5] (although mentioned in the related work, it is not added to table 2)?

  6. How does the drug pathway response association mining compare with the approach in DPADM[6]?

  7. Was any anaylsis done apriori to show that drug-pathway associations hold true across the datasets?

  8. In table 1, why is the result for Pair-random lower for EviDR, as compared to MAC-CDR?

  9. How do ranking methods like NDCG fare for these methods?

[1] Zeng, A. G., Bansal, S., Jin, L., Mitchell, A., Chen, W. C., Abbas, H. A., ... & Dick, J. E. (2022). A cellular hierarchy framework for understanding heterogeneity and predicting drug response in acute myeloid leukemia. Nature medicine, 28(6), 1212-1223.

[2] Sung, I., Bang, D., Kim, S., & Lee, S. (2026). Deep learning for predicting patient drug response by transferring gene-level and cell-level knowledge to tumors. npj Precision Oncology.

[3] Jayagopal, A., Zhang, Y., Walsh, R., Tan, T., Jeyasekharan, A., & Rajan, V. (2025, May). GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment. In International Conference on Learning Representations (Vol. 2025, pp. 41617-41637).

[4] Zhai, J., & Liu, H. (2024). Cross-domain feature disentanglement for interpretable modeling of tumor microenvironment impact on drug response. IEEE Journal of Biomedical and Health Informatics, 28(7), 4382-4392.

[5] Zhang, Y., Huang, Z., Fan, Z., Shen, S., Li, Y., Wu, S., ... & Deng, L. (2026, April). DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction. In International Conference on Learning Representations (Vol. 2026, pp. 116485-116513).

[6] Wang, Y., Qi, J., & Ai, D. (2023). DPADM: a novel algorithm for detecting drug-pathway associations based on high-throughput transcriptional response to compounds. Briefings in Bioinformatics, 24(1), bbac517.

6 comments
AuthorsOpenReview ↗

Response to Reviewer kLTR (Part 1/3): Zero-Shot TCGA-LAML Risk Association & Breast Cancer PDTC Transfer (Q1, Q2)

We sincerely thank you for your expert review and for recognizing the clinical relevance of within-drug patient ranking, our statistical reporting, and our Cytarabine biological case study. Following your nine constructive questions, we expanded our benchmarks to include recent 2024–2026 transfer methods (DeepSADR [ICLR 2026], Drug2TME [JBHI 2024]) and the AML cellular hierarchy signature (Zeng et al. [Nature Medicine 2022]), evaluated NDCG@5/10/20%, examined cross-dataset association compatibility, and evaluated zero-shot frozen EviDR ensembles on both clinical TCGA-LAML (N=173N=173) and non-AML Breast Cancer PDTC solid tumors. We present our responses in three parts below (marked in blue in the revised PDF).

Comment 1 (Q1 & Q2): 1. The authors have shown their approach on AML, but does this generalize to others like the TCGA-LAML cohort? 2. Does this approach also generalize to non-AML datasets?

Response: We thank the reviewer for this valuable suggestion. We evaluated frozen 5-fold EviDR ensembles zero-shot on both TCGA-LAML and the non-AML Bruna Breast Cancer PDTC pharmacogenomic biobank (Appendix A.9, Table 16):

  1. Clinical-Risk Association in TCGA-LAML (N=173N = 173 Patients): Because TCGA-LAML is a clinical cohort without unconfounded single-agent ex vivo AUC matrices—and clinical overall survival is heavily confounded by multi-agent regimens, stem-cell transplantation, and patient age (univariable Cox p=0.274p = 0.274)—we evaluated whether zero-shot EviDR predicted Cytarabine resistance scores associate with established pre-treatment CALGB cytogenetic risk tiers (testing clinical-risk alignment rather than direct clinical treatment efficacy):
  • CALGB Poor- vs. Favorable-Risk Separation: Patients in the CALGB Poor-Risk tier exhibit significantly higher zero-shot predicted Cytarabine resistance (median 0.4072) than patients in the Favorable-Risk tier (median 0.3679, two-sided p=0.0311p = 0.0311, n=68n = 68).
  • Rank Correlation Across All Risk Tiers: Across all n=171n = 171 risk-annotated patients, predicted Cytarabine resistance correlates positively with worsening cytogenetic risk category (Spearman ρ=0.1805,p=0.0181\rho = 0.1805, p = 0.0181), supporting biological alignment with clinical chemoresistance risk.
  1. Cross-Cancer Drug-Response Transfer to Non-AML Solid Tumors (Bruna Breast Cancer PDTCs): To directly evaluate ex vivo drug-response predictions outside AML, we tested the Bruna breast cancer patient-derived tumor cell (PDTC) biobank (25 expression profiles across 15 independent patient-origin groups, 40 PRISM-matched drugs, 646 ex vivo responses; 34 evaluable drugs), strictly grouping passage replicates (e.g., STG139/STG139M) by patient origin:
  • Zero-Shot Cross-Cancer Transfer: Directly applying the frozen EviDR ensemble achieves positive within-drug patient ranking (Macro Spearman = 0.1485, Macro Pearson = 0.1368) and a 1.2964x enrichment of Top-20% experimentally sensitive samples (precision 0.3113; remaining positive at Macro Spearman = 0.1221 and 1.2369x enrichment under unlabeled cohort expression alignment).
  • Small-Cohort Fine-Tuning Diagnostic (N=15N = 15 groups): A complementary 5-fold fine-tuning check (~3 test groups/fold) likewise confirms positive within-fold correlation (+0.0488), showing that EviDR's shared Reactome representation transfers positively to solid tumors.
External CohortSetting / EndpointSample / Drug ScaleEvaluation MetricResult / Significance
TCGA-LAMLClinical AML Risk Association173 patients (n=68n=68 Poor vs. Fav)CALGB Poor vs. Favorable RiskMedian 0.4072 vs. 0.3679 (p=0.0311p = 0.0311)
TCGA-LAMLClinical AML Risk Association171 risk-stratified patientsRisk-Tier Spearman Correlationρ=0.1805\rho = 0.1805 (p=0.0181p = 0.0181)
Bruna PDTC (Zero-Shot)Breast Cancer Ex Vivo Transfer25 samples / 15 groups / 34 drugsWithin-Drug Macro Spearman / Pearson0.1485 / 0.1368 (Positive Transfer)
Bruna PDTC (Zero-Shot)Breast Cancer Ex Vivo Transfer25 samples / 15 groups / 34 drugsTop-20% Sensitive Enrichment1.2964x (Precision 0.3113)
AuthorsOpenReview ↗

Response to Reviewer kLTR (Part 2/3): Comparisons with Zeng et al., Drug2TME, DeepSADR & NDCG Metrics (Q3, Q5, Q9)

Comment 2 (Q3, Q5 & Q9): 3. How does EviDR compare against Zeng et al. [1]? 5. How does EviDR compare against THERAPI [2], GANDALF [3], Drug2TME [4], DeepSADR [5]? 9. How do ranking methods like NDCG fare for these methods?

Response: We thank the reviewer for pointing out these important benchmarks and metrics. Under the exact same 5-fold patient-held-out protocol, we evaluated Zeng et al. [1], Drug2TME [4], and DeepSADR [5], and computed tie-aware NDCG@5, NDCG@10, and NDCG@20% (top 20% samples per drug, using normalized within-drug sensitivity rank as relevance gain) across all methods (Appendix A.4.3, Table 10):

  • Zeng et al. [1] (Nature Medicine 2022, Q3): Using the authors' official repository, we restored 4 signature genes outside the PRISM-shared intersection from the full BeatAML RNA-seq matrix and evaluated the published LinClass-7 cellular hierarchy signature with training-patient per-drug linear calibration. It achieves a Macro Spearman of 0.1188 +/- 0.0260 (vs. 0.3291 for EviDR), indicating that a fixed 7-gene hierarchy score is insufficient to rank patient responses across 99 mechanistically diverse compounds.
  • Drug2TME [4] & DeepSADR [5] (Q5): We adapted the official repositories of Drug2TME (retaining shared/private encoders, orthogonality/adversarial objectives, and pretrained molecular graph encoders) and DeepSADR (retaining BRICS substructures, GNN+, variational interaction graph, and SetTransformer readout over 13 label-free Reactome gene groups) to continuous AUC regression. They achieve Macro Spearman scores of 0.1461 +/- 0.0142 and 0.1232 +/- 0.0456, respectively, substantially below EviDR (0.3291).
  • THERAPI [2] & GANDALF [3] (Q5): Code inspection confirmed that THERAPI requires external pretrained post-perturbation gene-ranking representations and GANDALF requires multi-omic somatic mutation augmentation outside our baseline transcriptomics + molecular fingerprint modality boundary; we clarify these scope differences in Section 2.2 and Appendix A.4.1.
  • NDCG Ranking Evaluation (Q9): EviDR achieves the highest mean score across NDCG@5 (0.6936), NDCG@10 (0.7022), and NDCG@20% (0.6933), confirming strong patient prioritization at the sensitive head of each drug's ranking list.
MethodVenue / ParadigmMacro Spearman (Up)Macro Pearson (Up)NDCG@5 (Up)NDCG@10 (Up)NDCG@20% (Up)
Zeng et al. (LinClass-7)Nat. Med. 2022 (Hierarchy)0.1188 +/- 0.02600.1154 +/- 0.0261———
DeepSADR [5]ICLR 2026 (Subsequence)0.1232 +/- 0.04560.1268 +/- 0.04460.5661 +/- 0.04500.5897 +/- 0.03880.5857 +/- 0.0364
Drug2TME [4]JBHI 2024 (TME Disentangle)0.1461 +/- 0.01420.1509 +/- 0.01340.5694 +/- 0.03060.5984 +/- 0.02390.5954 +/- 0.0247
Elastic Net (Joint)Linear Baseline0.2198 +/- 0.04280.2219 +/- 0.04390.6068 +/- 0.02360.6317 +/- 0.02910.6344 +/- 0.0185
AITLBioinformatics 20200.2940 +/- 0.05720.2877 +/- 0.06120.6598 +/- 0.05690.6764 +/- 0.04400.6682 +/- 0.0371
CaDRReSBioinformatics 20180.3016 +/- 0.04880.2951 +/- 0.04350.6701 +/- 0.03450.6863 +/- 0.03020.6844 +/- 0.0238
WISERarXiv 20240.3042 +/- 0.04060.3054 +/- 0.03590.6783 +/- 0.01440.6915 +/- 0.01640.6836 +/- 0.0161
CODE-AENat. Mach. Intell. 20220.3080 +/- 0.03820.3043 +/- 0.04220.6671 +/- 0.04240.6848 +/- 0.03450.6778 +/- 0.0264
MAC-CDRPattern Recognit. 20260.3126 +/- 0.06310.3125 +/- 0.06600.6860 +/- 0.03430.6971 +/- 0.03170.6892 +/- 0.0318
TransDRPAAAI 20250.3139 +/- 0.03730.3090 +/- 0.04290.6747 +/- 0.03740.6856 +/- 0.02830.6792 +/- 0.0222
EviDR (Ours)Pathway-Evidence Transfer0.3291 +/- 0.04000.3200 +/- 0.04460.6936 +/- 0.02670.7022 +/- 0.02480.6933 +/- 0.0219
AuthorsOpenReview ↗

Response to Reviewer kLTR (Part 3/3): Significance p-Values, DPADM, Cross-Dataset Compatibility & Protocol Gap (Q4, Q6–Q8)

Comment 3 (Q4, Q6, Q7 & Q8): 4. Please provide p-values to address significance, given the small dataset size and overlapping SDs. 6. How does association mining compare with DPADM [6]? 7. Was any analysis done a priori to show that drug-pathway associations hold true across datasets? 8. Why is Pair-random lower for EviDR than MAC-CDR in Table 1?

Response:

  • Statistical Significance & Overlapping Fold SDs (Q4): Pooled Global Spearman is highly significant (r=0.695,p<10−100r = 0.695, p < 10^{-100}), and across all 470 evaluable drug–fold test groups, individual within-drug Spearman correlations have a median two-sided pp-value of 0.0028 (IQR: [2.21×10−5,0.0625][2.21 \times 10^{-5}, 0.0625]; Appendix A.4.4). Moreover, the ~0.04 SD in Table 2 is the marginal between-fold SD across the 5 patient partitions (reflecting shared fold-to-fold cohort variation). When paired on identical test folds, EviDR beats the overall runner-up TransDRP in all 5/5 folds (paired t-test p=0.0164p = 0.0164), wins 5/5 folds over 10 other baselines and 4/5 over AITL, beats MAC-CDR on 60.6% of drugs (paired drug-level bootstrap 95% CI: [+0.002, +0.023], while patient-clustered CI is [-0.012, +0.039]), and leads across 3 random seeds (0.3214 +/- 0.0073).
  • Comparison with DPADM [6] (Q6): DPADM (Wang et al., Brief. Bioinform. 2023) mines drug–pathway associations from post-treatment transcriptional perturbations (e.g., L1000 differential expression after compound exposure) via conditional dependence tests. In clinical pre-treatment prediction, post-treatment profiles are unavailable prior to therapy. EviDR instead estimates lineage-deconfounded associations between pre-treatment baseline pathway activities and source response AUCs via cross-fitting to guide a patient-specific predictor. We added this comparison in Section 2.1.
  • Exploratory Cross-Dataset Association Analysis (Q7): Using strictly BeatAML training patients (zero test-label access), we examined univariate drug–pathway sign agreement with PRISM Evidence across the 46 eligible Evidence-supported drugs (Appendix A.6.3). Mean sign agreement is 54.8% (drug-bootstrap 95% CI: [44.9%, 64.6%]) and 59.6% (95% CI: [47.3%, 71.5%]) across the broader two-thirds drug stratum (marginal association Spearman ρ=0.050\rho = 0.050). Because both confidence intervals include 50% and marginal rank correlation is near zero, univariate cell-line associations do not transfer as conserved 1:1 linear effects in patient cohorts. This exploratory finding is consistent with feeding Evidence into a jointly trained nonlinear multi-pathway interaction encoder (fintf_{\text{int}}) instead of applying fixed univariate weights.
  • Pair-Random vs. Patient-Held-Out Gap (Q8): Under Pair-random splitting, a test patient's RNA-seq profile already appears in the training set paired with ~60 other drugs. Memory-anchor models like MAC-CDR can exploit observed patient-identity sensitivity under Pair-random splitting (0.609) before dropping more steeply (-0.296, to 0.313) when evaluated on unseen patients. EviDR's 1,000-pathway bottleneck and cross-fitted Evidence curb patient-identity overfitting, yielding stronger generalization under strict patient-held-out evaluation (0.329 vs. 0.313).

We hope these additions fully resolve your concerns and earn your support.

Reviewer_pxE5OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper tackles predicting how AML patient samples respond to drugs (BeatAML ex vivo data) by transferring knowledge from a much larger cell-line screen (PRISM). The proposed model, EviDR, maps both datasets into a shared space of 1,000 Reactome pathway scores. From PRISM the authors computes evidence. For each drug, how strongly each pathway correlates with response. Cross-fitting avoids label leakage, and a reliability score down-weights shaky associations. Evidence then decides which pathways the model focuses on for each drug, and is combined with each patient’s pathway profile and a drug fingerprint encoder.

Critical strengths
  1. The evaluation setup is well thought out
  2. The results are transparent. The paper includes fold-level scores, two kinds of bootstrap intervals, a data-scaling experiment and component ablations. It keeps them in even where they don’t flatter the method, such as losing to MAC-CDR in two folds and near-zero ablation effects. The appendix is well detailed.
  3. Paper is well written and well motivated
Critical weaknesses
  1. Table 10 in the appendix undercuts the need for evidence. Removing reliability, relevance or signed evidence changes Macro Spearman by at most 0.002. The same holds even when shuffling which pathway each value belongs to.
  2. The headline result is 0.329 vs 0.314 (Table 2), against a fold-to-fold SD of about 0.04. The patient-clustered bootstrap interval vs MAC-CDR, [−0.012, 0.039] (Sec. 4.2.1, Fig. 2c), includes zero, and EviDR loses to MAC-CDR in Folds 4 and 5 (2/5) (Table 7)
  3. The submission's method is a collection of methods from prior work, where the novelty of using evidence to gate pathways is undercut by table 7
Questions for authors
  1. Under the same patient folds, can you add per-drug BeatAML-only models (e.g., elastic net and random forest per drug), a tuned tree ensemble, and a naive baseline? DrEval’s pipeline already supports a BeatAML2 dataset

  2. What does Evidence actually contribute? 2.1 What are the learned values of η_r and η_s? 2.2 Are the Table 9 perturbation effects (0.50–1.01) really on the original AUC scale? 2.3 Can you redo the stability and reliability analyses against a permutation null, i.e., Evidence computed from shuffled PRISM responses with the same fold structure?

7 comments
AuthorsOpenReview ↗

Response to Reviewer pxE5 (Part 1/3): Architectural Insights & DrEval BeatAML-Only Baselines (W3, Q1)

We sincerely appreciate your recognition of our well-thought-out evaluation setup, transparent reporting of all ablations and fold scores, and clear writing. To directly address your concerns regarding what EviDR contributes beyond combining existing ideas (W3) and whether Evidence and our ranking gains are statistically meaningful (W1, W2), we present our new experiments and clarified statistical interpretations in three parts below (marked in blue in the revised PDF).

Overview on W3 (What EviDR Contributes Beyond Combining Prior Components): Our new controlled experiments establish two concrete findings:

  1. Task-Aligned Pathway Bottleneck + Within-Drug Ranking Objective: Even Target-Only EviDR (0.3194, trained on BeatAML alone without PRISM; Table 11 & Table 15) already outperforms all 7 original baselines and all 6 newly added transfer, AML-hierarchy, and DrEval per-drug baselines (all 13 external predictive baselines: 0.1188–0.3139).
  2. Controlled Gain Over Capacity-Matched Unstructured Transfer: Under identical 100% PRISM + BeatAML supervision, Full EviDR (766,989 params) outperforms an exact capacity-matched Dual-Tower Concatenation MLP (766,988 params) on within-drug ranking (0.3291 vs. 0.3239, winning 4/5 folds; Macro Pearson 0.3200 vs. 0.3160). Thus, joint pathway gating and Evidence modulation specifically improve within-drug patient stratification rather than pooled error alone.

Comment 1 (Q1): Under the same patient folds, can you add per-drug BeatAML-only models (e.g., elastic net and random forest per drug), a tuned tree ensemble, and a naive baseline? DrEval's pipeline already supports a BeatAML2 dataset.

Response: Following DrEval's protocol, we implemented all four requested BeatAML-only models under the exact same 5-fold patient-held-out splits (fitted strictly on training patients, with hyperparameters selected on validation patients; Appendix A.4.3, Table 11):

  • Constant Naive Drug-Mean Reference: Predicting each drug's BeatAML training mean yields mathematically undefined within-drug Spearman correlation (zero patient-ranking capacity), yet achieves 0.6411 Global Spearman purely from between-drug potency differences, concretely illustrating why pooled correlation masks within-drug stratification failure.
  • BeatAML-Only Per-Drug & Tree Predictors: Per-Drug Elastic Net (0.2757 +/- 0.0302) and Per-Drug Random Forest (0.2771 +/- 0.0235) substantially outperform joint concatenated linear Elastic Net (0.2198), while Validation-Tuned Gradient Boosting achieves 0.2433 +/- 0.0206. Nevertheless, all fall well below both Target-Only EviDR (0.3194) and Full EviDR (0.3291 +/- 0.0400, +0.0520 over the best per-drug model).
Method (5-Fold Patient-Held-Out)Training DataMacro Spearman (Up)Global Spearman (Up)RMSE (Down)
Naive Drug-Mean ReferenceBeatAML-onlyUndefined (Constant within drug)0.6411 +/- 0.030149.228 +/- 2.966
Elastic Net (Joint Concatenated)PRISM + BeatAML0.2198 +/- 0.04280.6681 +/- 0.030648.113 +/- 3.162
Validation-Tuned Gradient BoostingBeatAML-only0.2433 +/- 0.02060.6783 +/- 0.025547.214 +/- 2.848
Per-Drug Elastic NetBeatAML-only0.2757 +/- 0.03020.6539 +/- 0.022249.316 +/- 2.027
Per-Drug Random ForestBeatAML-only0.2771 +/- 0.02350.6877 +/- 0.024146.308 +/- 2.716
Target-Only EviDR (Ours, w/o PRISM)BeatAML-only0.3194 +/- 0.0370——
Full EviDR (Ours)PRISM + BeatAML0.3291 +/- 0.04000.6952 +/- 0.021346.422 +/- 2.560
AuthorsOpenReview ↗

Response to Reviewer pxE5 (Part 2/3): Learned Scales, Original AUC Perturbations & 100x Permutation Null (Q2, W1)

Comment 2 (Q2.1, Q2.2, Q2.3 & W1/W3): 2. What does Evidence actually contribute? 2.1 What are the learned values of η_r and η_s? 2.2 Are the Table 9 perturbation effects (0.50–1.01) really on the original AUC scale? 2.3 Can you redo the stability and reliability analyses against a permutation null, i.e., Evidence computed from shuffled PRISM responses with the same fold structure?

Response: We address each sub-question directly below (Appendix A.6 & A.7, Figure 4, Tables 9 & 13):

  1. Learned Values of η_r and η_s (Q2.1): Across the 5 outer-fold checkpoints, the softplus-transformed Evidence scaling parameters converge consistently to substantial positive values: η_r = 2.0241 ± 0.0135 and η_s = 1.0100 ± 0.0130.

  2. Original AUC Scale & Within-Drug Ranking Impact of Perturbations (Q2.2): We verified in our evaluation pipeline that inverse standardization (inverse_labels) is applied before saving predictions; thus, the perturbation shifts in Appendix Table 9, namely 0.503 / 0.599 overall and 0.731 / 1.006 on high-quality drugs (upper third ranked by mean(|E_d| · R_d) per fold), are strictly in original BeatAML AUC units. Why are these shifts about 0.5–1.0 AUC units while pooled RMSE is about 46.4? Pooled AUC variance is dominated by large between-drug potency offsets, whereas within-drug prediction spread for a fixed compound is much narrower and Evidence is sparse (about 8.1 / 1,000 active pathways per drug). Evaluating within-drug patient orderings directly shows that perturbing just these 8.1 active pathways per drug without retraining flips 0.56%–0.58% of all pairwise patient rankings and replaces 1.30%–1.39% of Top-20% prioritized sensitive patients.

  3. 100x Structured PRISM Response-Permutation Null & Constrained Probes (Q2.3 & W1):

  • 100x Permutation Null: Within each drug × lineage stratum, we permuted PRISM responses and observation weights 100 times while preserving exact 5-fold cross-fitting splits, pathway matrices, and missingness, computing Top-K overlap over nonzero supported associations to avoid zero-padding artifacts (updated Figure 4a). Real PRISM responses recover 545 retained pairs vs. 142.6 under the null (3.8x enrichment, p = 0.0099), higher mean reliability (0.648 vs. 0.603, p = 0.0099), and significantly higher supported Top-5/10/20 overlaps (all p = 0.0099).
  • Why Retraining Absorbs Shuffling & What Constrained Probes Show: When retraining the full model from scratch (Appendix Table 14b), unmodulated sample/drug branches (z_x || z_d), Reactome target priors (p_d), and molecular gate projections (W_g z_d) allow trainable interaction/fusion layers to adapt around degraded Evidence. When we remove these bypass branches in a No-Bypass Neural Probe, Real Signed Evidence achieves a 4.2x higher coverage-adjusted ranking score (0.1094 vs. 0.0261) than the Response-Permuted Null (producing active rankings on 64/64 vs. 24/64 drugs). Likewise, in a Frozen-Projection Ridge Readout Probe (without retraining f_int), Real Signed Evidence (0.0431) outperforms Pathway-Shuffled (0.0357, +20.7%), Unsigned (0.0287, +50.2%), and Null (0.0116) controls.
Diagnostic Experiment / ConditionMetric / ObservationResponse-Permuted NullUnsigned EvidencePathway-ShuffledReal Signed Evidence (Ours)
100x PRISM Permutation TestRetained Pairs (R >= 0.5)142.6 pairs——545 pairs (p = 0.0099)
100x PRISM Permutation TestSupported Top-5 / 10 / 20 Overlap19.1% / 18.4% / 16.0%——28.1% / 30.8% / 29.4% (p = 0.0099)
No-Bypass Neural Probe (5-Fold)Coverage-Adjusted Spearman0.0261 +/- 0.0132 (24/64)——0.1094 +/- 0.0113 (64/64, 4.2x)
Frozen-Projection Readout ProbeCoverage-Adjusted Spearman0.0116 +/- 0.01070.0287 +/- 0.01630.0357 +/- 0.01580.0431 +/- 0.0129 (+20.7% / +50.2%)
AuthorsOpenReview ↗

Response to Reviewer pxE5 (Part 3/3): Patient-Clustered vs. Paired Fold, Multi-Seed & Drug-Level Evidence (W2)

Comment 3 (W2): The headline result is 0.329 vs 0.314 (Table 2), against a fold-to-fold SD of about 0.04. The patient-clustered bootstrap interval vs MAC-CDR, [-0.012, 0.039], includes zero, and EviDR loses to MAC-CDR in Folds 4 and 5 (2/5) (Table 8, previously Table 7).

Response: We appreciate this careful statistical critique and address both points directly (revised Section 4.2.1 & Appendix A.4.4):

  1. Explicitly Addressing the Patient-Clustered Bootstrap vs. MAC-CDR: You are completely right that under patient-clustered resampling of pooled out-of-fold predictions, the 95% confidence interval against MAC-CDR (+0.013, [-0.012, +0.039]) includes zero, and our drug-level bootstrap interval ([+0.002, +0.023]) uses a different resampling unit (compounds rather than patient clusters) and therefore complements rather than replaces the patient-clustered result. Why is the patient-clustered interval wide against MAC-CDR? In Table 8, MAC-CDR exhibits the highest fold-to-fold sensitivity among all models (SD = 0.0631), scoring higher on held-out Folds 4–5 (0.345 and 0.404) but dropping sharply on Folds 1–3 (e.g., 0.241 vs. EviDR's 0.293 in Fold 3). We state this explicitly in the revised text.

  2. Complementary Paired Fold, Multi-Seed, Compound-Level & NDCG Evidence: At the same time, four complementary evaluations consistently support EviDR's ranking advantage and stability:

  • 5/5 Fold Consistency Over the Actual Overall Runner-Up (TransDRP): Note that in Table 2, the runner-up baseline with the highest mean Macro Spearman (0.314) is TransDRP (0.3139 +/- 0.0373), followed closely by MAC-CDR (0.3126 +/- 0.0631). Because the marginal SD of about 0.04 largely reflects shared fold-to-fold cohort variation (e.g., Fold 5 yields the highest score across all 8 models in Table 8), paired fold comparison is informative: EviDR beats TransDRP in all 5 out of 5 outer folds (+0.013, +0.003, +0.023, +0.014, +0.023; two-sided paired t-test p = 0.0164), wins 5/5 folds over CaDRReS, CODE-AE, WISER, Drug2TME, DeepSADR, LinClass-7, joint Elastic Net, and all BeatAML-only per-drug models, and wins 4/5 folds over AITL.
  • 3-Seed x 5-Fold Stability (15 Runs per Method): Repeating full 5-fold training across three predetermined random seeds (2026, 2027, and 2028; Appendix Table 12), EviDR achieves the highest 3-seed mean Macro Spearman and lowest cross-seed SD (0.3214 +/- 0.0073), outperforming TransDRP (0.3050 +/- 0.0090) on every seed and leading MAC-CDR (0.3089 +/- 0.0099) by +0.0125 overall.
  • Compound-Level Breadth & Top-Heavy NDCG: Across all 94 evaluable drugs, EviDR outperforms MAC-CDR on 60.6% of compounds (paired drug-level bootstrap 95% CI: [+0.002, +0.023]), and achieves the highest five-fold mean across NDCG@5 (0.6936 vs. 0.6860), NDCG@10 (0.7022 vs. 0.6971), and NDCG@20% (0.6933 vs. 0.6892) (Appendix Table 10).
MethodFold 1Fold 2Fold 3Fold 4Fold 55-Fold Mean +/- SDEviDR Fold WinsPaired Diff (EviDR - Method)
CODE-AE0.3260.2930.2650.2920.3640.3080 +/- 0.03825 / 5+0.0211
MAC-CDR0.2910.2820.2410.3450.4040.3126 +/- 0.06313 / 5+0.0165 (60.6% drugs win)
TransDRP0.3420.2940.2700.3020.3620.3139 +/- 0.03735 / 5+0.0152 (p = 0.0164)
EviDR (Ours)0.3550.2970.2930.3160.3850.3291 +/- 0.0400——
MethodSeed 2026Seed 2027Seed 20283-Seed Mean +/- SDDrug-Level 95% CIPatient-Clustered 95% CINDCG@5 / @10 / @20%
TransDRP0.31390.30510.29590.3050 +/- 0.0090——0.6747 / 0.6856 / 0.6792
MAC-CDR0.31260.31640.29760.3089 +/- 0.0099[+0.002, +0.023][-0.012, +0.039]0.6860 / 0.6971 / 0.6892
EviDR (Ours)0.32910.31460.32050.3214 +/- 0.0073Wins 60.6% drugs+0.013 mean diff0.6936 / 0.7022 / 0.6933

We hope our forthright statistical framing and the full suite of DrEval baselines, permutation nulls, and controlled probes address your reservations and merit a positive reassessment of our work.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Response to Reviewers

Dear Area Chair,

Thank you and the reviewers for your constructive feedback, which helped us substantially strengthen both the empirical scope and the statistical precision of our manuscript. Across over 100 new training and diagnostic runs strictly under our patient-held-out five-fold protocol, we have addressed every reviewer question and synchronized all interpretations in the revised paper (marked in blue in the revised PDF):

1. Expanded Benchmarks, Recent SOTA Baselines & Ranking Metrics (Reviewers kLTR & pxE5)

  • New Transfer & AML Baselines: We added comparisons with recent transfer methods DeepSADR [ICLR 2026: 0.1232] and Drug2TME [JBHI 2024: 0.1461], plus the AML cellular hierarchy signature LinClass-7 [Zeng et al., Nature Medicine 2022: 0.1188] (requested by Reviewer kLTR).
  • DrEval-Aligned BeatAML-Only Baselines: Following DrEval protocols, we added BeatAML-only Per-Drug Elastic Net (0.2757), Per-Drug Random Forest (0.2771), and Validation-Tuned Gradient Boosting (0.2433), plus a constant Naive drug-mean reference (0.6411 Global Spearman despite zero within-drug ranking; requested by Reviewer pxE5). Even Target-Only EviDR (0.3194) and Full EviDR (0.3291) surpass all 13 external predictive baselines (0.1188–0.3139) (Appendix Tables 10–11).
  • Top-Heavy Ranking (NDCG): Across tie-aware NDCG@5 (0.6936), NDCG@10 (0.7022), and NDCG@20% (0.6933), EviDR ranks 1st among all evaluated baselines (requested by Reviewer kLTR; Appendix Table 10).

2. Controlled Source-Supervision Scaling, Capacity Matching & Evidence Probes (Reviewers xG7y & pxE5)

  • 5-Fold Source Scaling & Capacity-Matched Comparison: Across all 5 folds (Appendix A.7.1, Table 15 & Figure 6), increasing direct PRISM supervision within fixed EviDR lifts Macro Spearman from 0.3214 (0% loss) to 0.3291 (+0.0077; +0.0097 over 0.3194 Target-Only EviDR) and lowers RMSE to 46.422. Under identical 100% PRISM + BeatAML losses, EviDR (766,989 params) also beats an exact capacity-matched unstructured Concatenation MLP (766,988 params) on within-drug ranking (0.3291 vs. 0.3239, winning 4/5 folds; Macro Pearson 0.3200 vs. 0.3160), even though the MLP attains slightly lower pooled RMSE (46.151), highlighting the distinction between pooled fit and within-drug stratification.
  • 100x Structured PRISM Permutation Null & Constrained Readout Probes: Across 100 drug ×\times lineage response permutations preserving exact 5-fold splits, real PRISM data recovers 3.8x more high-reliability associations (545 vs. 142.6, p=0.0099p = 0.0099) and higher supported Top-5/10/20 cross-fold overlaps (all p=0.0099p = 0.0099; Figure 4a). Because end-to-end retraining adapts around perturbed Evidence through unmodulated sample/drug and target-prior branches, we evaluate Evidence directly via fixed-model perturbations (Figure 4b) and constrained probes without bypasses (Appendix Table 13): a No-Bypass Neural Probe achieves a 4.2x gain over Null (0.1094 vs. 0.0261), and a Frozen-Projection Ridge Probe yields Null (0.0116) < Unsigned (0.0287) < Pathway-Shuffled (0.0357) < Real Signed Evidence (0.0431).

3. Multi-Seed Stability & Paired Statistical Comparisons (Reviewers xG7y, pxE5 & kLTR)

  • Paired Fold & Resampling Analyses: EviDR beats the overall runner-up TransDRP (0.3139) in all 5/5 outer folds (paired t-test p=0.0164p = 0.0164), wins 5/5 folds over 10 additional baselines, and wins 4/5 folds over AITL. Against MAC-CDR (0.3126, fold SD 0.0631), while the patient-clustered bootstrap 95% CI ([-0.012, +0.039]) crosses zero due to Folds 4–5 variance, EviDR wins on 60.6% of evaluable drugs (drug-level 95% CI: [+0.002, +0.023]), leads on NDCG@5/10/20%, and achieves the highest 3-seed mean and lowest seed SD across 15 runs (0.3214 +/- 0.0073 vs. 0.3089 for MAC-CDR and 0.3050 for TransDRP; Appendix Table 12).

4. External Clinical-Risk Association & Solid-Tumor Transfer (Reviewer kLTR)

  • TCGA-LAML (N=173N = 173) & Bruna Breast Cancer PDTCs: Zero-shot frozen EviDR cytarabine resistance scores associate significantly with clinical CALGB cytogenetic risk tiers in TCGA-LAML (ρ=0.1805,p=0.0181\rho = 0.1805, p = 0.0181; Poor vs. Favorable p=0.0311p = 0.0311), and zero-shot transfer to Bruna breast cancer PDTC ex vivo responses achieves positive within-drug ranking (Macro Spearman = 0.1485) with 1.2964x Top-20% sensitive enrichment (Appendix A.9, Table 16).

Best regards,

The Authors

AuthorsOpenReview ↗

Closing Summary of Discussion, Core Takeaways & Appreciation to Reviewers

Dear Area Chair and Reviewers,

As the discussion period comes to a close, we would like to sincerely thank the Area Chair and Reviewers kLTR, xG7y, and pxE5 for their time, thoughtful critiques, and constructive engagement. We are grateful that all three reviewers engaged with our rebuttal and positively updated their evaluations following our extensive new baseline, diagnostic, and external-cohort experiments. Below is a concise summary of what our study establishes and how each reviewer's feedback strengthened the work.

1. Core Contributions & Validated Empirical Takeaways:

  • Shared Pathway Ranking Backbone Over Pooled Shortcuts: Under strict patient-held-out evaluation, pooled correlation is easily inflated by between-drug potency shifts (a constant drug-mean reference reaches 0.6411 Global Spearman with zero within-drug ranking). Mapping both domains into a shared 1,000-Reactome coordinate space with a drug-balanced pairwise ranking objective enables strong unseen-patient stratification: Target-Only EviDR (0.3194) surpasses all DrEval BeatAML-only per-drug models (0.2433 to 0.2771), while Full EviDR (0.3291) leads all native continuous transfer baselines, beats runner-up TransDRP in all 5/5 outer folds (paired p = 0.0164), achieves the highest 3-seed mean and lowest variance (0.3214 +/- 0.0073), wins on 60.6% of drugs vs. MAC-CDR, and ranks 1st across NDCG@5/10/20%.
  • Validated Source Evidence Signal & Amplified Gains in Scarce-Data Regimes: Cross-fitted, lineage-deconfounded PRISM Evidence recovers 3.8x more reliable pathway associations than a 100x structured response-permutation null (p = 0.0099) and yields a 4.2x ranking gain in bypass-free probes. Under identical full supervision, EviDR beats an exact capacity-matched unstructured Concatenation MLP (766,988 vs. 766,989 parameters) in 4/5 folds while preserving pathway-level traceability. Crucially, regime stratification clarifies where Evidence matters most: while curated priors suffice on well-sampled drugs, EviDR roughly doubles its structural margin over the capacity-matched MLP on label-scarce (+0.0119) and sparse-prior (+0.0092) compounds and expands its lead over MAC-CDR (+0.0416).
  • Zero-Shot Clinical & Cross-Cancer Generalization: Without target retraining, frozen five-fold EviDR ensembles align significantly with pre-treatment CALGB cytogenetic risk tiers in clinical TCGA-LAML (N = 173; Poor vs. Favorable p = 0.0311; risk-tier rho = 0.1805, p = 0.0181) and transfer positively to non-AML Bruna breast cancer solid-tumor PDTCs (Macro Spearman = 0.1485, 1.2964x Top-20% sensitive enrichment).

2. Summary of Resolved Questions & Final Framing Refinements:

  • Reviewer kLTR: Expanded benchmarks with 2024 to 2026 methods (Drug2TME, DeepSADR, and AML hierarchy signature LinClass-7), added NDCG@5/10/20% rankings and paired/per-drug significance p-values, compared pre-treatment Evidence estimation with post-perturbation mining (DPADM), analyzed cross-dataset sign compatibility, and validated zero-shot generalization on TCGA-LAML and breast cancer PDTCs.
  • Reviewer xG7y: Isolated Evidence pathway alignment and directional sign via 100x permutation nulls and constrained readout probes, verified multi-seed stability across 15 runs and paired fold tests, completed 5-fold PRISM supervision scaling alongside an exact capacity-matched baseline, and streamlined the step-by-step method presentation and notation table.
  • Reviewer pxE5: Added DrEval BeatAML-only per-drug models and naive drug-mean references, reported learned Evidence scales and verified original-AUC perturbation impacts, and identified label-scarce and sparse-prior regimes as the primary locus of Evidence gains. Following Reviewer pxE5's helpful framing feedback, we will also ensure the final version centers overall ranking claims on the shared pathway backbone, qualifies binary-to-continuous adaptations (DeepSADR and Drug2TME) as scope-limited, and revises the main-text source-scaling claim to the 5-fold curve descriptively without post-hoc speculation.

Thank you once again to all three reviewers and the Area Chair for helping us make this manuscript much stronger, clearer, and more rigorous.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.