MolDeTox: Evaluating Language Model's Stepwise Fragment Editing for Molecular Detoxification

2026-05-12

AI Summary

Existing molecular optimization methods for toxicity mitigation suffer from limited data diversity, low validity of generated structures, and overreliance on surrogate toxicity predictors, while lacking fine-grained and interpretable evaluation mechanisms. To address these limitations, this work introduces MolDeTox—the first fine-grained benchmark specifically designed for molecular detoxification. MolDeTox constructs a diverse dataset grounded in real toxicity labels and incorporates fragment-level editing tasks to enhance both the plausibility and interpretability of generated molecules. Systematic evaluations using large language models (LLMs) and vision-language models (VLMs) across multiple settings demonstrate that fragment-level understanding substantially improves structural validity. MolDeTox thus provides a reliable and interpretable evaluation framework for toxicity-aware molecular optimization, with all associated data publicly released.

Abstract

Large Language Models (LLMs) and Vision Language Models (VLMs) have recently shown promising capabilities in various scientific domain. In particular, these advances have opened new opportunities in drug discovery, where the ability to understand and modify molecular structures is critical for optimizing drug properties such as efficacy and toxicity. However, existing models and benchmarks often overlook toxicity-related challenges, focusing primarily on general property optimization without adequately addressing safety concerns. In addition, even existing toxicity repair benchmarks suffer from limited data diversity, low structural validity of generated molecules, and heavy reliance on proxy models for toxicity assessment. To address these limitations, we propose MolDeTox, a novel benchmark for molecular detoxification, designed to enable fine-grained and reliable evaluation of toxicity-aware molecular optimization across stepwise tasks. We evaluate a wide range of general-purpose LLMs and VLMs under diverse settings, and demonstrate that understanding and generating molecules at the fragment-level improves structural validity and enhances the quality of generated molecules. Moreover, through detailed task-level performance analysis, MolDeTox provides an interpretable benchmark that enables a deeper understanding of the detoxification process. Our dataset is available at : https://huggingface.co/datasets/MolDeTox/MolDeTox

Problem

• molecular detoxification
• toxicity assessment
• language models
• molecular optimization
• benchmark

Innovation

• molecular detoxification
• fragment-level editing
• toxicity-aware optimization
• stepwise molecular generation
• language models for drug discovery