acceptodds
Under review as a conference paper at ICLR 2027

Generalizing Drug Response Prediction through Evidence-Grounded Reasoning

Abstract

Accurate drug response prediction for previously unseen drugs and cell lines is a central challenge in computational pharmacology. In this setting, a few measured responses provide useful context, but integrating these observations with molecular properties and cellular characteristics remains difficult. We propose a framework for generalizable drug response prediction through evidence-grounded reasoning. The framework uses a large language model to combine molecular and cellular representations with biological facts and a small set of observed responses. To connect these heterogeneous inputs to prediction, we construct chain-of-thought supervision from available drug and cell facts and support-response summaries, without using the query response in reasoning construction. A multi-stage training strategy first aligns biological inputs through response-free pretraining, then initializes evidence-grounded reasoning jointly with response prediction through a teacher-guided warmup. Subsequent numerical prediction updates alternate with reinforcement learning using REINFORCE, where the predictive loss relative to a reference reasoning trace guides the optimization of generated reasoning. A probability-weighted numerical token interface produces the response estimate. This links reasoning to its usefulness for response prediction. At inference, the framework uses the available biological evidence and measured responses to generate a reasoning trace and estimate the response of an unmeasured drug–cell pair. Across cell-blind and drug-blind GDSC2 evaluation with zero or five measurements, the framework records the lowest macro RMSE in the reported comparison, while contextual measurements improve both RMSE and PCC in each protocol.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Drug response prediction, Few-shot generalization, Biological representation learning, Evidence-grounded reasoning, Reinforcement learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_txiJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

P2R-LLM studies prediction of GDSC2 drug responses for held-out drugs or cell lines, with either zero or five measured responses from the unseen entity. It combines frozen molecular and expression features, textual biological facts, and optional support measurements in a Qwen3-14B language model. Training proceeds through response-free biological alignment, teacher-guided reasoning warmup, and alternating numerical prediction and REINFORCE updates. A 100-token response interface converts the reasoning-conditioned hidden state into a continuous LN IC50 estimate. The method achieves the lowest entity-macro RMSE among the reported baselines in all four evaluation settings, with further cell-blind component ablations.

Critical strengths
  1. A carefully specified entity-held-out evaluation. Drug-blind and cell-blind protocols are reported separately, with zero- and five-measurement settings, fixed support panels, explicit query exclusions, and train-fitted response bins (Sections 3.1, 3.3, and 4; Appendix A). These details make the claimed operating regime much easier to assess than a generic random drug–cell pair split.
  2. Competitive prediction results with relevant controls. Tables 1–2 report the best RMSE in all four comparisons, including 1.6109 versus 1.7116 in drug-blind zero-shot prediction. The model also improves from 1.6109 to 1.2632 when five same-drug responses are supplied. Table 3 retains a direct no-CoT predictor with the same biological and support inputs, which is particularly useful for judging what the trained reasoning pathway adds.
Critical weaknesses
  1. The evidence does not yet isolate the value of the reasoning content. In Table 3, the no-CoT model already reaches 1.2461 cell-blind RMSE at k = 0, versus 1.3261 for MLP; generated reasoning lowers this further to 1.2337. The authors correctly note that this ablation changes intermediate autoregressive computation as well as text content. Likewise, the w/o RL control removes prediction-error feedback together with format and length rewards, so it cannot isolate the contribution of outcome-based reasoning optimization. The experiments support the complete training pathway, but leave unclear whether the biological comparisons expressed in the trace are needed for the prediction improvements.
  2. The margins and comparison design leave the headline generalization advantage uncertain. In drug-blind five-shot prediction, the improvement over the strongest baseline is only 1.2753 to 1.2632 (0.95%; Table 2). Appendix B reports one training seed per protocol and a fixed support panel, without uncertainty across seeds, support draws, or held-out entities. The baselines legitimately use the same support measurements, but typically adapt via residual KRR and do not share the pretrained language backbone or textual facts (Section 4; Appendix C). These comparisons establish a promising full-system result, not a controlled advantage of in-context reasoning over simpler support-aware predictors. More evidence is needed before treating the small reported gaps as robust.
Questions for authors
  1. Are there existing evaluations in which generated traces are replaced, shuffled, or otherwise perturbed while the same model, biological inputs, and approximate inference computation are retained? If not, which result most directly supports the need for semantically meaningful intermediate text?
  2. Does the reported w/o RL result include removal of both the response-error term and the format/length terms in Eqs. (7)–(8)? Are any existing separate-component results available?
  3. For the drug-blind k = 5 comparison, can the authors report paired per-drug RMSE differences, or any existing uncertainty estimates across held-out drugs or support-panel selections? How many of the 24 test drugs improve over the strongest baseline?
  4. Is a same-input direct predictor with a simple support-based calibration available beyond the no-CoT control, and are there any drug-blind component ablations? These would help distinguish in-context adaptation from the effect of the pretrained representations and the broader training recipe.
6 comments

Thank you for asking us to clarify the role of the reasoning text and the reward. We report the trace substitutions, define the RL ablation, and give the paired drug results and the comparison with a direct predictor.

1. Trace replacement at a fixed checkpoint

We replaced the generated trace with a similarly long trace from a different drug on the same drug-blind development queries, keeping the model, biological inputs, and supports fixed.

Tracek=0k=0 Micro RMSEMacro RMSEk=5k=5 Micro RMSEMacro RMSE
Original1.70371.57781.21891.1443
Shuffled across drugs1.74861.61271.33221.2392

At k=5k=5, shuffling raises Micro RMSE by 0.1133 (95% paired interval [0.0197, 0.2330]). The original trace has lower error on 15 of 24 drugs, showing that the predictor responds to whether the trace fits the query.

In the separate drug-blind training ablation, No-CoT scores 1.8360 / 1.4839 Macro RMSE at k=0/5k=0/5, versus 1.6908 / 1.4315 for No-RL. The ablation changes the training pathway, while shuffling changes the text at fixed weights.

2. What the w/o RL ablation removes

Yes. The Table 3 w/o RL ablation removes the entire REINFORCE term, including prediction-error feedback, the completion bonus, and the length penalty. It retains teacher warmup, trace sampling, and numerical supervision on sampled traces.

RL accounts for roughly one tenth of the response-training updates up to the evaluated drug-blind checkpoint. Isolating the effects of individual reward terms at this share would require repeated runs. Time and compute precluded those runs during rebuttal. We will add a detailed sensitivity analysis of the loss weights and reward terms in the revised manuscript.

3. Per-drug differences and paired uncertainty

In the drug-blind table at k=5k=5, our Macro RMSE is 1.2632 versus 1.2753 for DeepCDR. Our model has lower error on 15 of 24 drugs than DeepCDR. The table gives the paired differences for all 24 drugs against DeepCDRat k=5k=5. A positive difference means our model has lower RMSE.

DrugQueriesDeepCDR RMSE minus P2R(Ours) RMSE
2,3,4,9-Tetrahydro-1H-pyrido(3,4-b)indol-1-one530+0.0766
Afatinib680−0.2168
BMS-345541 free base631+0.0082
Bromosporine530+0.0378
Dabrafenib672+0.0356
Dihydrorotenone529+0.3159
Erlotinib672+0.1111
Gsk-269962A165+0.0050
JNK Inhibitor VIII670+0.1200
N-((2R)-2,3-Dihydroxypropoxy)-3,4-difluoro-2-((2-fluoro-4-iodophenyl)amino)benzamide682−0.1795
NU-7441665−0.1911
Niraparib632−0.1356
Podophyllotoxin bromide530+0.0889
Rucaparib667−0.1596
Schweinfurthin A529−0.2489
Selumetinib681−0.0302
Sgc-cbp30530+0.1620
Sgc0946669+0.3688
Tamoxifen675−0.2349
Teniposide530−0.2459
Umi-77554+0.0836
Veliparib668+0.1902
Zorifertinib680+0.1312
alpha Lipoic acid552+0.1980

Similarly, compared with TopDr, our model shows an advantage on 16 out of 24 drugs, and compared with PASO, it shows an advantage on 18 out of 24 drugs. This indicates that although the numerical RMSE advantage of our method in the drug-blind setting is not substantial, our method maintains a consistent advantage across different drugs.

4. Direct prediction and drug-blind ablations

The full model has the lowest Macro RMSE at both support levels in the drug-blind ablations.

Variantk=0k=0 Macro RMSEk=5k=5 Macro RMSE
P2R(Ours)1.61091.2632
No-RL1.69081.4315
No-CoT1.83601.4839
No-BIO1.75221.4331

Compared with No-RL, CoT and biological alignment each lower error. CoT links evidence to the prediction, biological alignment helps the model read molecular and expression features, and RL guides trace generation using prediction error.

Methodk=5k=5 Macro RMSEk=5k=5 PCC
P2R(Ours)1.26320.6009
Direct predictor + simple calibration1.27070.5779

We used the baselines' residual kernel ridge regression to calibrate the direct predictor using five support measurements per drug.Simple calibration also substantially reduces Macro RMSE (from 1.6109 to 1.2707), but it remains inferior to our method (1.2632). Moreover, simple calibration yields only limited PCC improvement (0.5779), consistent with the baseline results in the main table, whereas our ICL-based approach improves PCC to 0.6009.

Thank you for the additional analyses. The trace-replacement experiment, drug-blind ablations, and calibrated direct-prediction comparison provide useful additional evidence. However, the trace experiment does not establish the biological validity of the generated reasoning, and the small RMSE advantage over the calibrated direct predictor remains uncertain without repeated seeds or support-panel resampling. The description of the No-RL ablation and RL update frequency also needs clarification. I therefore maintain my original assessment.

Thank you for the follow-up. We address the evidence for the content of the generated reasoning, the stability of the comparison with the calibrated direct predictor, and the distinction between the No-RL controls and the full training schedule.

1. What the reasoning contributes

We understand the concern about whether the generated reasoning contains biologically meaningful inferences. We found two zero-shot examples involving drugs held out of response training in which the model inferred downstream effects consistent with published experiments in the same cell lines. For BMS-345541 in SK-MEL-5, the trace linked IKK inhibition to reduced NF-κB signaling, consistent with the decrease in NF-κB activity measured by Yang et al. [1]. For UMI-77 in BxPC-3, the trace linked MCL1 inhibition to the release of pro-apoptotic proteins, consistent with the disruption of MCL1 binding to BAX and BAK reported by Abulwerdi et al. [2]. In both cases, the model used the supplied target annotation to generate a testable hypothesis about a downstream effect.

To test whether predictions respond to the direction of that biological relationship, we selected several zero-shot drug-blind queries (k=0k=0) and made two edits to each generated trace. One rewrites the mechanism in different words while preserving its meaning. The other uses the same rewritten sentence structure and reverses the direction of the effect. We held the checkpoint, query, biological inputs, and surrounding reasoning fixed. The two edits preserve the same factual observations and have equal token counts, so their comparison tests sensitivity to the direction of the stated mechanism.

BMS-345541 in SK-MEL-5

Original trace

BMS-345541 is an IKK inhibitor, and CHUK expression is 4.1 log2(TPM+1) in this melanoma. No measured supports are available; expression alone does not establish sensitivity. IKK inhibition could reduce NF-κB signaling, but the conditional expectation remains uncertain.

Meaning-preserving rewrite

BMS-345541 is an IKK inhibitor, and CHUK expression is 4.1 log2(TPM+1) in this melanoma. No measured supports are available; expression alone does not establish sensitivity. When IKK is blocked, the NF-κB pathway could become less active. The resulting sensitivity in this cell line remains uncertain.

Reversed mechanism

BMS-345541 is an IKK inhibitor, and CHUK expression is 4.1 log2(TPM+1) in this melanoma. No measured supports are available; expression alone does not establish sensitivity. When IKK is blocked, the NF-κB pathway could become more active. The resulting sensitivity in this cell line remains uncertain.

Reasoning textRMSE on selected k=0k=0 queries (log IC50) ↓
Original trace1.4607
Meaning-preserving rewrite1.4671
Reversed mechanism1.4812

We further extended this edit-based analysis to a total of eight samples: the two examples with explicit evidence support described above, plus six samples randomly selected from cases whose reasoning sentences contained clear mechanistic statements. For the meaning-preserving rewrites, RMSE decreased on 4 samples and increased on 4 samples, with the overall RMSE remaining essentially unchanged (only 0.0064 higher on average). In contrast, for the mechanism-reversed edits, RMSE increased on 6 samples and decreased on 2 samples, with an average increase of 0.0205. These results indicate that our reasoning is direction-sensitive and carries genuine inferential meaning, rather than merely superficial paraphrasing, and thus has reference value for real biological research.

References

[1] Yang J, Amiri KI, Burke JR, Schmid JA, Richmond A. BMS-345541 targets inhibitor of κB kinase and induces apoptosis in melanoma: involvement of nuclear factor κB and mitochondria pathways. Clin Cancer Res. 2006;12(3 Pt 1):950–960. DOI

[2] Abulwerdi F, Liao C, Liu M, et al. A novel small-molecule inhibitor of Mcl-1 blocks pancreatic cancer growth in vitro and in vivo. Mol Cancer Ther. 2014;13(3):565–575. DOI

Reviewer_r7q3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes P2R-LLM, a framework for generalizable drug response prediction that integrates molecular and cellular information, biological knowledge, and a small number of observed responses through LLM-based reasoning. The approach employs a multi-stage training strategy involving biological alignment, teacher-guided warmup, and prediction-guided reinforcement learning to connect generated reasoning with numerical predictions. Experiments on GDSC2 under cell-blind and drug-blind settings, with zero or five observed responses, report improvements in prediction accuracy over the evaluated baselines, particularly when contextual measurements are available.

Critical strengths

The paper tackles the challenging problem of predicting drug responses for previously unseen drugs and cell lines. A key strength is the integration of molecular and cellular information, biological knowledge, and limited experimental measurements within a unified LLM framework. The prediction-guided training strategy provides an interesting way to connect generated reasoning with quantitative outcomes. Furthermore, the evaluation covers both cell-blind and drug-blind settings, with zero-shot and few-shot scenarios, demonstrating the potential benefits of incorporating contextual measurements.

Critical weaknesses

The proposed framework involves multiple training stages and optimization mechanisms, making it difficult to assess the necessity and individual contributions of each component. Although the reported results are encouraging, the extent to which the generated reasoning reflects meaningful biological relationships remains unclear. The evaluation is also limited to a single dataset, raising questions about robustness and generalisation across different pharmacogenomic settings. Further comparisons with simpler approaches and a more detailed analysis of computational efficiency would strengthen the paper.

Questions for authors
  1. How much does each training stage contribute to the overall performance, particularly the reinforcement learning component?

  2. How do the authors validate that the generated reasoning captures meaningful biological relationships rather than merely improving numerical predictions?

  3. How well does the proposed framework generalise to independent drug response datasets, and how does its computational efficiency compare with simpler approaches?

1 comment

Thank you for asking about the training stages, the biological traces, and transfer across screens. We report the ablations, two biological case studies, results on CTRPv2 and gCSI, and a runtime comparison with a direct predictor.

1. Contribution of the training stages

The paper's cell-blind ablations separate the full model from the variants by at most 0.0124 Macro RMSE at k=0k=0 and 0.0246 at k=5k=5. At five supports, the largest change comes from removing the shaped policy update in Table 3, which increases Macro RMSE from 1.2106 to 1.2352 and lowers Macro PCC from 0.9110 to 0.8988. The corresponding increases in RMSE for No-CoT and No-BIO are 0.0204 and 0.0063.

In the drug-blind follow-up, the full configuration lowers Macro RMSE relative to No-RL by 0.0461 at zero supports and 0.1635 at five supports. This comparison also changes teacher supervision and initialization. At five supports, No-BIO differs from No-RL by 0.0016 Macro RMSE.

2. Biological interpretation of the generated reasoning

Generated traces offer testable hypotheses about drug action. Their biological accuracy requires experimental validation. We examined two zero-shot traces that our model generated for drugs held out of response training.

  • BMS-345541 in SK-MEL-5. The model inferred that IKK inhibition could reduce NF-κB signaling. Experiments in the same melanoma cell line found reduced NF-κB activity [1].
  • UMI-77 in BxPC-3. The model inferred that MCL1 inhibition could release pro-apoptotic proteins. Experiments in the same pancreatic cancer cell line found disrupted MCL1 binding to BAX and BAK [2].

Both cases show how the model can connect supplied target annotations to downstream effects that researchers can test.

3. Independent-screen transfer and computational cost

CTRPv2 and gCSI are independent drug screens from the Broad Institute and Genentech, respectively. We evaluated the trained GDSC2 drug-blind model on both screens without retraining or fine-tuning. All evaluated cell lines occur in GDSC2 training. The drugs were held out of response training but appeared in GDSC2 checkpoint selection. We score log IC50 from each screen for curves that cross 50% viability in range. CTRPv2 contributes 1,710 queries across seven drugs and gCSI 256 across three. Of these, 1,665 and 251 pairs also have GDSC2 measurements, respectively. At k=5k=5, all methods use the same five support measurements per drug from the target screen.

MethodCTRPv2 k=0k=0 Micro RMSECTRPv2 k=5k=5gCSI k=0k=0 Micro RMSEgCSI k=5k=5
P2R(Ours)2.03751.64963.31151.9804
TopDr2.20741.83003.44072.2919
GCNPath2.47201.69243.82022.4513
DiSPA2.18301.83663.41482.1286

We compared P2R with TopDr, GCNPath, and DiSPA on these two external datasets, CTRPv2 and gCSI, under both the k=0k=0 and k=5k=5 settings. As shown in Table, our method achieves the lowest Micro RMSE in all four settings and consistently outperforms every baseline. On CTRPv2, P2R reduces the Micro RMSE to 2.0375 at k=0k=0 and 1.6496 at k=5k=5, corresponding to 6.67% and 2.53% improvements over the strongest baseline in each setting. On gCSI, P2R attains 3.3115 at k=0k=0 and 1.9804 at k=5k=5, yielding 3.03% and 6.96% reductions relative to the strongest baselines. These results show that the advantage of P2R is not limited to the original experimental setting but generalizes robustly to new external datasets and both evaluation settings. Even in the setting with the smallest margin (CTRPv2, k=5k=5), P2R still ranks first, further supporting the effectiveness and generalization of our approach.

References

[1] Yang J, Amiri KI, Burke JR, Schmid JA, Richmond A. BMS-345541 targets inhibitor of κB kinase and induces apoptosis in melanoma: involvement of nuclear factor κB and mitochondria pathways. Clin Cancer Res. 2006;12(3 Pt 1):950–960. DOI

[2] Abulwerdi F, Liao C, Liu M, et al. A novel small-molecule inhibitor of Mcl-1 blocks pancreatic cancer growth in vitro and in vivo. Mol Cancer Ther. 2014;13(3):565–575. DOI

Reviewer_NG7XOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes P2R-LLM, a framework for predicting drug responses in unseen drugs or cell lines by combining molecular embeddings, gene-expression features, biological facts, and optional measured responses within an LLM. Its three-stage training procedure consists of response-free biological input alignment, teacher-supervised reasoning and prediction warmup, and alternating numerical prediction and REINFORCE updates. The reinforcement learning signal rewards generated reasoning that reduces prediction error relative to a deterministic evidence summary. A numerical output head estimates continuous responses through a probability-weighted average over 100 response bins. Evaluation on GDSC2 covers drug-blind and cell-blind settings with zero or five support measurements. The method achieves the lowest reported macro RMSE across all four settings, while support measurements improve both RMSE and correlation without inference-time parameter updates. Cell-blind ablations assess the contributions of biological alignment, explicit reasoning, and reinforcement learning.

Critical strengths
  1. The method integrates molecular representations, gene-expression features, biological facts, and measured responses within a shared LLM. Its staged training procedure connects reasoning generation to downstream response prediction.
  2. The paper evaluates both unseen drugs and unseen cell lines, with either zero or five support measurements. This tests prediction from biological information alone and adaptation when limited experimental evidence becomes available. The reported improvements in both RMSE and correlation with support measurements, without updating model parameters, demonstrate the practical value of incorporating new observations through context.
Critical weaknesses
  1. Although the paper describes feature projectors and a numerical readout, the overview figures do not clearly show how the projected biological features enter the LLM or how the reasoning-conditioned hidden state produces the response prediction. The transitions between the “aligned model” and “initialized predictor” can also make it unclear whether these are separate components or successive states of the same model. A more explicit computational diagram showing the input embeddings, shared LLM, response head, and trainable parameters at each stage would make the method easier to understand and reproduce.

  2. Equations 6–8 introduce several fixed coefficients. The paper specifies their values but does not explain how they were selected or assess sensitivity to these choices. Because these coefficients control the balance between numerical accuracy, reasoning imitation, and generation behavior, their influence on the reported gains remains unclear. A description of the tuning procedure and a focused sensitivity analysis would strengthen confidence in the training recipe.

Questions for authors
  1. Could you clarify the end-to-end architecture, particularly how the projected drug and cell features enter the LLM and how the generated reasoning connects to the numerical prediction head?
  2. Could you clarify how the weight parameters are selected in loss functions?
  3. Could you include comparisons with stronger general-purpose LLM agents, such as Opus or Sol, equipped with tools to analyze the same biological features and support measurements? This would help assess whether the proposed feature integration and training framework offers advantages over a strong tool-using agent
5 comments

Thank you for identifying where the architecture and loss choices need more explanation and for suggesting a comparison with tool-enabled language models. We describe the computation from projected features to the numerical answer, clarify which parameters the three stages share, and evaluate our model, Sol, and DeepSeek Flash on the same drug-blind queries.

1. End-to-end computation and parameter sharing

The model follows one sequence from input to numerical prediction. Separate projectors turn the drug and cell vectors into soft tokens at designated prompt positions. The text embeddings describe the entities and the available biological facts. Each support block pairs a measured response with the features and facts of its counterpart. In drug-blind prediction, for example, the sequence starts with the shared drug, then includes the measured support cells and the query cell.

The Qwen3-14B-Base backbone with low-rank adaptation (LoRA) processes the combined sequence and generates the trace. After the trace closes, the controller appends <RESP>. The hidden state at that position can attend to the projected features, the biological facts, the support observations, and the trace tokens. Multiplication by the 100 learned response output rows gives the bin logits. A softmax over those logits followed by a weighted average of the training bin representatives and rescaling gives the prediction in log IC50 units. Sections 3.2–3.3 give the projection and readout equations.

“Aligned model” and “initialized predictor” name successive parameter states of that shared architecture. During biological alignment, the first phase updates LoRA and the joint phase updates LoRA and the projectors. The response-token rows are trained in Stages II and III alongside those components. The base backbone weights stay frozen throughout. Appendix A gives the stage-specific settings.

We have revised Figure 3 to show how the projected drug and cell inputs enter the shared LLM and how the model generates a reasoning trace before the O100 head outputs a continuous response prediction. The caption also specifies which parameters are updated at each training stage.

2. Choice of loss weights

We set the coefficients based on the scale and role of each term, keeping squared error as the primary objective. We did not use grid search or tune the coefficients across repeated runs. The bin cross-entropy weight (0.1) guides the distribution toward the target bin. The teacher imitation weight (0.25) teaches trace structure during warmup, and the REINFORCE weight (0.05) scales the update based on prediction error relative to the reference. The completion bonus (0.02) and length penalty (up to 0.05) encourage complete, concise traces.

Time and compute limited additional training during the rebuttal. We will add a focused sensitivity study of these coefficients in the revised manuscript.

3. Comparison with general-purpose agents

Our model has the lowest Micro and Macro RMSE at both support levels in the 50-query comparison. At k=0k=0, its Micro and Macro RMSE are 11.8% and 14.0% lower than those of tool-enabled DeepSeek Flash. At k=5k=5, the reductions are 3.8% and 8.2% relative to tool-enabled Sol. Our model beats the strongest comparator on 14 of 24 drugs at each support level.

Methodk=0k=0 Micro RMSEMacro RMSEPooled PCCk=5k=5 Micro RMSEMacro RMSEPooled PCC
GPT-6.1 Sol (xhigh), text2.40222.06250.17481.52931.32900.7783
GPT-6.1 Sol (xhigh), feature tool2.43672.09970.13091.51611.29450.7846
Gemini 3.8 Flash (high), text2.40932.11670.22401.53551.34390.7763
Gemini 3.8 Flash (high), feature tool2.28062.00590.38471.53931.31120.7765
P2R(Ours)2.01101.72540.56871.45861.18780.8087

The comparison uses the same 50 queries from 24 drugs and the same available supports for each method. We give each agent one query in a fresh context and use the reasoning instructions from our model's prompt. The text setting supplies the biological facts and the available measurements. The feature-tool setting also lets the agent inspect the frozen molecular and expression vectors for that query and its supports. We map each agent's integer bin to its training representative.

Micro RMSE and pooled Pearson correlation (PCC) pool all 50 queries. Macro RMSE gives each of the 24 drugs equal weight. The two error summaries show that the advantage on this subset is present both across pooled queries and when each drug contributes equally.

Thank you for the detailed response and additional experiments. Clarifying that the training stages represent successive states of the same architecture is particularly helpful. Please also clarify the discrepancy between “DeepSeek Flash” in the text and “Gemini 3.8 Flash” in the table.

Regarding the loss coefficients, I appreciate the transparent explanation that they were chosen manually according to the scale and role of each term. This clarifies their selection, but their sensitivity remains unresolved. A focused analysis, particularly of the teacher-imitation and REINFORCE weights, would help establish whether the reported gains are robust to these choices. But given the time window for the discussion period I understand it might be hard to conduct comprehensive analysis. I therefore consider this concern partially addressed.

I will reconsider my score in light of further clarifications and any additional sensitivity results (if possible).

1. Clarification about general-purpose agents compare results

We sincerely apologize for the confusion caused by our typographical error. In fact, we initially selected Sol and DeepSeek as the stronger general-purpose LLM agent baselines. However, relative to the Sol/Opus baselines recommended by the reviewer, we were concerned that this substitution might inadvertently cause misunderstanding regarding the double-blind review policy, and we therefore switched to Gemini. We further apologize that, during our revision, we overlooked updating the model name referenced in the body of the response. The DeepSeek results are provided below.

Methodk=0k=0 Micro RMSEMacro RMSEPooled PCCk=5k=5 Micro RMSEMacro RMSEPooled PCC
DeepSeek Flash (max), text2.47402.20050.30681.54691.33200.7436
DeepSeek Flash (max), feature tool2.21972.01140.36511.53171.30880.7651
P2R(Ours)2.01101.72540.56871.45861.18780.8087

2. Follow-up on loss coefficients

A full comparison at the submitted scale is not possible within the discussion period (nine full retrainings). Within that limit we ran a small-scale controlled probe, read qualitatively, to test two questions: is each coefficient necessary, and does the recipe depend on its exact value? Each sweep starts from a fixed checkpoint, changes one coefficient at a time, and holds the data order, batch size, prompt, evaluation protocol, and learning-rate schedule fixed. The τ runs branch from the biological-alignment checkpoint (200 teacher-supervised warmup updates); the β runs branch from the submitted warmup checkpoint (τ=0.25, 200 post updates). Each run is evaluated at k=0 and k=5 on the 400 development queries used for checkpoint selection (24 drug units, so micro RMSE; mean is the k=0/k=5 average used for selection).

βk=0k=5mean
0.01.79721.40281.6000
0.0251.81141.33221.5718
0.05 (submitted)1.87261.32101.5968
0.11.86891.35681.6128

Zeroing the actor term does not merely cost accuracy; it changes the reasoning the model produces: at β=0, 4.3%/2.9% of generated 4-grams are repeats and 4.5%/2.5% of traces hit the length cap, against 0.1%/0.5% and 0%/0% at β=0.05. Without the prediction-guided update the trace degrades into repetition and truncation, so the coefficient cannot be set to zero.

Above zero the recipe is insensitive to the exact value: the mean stays within 2.6% across 0–0.1, and the only systematic direction is the intended one — with five measured responses the trace must decide how far to rely on them, so β=0→0.05 improves k=5 (1.4028→1.3210). The support-free column gives back 4.2%, the cost side of the same trade-off (k=0 is where relying on responses cannot pay off); we read k=5 as the column measuring the targeted mechanism and do not treat the k=0 ordering as converged. The sweep does not single out our value as optimal (β=0.025 is 1.6% lower on the mean), consistent with a coefficient fixed a priori rather than tuned.

τk=0k=5mean
0.01.92831.41331.6708
0.1251.82601.41441.6202
0.25 (submitted)1.87261.32101.5968
0.51.87191.39191.6319
1.01.69431.29081.4925

Zeroing the teacher term is the only clearly worse setting: highest mean RMSE (1.6708) and highest k=0, with every non-zero setting 2.3–10.7% lower. Without teacher supervision the model does not learn to produce a usable reasoning trace, which is why the term is needed in warmup.

Among the non-zero settings the response is flat — all four within 7% of the submitted value (1.4925–1.6319) — and their ordering is not monotone in τ, so we do not read it as a ranking: the probe observes only 200 of the 3,000 post updates and later B-phase updates still use the submitted weight. The largest weight gives the lowest probe mean (τ=1.0, 1.4925, 6.5% below ours); we report it rather than recommend it, since the probe stops at 1/15 of the post schedule. The submitted 0.25 was fixed before these runs, and the sweep does not single it out as the optimum.

Taken together, both coefficients are necessary — zeroing either visibly degrades the result — and neither is fragile above zero, where the mean moves by a few percent. Each setting is a single run, so we read the sweeps as bounds on each coefficient's influence, not as fine-grained rankings.

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

This work proposes P2R-LLM, a biological alignment + SFT + RL framework designed to generalize drug response prediction (DRP) to unseen drugs and cell lines across both zero-shot and few-shot in-context settings without requiring test-time training. To address the key challenges in DRP, heterogeneous evidence integration,the proposed method first discretizes continuous drug responses into binned special tokens (O100) to integrate them into the LLM vocabulary. Next, it projects biological representations (molecular and cellular embeddings) into soft tokens and conducts biological alignment training to capture non-textual information. Finally, the model learns evidence-grounded reasoning via SFT warmup on high-quality teacher traces, followed by an alternating prediction-and-RL optimization stage to generalize and scale its predictive capability.

Experiments on Qwen3-14B-Base on the GDSC2 benchmark under cell-blind and drug-blind protocols (k=0 and k=5) show that while unguided sampling without RL struggles to integrate few-shot context effectively, P2R-LLM improves both entity-macro RMSE and within-entity Pearson correlation over existing deep learning, pathway-based baselines.

Critical strengths

Zero-to-Few-Shot Generalization. The model can integrate empirical samples via in-context learning (ICL)—without requiring test-time training, improving prediction accuracy for unseen drugs and cell lines.

Unified Heterogeneous Evidence Integration. A comprehensive approach that employs trainable soft token projectors to fuse diverse data types—continuous molecular graph, gene expression vectors, textual facts, and few-shot assay observations—into a unified input interface for LLMs.

reasonable RL design(stage III). Building upon the warmup phase, the alternating prediction-and-RL approach anchors the advantage in a reference-conditioned prediction, the reliability of the reward signal is thus enhanced.

Transparency. Rigorous cell-blind and drug-blind evaluations were conducted on the GDSC2, with hyperparameters and computational overheads detailed in the appendix.

Critical weaknesses

W1. The clarity of the writing could be improved. In many places, an excessive accumulation of technical terms makes the text difficult to read. Sec 3. reads more like a technical report on the model’s training recipe rather than a clear presentation aligned with the underlying motivation and the specific problem being addressed. This may hinder readers from intuitively grasping the rationale and purpose behind the design of each module.

W2. The selection of hyperparameters lacks justification. The proposed method involves several hyperparameters, yet for most of them, there is a lack of a justification for their selection. For example, the dimension of the output vector, in Sec 3.2, which is set to 4×h4 \times h; coefficients of terms in loss functions, eq. (5), (6), (7). This limits the reliability of the results in the experimental section to some extent. While I understand that constraints on computational resources may be the primary bottleneck, conducting a sensitivity analysis for a select subset of the most critical hyperparameters would help clarify the scope.

W3. Insufficient justification of module design.

  • The O100 interface is a key module of the proposed method; however, the necessity of this design lacks empirical support. For instance: (1) Could a coarser binning strategy achieve similar results? (Consider O100 vs. O25.) (2) Is the binning design itself necessary? (Consider the O100 interface versus a direct scalar linear layer—i.e., a "scalar head.")

  • The benefit of optimizing the trace rr as a reasoning process remains unclear; consider comparing r against directly using r0r_0.

Questions for authors

See weakness

In addition, could the author provide a more detailed description of the probability distribution for O100? For instance, do the shapes of the distributions vary significantly across different samples? Is the probability distribution smooth and unimodal, or does it take another form? Furthermore, are there any differences in the probability distribution between the cases where k=0 and k=5?

3 comments

Thank you for asking us to clarify the motivation for each stage, the projector and loss settings, and the behavior of the response readout.

1. Motivation for each training stage

We use a pretrained LLM to integrate biological evidence and generalize to unseen drugs and cell lines. When a full response profile is impractical to obtain, in-context learning lets the model use a few measurements without updating its weights.

Stage I, biological alignment. Soft tokens bring molecular and gene expression features into the LLM alongside text. Biological property tasks teach the model to interpret these features before it learns to predict drug response.

Stage II, teacher-guided warmup. We use chain-of-thought to help the model weigh the evidence and make a more accurate prediction. The trace also gives researchers a hypothesis to test. Few drug–cell pairs have enough documented mechanistic evidence for a well-grounded trace. Teacher examples show the model how to draw a conditional conclusion and predict from it.

Stage III, prediction-guided reasoning. Undocumented mechanisms can still shape responses. The model generates a hypothesis for each query using the available evidence and any support measurements. A fixed evidence summary gives the predictor a consistent input for direct supervision. We alternate this with REINFORCE updates that favor generated traces yielding lower error than the summary on the same query. This trains generation from measured responses even when teacher traces are unavailable.

We have restructured Methods around these design goals and put detailed training settings in the appendix.

2. Why four soft tokens per entity

KPGT supplies a 2,304-dimensional drug embedding, and DepMap supplies a 978-gene cell expression profile. Separate projectors map each vector to four soft tokens of width hh, the language model's embedding width. Four positions let each projector distribute information across several tokens while keeping the prompt compact as each support adds a cell or drug profile.

3. How we set the loss coefficients

We set the weights directly, with squared error as the main response objective. The 0.1 weight on bin cross-entropy directs probability mass toward the target bin. The 0.25 weight on teacher imitation teaches trace structure during warmup, and the 0.05 weight on REINFORCE adds prediction-guided policy learning alongside numerical supervision.

The 0.02 completion bonus encourages a complete trace, and the length penalty rises to 0.05 near the generation cap. Limited compute and time prevented us from running the multiple training configurations during rebuttal. We will add a focused sensitivity analysis of the loss weights in the revised manuscript.

3.1 O100 versus O25 and scalar regression

The table reports drug-blind Macro RMSE for all three readouts.

Readoutk=0k=0 Macro RMSEk=5k=5 Macro RMSE
O100 (submitted Table 2)1.61091.2632
O251.75341.2987
Scalar1.81671.3759

O25 groups more response values into each bin, so its training targets distinguish nearby response levels less precisely. Each O100 response token denotes a narrower interval on the response scale. The model sees support measurements in this vocabulary and uses its next-token probabilities to express uncertainty over the same response levels for the query. Several plausible bins can contribute to the continuous estimate. Scalar regression directly predicts a single value, bypassing this token-based formulation for the answer. The finer targets and shared vocabulary may help explain the lower error with O100.

3.2 Shape of the O100 distribution

At both k=0k=0 and k=5k=5, individual O100 distributions typically have one broad region of high probability across neighboring bins. The location and width of this region vary by query, and probabilities fluctuate from bin to bin. With five support measurements, the probability mass within one log IC50 unit of the measured response increases for 17 of 24 drugs.

3.3 Generated trace versus the fixed reference

Reasoning Tracek=0k=0 Macro RMSEk=5k=5 Macro RMSE
Model Generated1.61091.2632
Default r0r_01.69731.2880

We compared model-generated reasoning with default reasoning in the drug-blind setting. Model-generated reasoning achieved substantial improvements over the default reasoning. The reason is that the default reasoning statements merely repeat and state the features in a simple manner, without any analysis or inference. Therefore, they are not only less effective, but also provide no guidance or insight.

Thank you for providing further explanations and evidence. I am willing to raise my score to 3, primarily because Strengths 2 and 3, and find the rebuttal to "Weakness 3" convincing. I must emphasize, however, that there is still significant room for improvement in the writing of this paper; I suggest that the authors carefully revise the methodology section for the camera-ready version or any subsequent revision.

Thank you for taking the time to reconsider our submission and for your feedback on the Methods section. We will futher revise the section to explain the rationale for each training stage and make it clearer how the stages fit together.

Reviewer_BkjjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce a reinforcement learning approach to enable reasoning on retrieved biological evidence. The model is evaluated on one dataset (GDSC2) and compared against simple baselines as well as models from 2020, 2022, 2025, and 2026. There are some ablation analyses to quantify the contribution of each component to the performance.

Critical strengths

The authors evaluate against 8 baselines and run ablation analyses to understand the contributions of each component to performance gains.

Critical weaknesses

There is no standard deviation on multiple seeded runs to understand the variability of performance and significance of performance gains. Performance gains in Table 1-2 seem limited for cell- and drug-blind prediction, especially compared to very simple baselines such as MLP and XGBoost.

Ablation analyses (Table 3) show that several components lead to minimal/no performance gains. Again, there is no standard deviation to understand whether improvements in performance are significant.

It is unclear what the novelty and impact of this work is.

Questions for authors

Please clarify the above weaknesses.

1 comment

Thank you for asking us to clarify the size and reliability of the gains, the contribution of individual components, and what is new in the method. We start with the strongest baselines in the paper, then use the additional drug-blind ablations to compare the effects of changing different parts of the training procedure. We also explain why we combine contextual measurements, biological evidence, and prediction-guided reasoning in one model.

1. Variability across training seeds

Our current compute resources are insufficient to complete the additional 14B training runs needed for a multi-seed evaluation within the rebuttal period. The submitted tables therefore report results from one training seed per protocol. We used the available compute for drug-blind ablations because the paper already reports cell-blind ablations and the reviews ask how the components perform when the drug is held out. As a follow-up, we plan to evaluate the main results with three training seeds per protocol and report the mean and standard deviation of macro root mean squared error (RMSE) and Pearson correlation at zero and five supports. We will keep the data splits and support panels fixed to measure variation across training runs.

2. Magnitude of the main results

Our method achieves the lowest Macro RMSE at both support levels in Tables 1–2, with larger margins at k=0k=0. MLP denotes the multilayer perceptron.

ProtocolSupportsP2R(Ours) Macro RMSEBest baseline Macro RMSEReduction
Cell-blind01.23371.3261 (MLP)6.97%
Cell-blind51.21061.2471 (MLP)2.93%
Drug-blind01.61091.7116 (PASO)5.88%
Drug-blind51.26321.2753 (DeepCDR)0.95%

At drug-blind k=0k=0, MLP and XGBoost give 1.7520 and 1.9298. With five supports, our model's Macro RMSE falls by 1.87% for cells and 21.58% for drugs.

Our model's Macro Pearson correlation (PCC) also improves with five supports in both protocols. The strongest PCC baselines decline.

ProtocolMethodk=0k=0 Macro PCCk=5k=5 Macro PCCChange
Cell-blindP2R(Ours)0.90110.9110+0.0099
Cell-blindMLP0.89770.8973−0.0004
Drug-blindP2R(Ours)0.57760.6009+0.0233
Drug-blindTopDr0.58060.5755−0.0051

At the reported precision, PCC declines for seven of eight cell-blind baselines and five of eight drug-blind baselines. Macro PCC measures how well predictions track response variation within each held-out entity.

3. Follow-up drug-blind ablations

The full model has the lowest Macro RMSE at both support levels in the follow-up drug-blind development evaluation.

Variantk=0k=0 Macro RMSEk=5k=5 Macro RMSE
P2R(Ours)1.61091.2632
No-RL1.69081.4315
No-CoT1.83601.4839
No-BIO1.75221.4331

Against No-RL, chain-of-thought and biological alignment each reduce error. Biological alignment teaches the model to read molecular and expression features, chain-of-thought connects biological evidence with any available measurements, and reinforcement learning guides reasoning using prediction error.

4. Motivation and novelty

Drug response prediction is hard to generalize because a new compound or cell line can behave differently from those seen during training. We use a pretrained large language model to combine molecular features, expression profiles, and biological facts and apply what it has learned to unseen drugs and cells.

Zero-shot prediction is especially difficult. In drug development and precision medicine research, a few response measurements may be available even when a full profile is impractical to obtain. With in-context learning (ICL), the model uses those measurements to improve few-shot predictions while keeping its weights fixed.

We use known biological mechanisms, such as annotated drug–target relationships, to guide chain-of-thought (CoT) training. Grounding the reasoning in relevant evidence is intended to improve prediction accuracy and reliability. The traces also show which evidence informed the prediction and can suggest hypotheses for follow-up assays. Our training contribution is to alternate direct response supervision on a structured evidence summary with REINFORCE updates that reward generated traces for reducing prediction error relative to that summary.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.