New Benchmark Uncovers Safety Risks in AI-Generated Molecules

The development of AI models for generating new molecules has primarily focused on efficacy and novelty, often overlooking potential safety hazards. A new benchmark, MolSafeEval, aims to address this critical gap by providing a systematic framework for identifying and explaining unsafe characteristics in AI-designed compounds. This tool moves beyond simple toxicity predictors by incorporating a broad range of safety data, from toxicological databases to hazard rules, into a comprehensive molecular safety knowledge graph. MolSafeEval leverages large language models to reason over this knowledge graph, enabling detailed detection and explanation of hazardous molecular features. The benchmark categorizes generative models into four types—unconditional generation, property optimization, target protein-based design, and text-based generation—and offers standardized datasets and evaluation protocols for each. By exposing the safety vulnerabilities of current AI approaches, MolSafeEval provides crucial guidance for developing more reliable and secure molecular design processes.