acceptodds
Under review as a conference paper at ICLR 2027

Translating In Silico Toxicology: Benchmarking Large Language Models for Interpreting Machine Learning-Driven Drug Discovery

Abstract

Clinical trials of potential new medicines are exceptionally costly and often end in failure, with approximately 30% of these failures attributed to drug toxicity. In silico drug toxicity prediction has been an ever-growing field in recent years. However, computational toxicity models often lack interpretability for drug design specialists. In this paper, we construct a framework using Large Language Models (LLMs) to interpret machine learning toxicity predictions and feature importances to clearly communicate the outputs to drug design specialists. Using a PubChem derived dataset, we test four LLMs across multiple structured prompts for varying tasks. These outputs are evaluated using a rigorous three-tier framework; deterministic evaluation, error analysis, and LLM-as-a-judge. The evaluation of the framework identified Gemma3 as the best performing model according to deterministic metrics, while LLM-as-a-judge credited Deepseek-r1 with superior reasoning depth and structural verifiability. In addition, performance was seen to be robust to random variation. Crucially the error analysis identified zero occurrences of feature hallucinations or magnitude distortions, thus establishing the reliability of LLM-generated interpretations for drug discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Large Langauge Models, Evaluation Frameworks, Drug Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_JzEGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an LLM-based framework for interpreting machine learning (ML) toxicity predictions and feature importance results to support drug design specialists. Using a PubChem cytotoxicity assay dataset, the authors train 10 ML models and generate predictive performance metrics and feature importance scores. These results are then provided to four LLMs, which generate interpretations across six predefined tasks using three prompt variants. The generated responses are evaluated through a three-tier framework consisting of deterministic evaluation, claim and hallucination analysis, and LLM-as-a-judge evaluation. The authors analyze the faithfulness, clarity, and consistency of LLM-generated interpretations across different models, tasks, and prompting strategies.

Critical strengths

Well-motivated research

The paper addresses an important challenge in bridging the gap between ML-based toxicity predictions and their practical use by drug design specialists.

Critical weaknesses

W1. Unclear practical utility for drug design specialists

It is unclear how the proposed framework actually benefits drug design specialists. What specific decisions is the framework intended to support? Is its primary goal to help specialists select among different toxicity prediction models, or to interpret individual toxicity predictions?


W2. Limited evaluation across toxicity endpoints

The experiments are conducted on a single PubChem cytotoxicity assay dataset. Since the framework is intended for toxicity interpretation in drug discovery, the authors should evaluate it across multiple toxicity endpoints and investigate whether the generated explanations provide endpoint-specific insights.


W3. Lack of qualitative analysis

The paper lacks representative case studies demonstrating how the framework generates and interprets toxicity predictions. Providing examples of LLM-generated explanations, including both successful and failure cases, would help readers better understand the framework's behavior and assess the quality and practical usefulness of its outputs.


W4. Lack of human evaluation

Despite targeting drug design specialists, the paper does not include evaluations by domain experts. Although the authors mention human evaluation as future work, I believe it is essential for this study to demonstrate whether the generated explanations are actually useful for the intended users, rather than leaving it for future work.


W5. Missing overview figure

The paper lacks a main figure illustrating the proposed framework. An overview figure showing the complete pipeline, from ML toxicity predictions to LLM-generated explanations and their evaluation, would improve the clarity of the methodology.

Questions for authors

Please refer to the weaknesses listed above for my main questions and concerns.

Additionally, I could not access the GitHub repository linked in the Reproducibility Statement, and the supplementary materials were also unavailable. Therefore, I was unable to inspect the code or verify the implementation details. Could the authors provide accessible links to the code and supplementary materials?

Reviewer_7TpZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce an LLM-based framework for the interpretation of ML models' toxicity predictions in a way comprehensive for drug design specialists. Four LLMs are given ML models predictions as well as local and global feature importances and asked to interpret, compare and analyze them across 5 task types. Then LLMs' outputs are evaluated using three evaluation tiers. The authors conclude that LLMs are generally reliable for this task.

Critical strengths

The problem addressed in the study is practically important.

Critical weaknesses
  1. The overall presentation of the paper requires substantial improvement. The manuscript lacks a dedicated discussion section (which would be very interesting to see here) and a clear formulation of its main contributions, while the experimental and implementation details are insufficient to ensure reproducibility; moreover, although a code repository is mentioned, it seems to be unavailable. The paper also contains some typos and formatting issues, e.g. the reference to Figure 1 in Section 2.1.1 appears incorrect.
  2. Despite positioning the framework as a tool for improving the interpretability of toxicity predictions for drug discovery specialists, the paper provides no human expert evaluation. The authors outline this limitation in the paper, but it seems to be crucial for the central claim of the study.
  3. The authors could more clearly explain the motivation behind the current study design. In particular, the target users and their expected level of ML expertise remain ambiguous. If the framework is intended to support drug discovery specialists without ML expertise, it is unclear who is expected to train the predictive models and perform feature importance analysis beforehand. Conversely, if these steps are performed by ML specialists, the authors should clarify the envisioned workflow and explain how LLM-generated interpretations would facilitate communication between ML experts and domain specialists.
  4. The study uses a single cytotoxicity assay comprising 2,294 compounds and ten molecular descriptors. It remains unclear whether the observations generalize to other toxicity endpoints, datasets, molecular representations, or prediction settings.
  5. The results section is largely descriptive, primarily reporting quantitative metrics and statistical findings without providing sufficient interpretation or critical analysis. The authors could offer deeper insights into the observed trends, discuss possible explanations for differences across models and tasks, and elaborate on the implications of their findings. As a result, it is difficult to identify the key takeaways of the study and understand what broader conclusions can be drawn from the experiments.
Questions for authors
  1. Could the authors clarify the intended practical workflow of the proposed framework?
  2. Could the authors provide a more detailed interpretation of the experimental findings? In particular, what might explain the substantial differences in performance across tasks and models, and what practical conclusions should be drawn from these observations?
Reviewer_9DtsOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether LLMs can explain toxicity-prediction results to drug-discovery specialists. It builds classifiers for a PubChem THP-1 cytotoxicity dataset of 2,294 compounds, then gives model metrics and feature-attribution results to four local LLMs. Each LLM addresses six tasks using three prompt variants and five random seeds. Responses are assessed by rule-based checks, an automated claim/error audit, and a Llama 3.3 judge. Gemma 3 has the highest reported mean overall rule-based score (2.720/4). The audit reports seven logical contradictions and no detected feature hallucinations, directional inversions, or magnitude distortions. Local SHAP interpretation (Task D) is particularly difficult. The paper acknowledges that it does not test usefulness with domain experts (Sections 2–4).

Critical strengths
  1. The study compares four LLMs over six interpretation tasks, three prompt styles, and five seeds, rather than presenting only a few selected examples. This design reveals task-specific weaknesses, especially for local SHAP explanations (Table 1; Figures 2 and 4).
  2. The authors try to separate numerical and feature agreement from readability, and they examine claims in addition to aggregate scores (Sections 2.3–2.4). This is a useful starting framework for auditing generated scientific explanations.
Critical weaknesses
  1. The paper does not validate its evaluation measures against independent expert judgments. Keyword matching, numerical checks, and a lexicon-filtered claim audit can miss unsupported chemical or biological statements; sentence splitting alone does not ensure atomic claims. The LLM judge is also not calibrated against human labels. Therefore, “zero detected” errors in three categories cannot establish that the explanations are reliable for drug discovery. The gap is especially consequential because the study uses one cell-line assay and has no end-user test (Sections 2.1, 2.3–2.4, and 4).
  2. The analysis of missing and failed outputs is unclear. The design implies 4 × 6 × 3 × 5 = 360 responses, and the conclusion refers to 360 outputs, yet the paper says DeepSeek-R1 timed out in more than half of its Task D runs (Section 3.3). It does not explain whether those runs were retried, excluded, or scored as failures. This could change the model comparison and the reported error rate. More broadly, the abstract credits DeepSeek-R1 with superior “reasoning depth and structural verifiability,” although the stated judge rubric measures only faithfulness and clarity (Figure 1).
Questions for authors
  1. What are the numbers of attempted, completed, retried, and excluded runs for every model–task–prompt cell? How do the model comparisons and error rates change when timeouts are counted as failures?
  2. Can independent toxicology or XAI experts label a representative sample of responses and claims, then report the rule-based audit's error-detection recall and agreement with the LLM judge? Please include the size of the lexicon-filtered claim set and examples of errors the audit misses.
  3. What evidence supports the abstract's claim about reasoning depth and structural verifiability, and what evidence shows that these explanations help drug-discovery specialists make better decisions than the original metrics and attribution plots?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.