Breaking Bad Molecules: Are MLLMs Ready for Structure-Level Molecular Detoxification?

Researchers introduce ToxiMol, the first benchmark dataset and evaluation framework for assessing Multimodal Large Language Models (MLLMs) on molecular toxicity repair—the task of generating structurally valid alternatives to toxic compounds. Testing 43 mainstream MLLMs reveals current models show promise in toxicity understanding and constraint adherence but face significant challenges in this specialized pharmaceutical application.

The pharmaceutical industry loses billions annually to drug candidates failing due to toxicity in late-stage development, making early-stage toxicity mitigation critical. This research addresses a previously unmeasured capability in general-purpose AI models: their ability to understand and redesign molecular structures to eliminate harmful properties while maintaining drug efficacy. The ToxiMol benchmark represents a systematic attempt to formalize what has been an ad-hoc process in medicinal chemistry, using expert toxicological knowledge to create 660 diverse test cases across 11 task categories.

The introduction of ToxiEval, an automated evaluation framework integrating toxicity prediction, synthetic accessibility, drug-likeness, and structural similarity metrics, establishes quantifiable standards for molecular repair success. This addresses a gap in AI benchmarking where pharmaceutical applications have traditionally relied on specialized domain models rather than general-purpose systems. The assessment of 43 MLLMs provides valuable baseline data showing these models are beginning to grasp molecular toxicity concepts, though they remain far from production-ready for this task.

For the pharmaceutical and biotech sectors, this research indicates that general-purpose AI systems could eventually augment or accelerate early-stage drug design workflows, potentially reducing development timelines and costs. The benchmark enables continued evaluation of newer MLLM iterations as model capabilities improve. The open-ended nature of the dataset means it will likely become a standard evaluation tool in AI research, driving competition and innovation around molecular understanding in large language models over coming years.

• →ToxiMol establishes the first standardized benchmark for evaluating MLLMs on molecular toxicity repair tasks with 660 diverse toxic compounds.
• →Current general-purpose MLLMs demonstrate early promise in toxicity understanding but face significant challenges in structure-level molecular design.
• →ToxiEval framework provides automated, quantitative evaluation metrics combining toxicity prediction, drug-likeness, and synthetic feasibility assessments.
• →Testing of 43 mainstream MLLMs reveals these models can adhere to semantic constraints but struggle with complex molecular editing requirements.
• →The benchmark addresses a critical pharmaceutical bottleneck where toxicity causes early-stage drug development failures worth billions annually.