Beyond accuracy: Why molecular AI needs comprehensive alignment
AI for SciencePeer-Reviewed Publication
Artificial intelligence (AI) is transforming biomolecular and material sciences, enabling rapid predictions and generative design of proteins, drugs, and materials. However, current AI models often fail to adhere to fundamental physical laws, scientific objectives, and safety principles, leading to impractical or unsafe outcomes. A research team led by Qiang Zhang at Zhejiang University and collaborators from Tongji University, Shanghai AI Laboratory, Duke University, the National University of Singapore, The Chinese University of Hong Kong, Mingdu Tech, and University College London, has proposed a comprehensive alignment framework. Their perspective argues that AI systems for biomolecular and materials design should be aligned not only with data distributions or benchmark performance, but also with natural laws, scientific goals, and responsible research principles. By analyzing examples across protein engineering, drug discovery, and materials science, the authors show that many AI-generated candidates fail not because models are useless, but because the objectives being optimized are often disconnected from the physical, functional, and regulatory realities of the laboratory and the real world. This approach aims to make AI a trustworthy and effective partner in scientific discovery.