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

The preprint operates within the dominant paradigm of applying machine learning, specifically Multimodal Large Language Models (MLLMs), to address challenges in drug discovery and development. It adheres to the established paradigm of using computational methods for molecular design and property prediction.

This preprint presents a novel benchmark and dataset for evaluating the ability of MLLMs to perform molecular toxicity repair. While this task is related to existing paradigms of toxicity prediction and molecular editing, framing it as "toxicity repair" introduces a nuanced focus within the existing paradigm. The preprint does not challenge the current paradigms of drug discovery or machine learning. It contributes to the ongoing "normal science" by refining methods and developing specialized benchmarks within the existing framework, aiming to improve the application of MLLMs in drug discovery. The introduction of a new benchmark dataset, ToxiMol, and evaluation framework, ToxiEval, facilitates further research and development within the existing paradigm, characterizing a "normal science" contribution that seeks to puzzle-solve and improve current techniques without overturning existing theoretical frameworks or methodologies. Because the success rate of existing MLLMs on this task is still low, and the benchmark is a novel contribution within an existing field, I could also see a case made for “pre-paradigm” science. However, the methods, framing, and evaluation all conform to accepted paradigms in computational drug discovery and MLLM application, with the benchmark itself a product of “normal science” activity. Thus, “normal science” is the most appropriate classification.