Scientists are calling for more explainable AI in protein language models, a technology with immense potential to address global challenges. These models, which help engineer proteins with novel properties, are currently operating as black boxes, making it difficult to understand their decision-making processes and assess their reliability and safety. In a new perspective paper published in Nature Machine Intelligence, researchers at the Centre for Genomic Regulation (CRG) explore the application of explainable AI techniques to protein language models, highlighting the need for greater transparency and trustworthiness in these powerful tools. The paper emphasizes the importance of explainability in the rapidly evolving field of protein language models, where breakthroughs in protein design are outpacing our understanding of fundamental biological processes. Without better explanations of how these models learn and make decisions, there is a risk of building tools that we cannot fully trust. The authors call for the research community to prioritize making protein-design systems more transparent, trustworthy, and secure, emphasizing that explainability should not be an afterthought but a fundamental aspect of the development process. The paper outlines four key areas for explaining a protein language model's decision-making: the training data, the specific protein sequence, the model's architecture, and input-output behavior. By examining these factors, researchers can gain insights into the model's biases, the influence of specific amino acids or protein regions, and the internal processing of information. The authors also introduce the concept of explainable AI playing different roles in protein research, such as Evaluator, Multitasker, Engineer, and Coach. However, the most ambitious and least realized role is that of the 'Teacher', where explainable AI can help uncover entirely new biological principles. This stage, akin to AI systems in other fields like chess or ancient text deciphering, would mean AI systems providing new insights into protein folding, catalysis, or molecular interaction, revolutionizing the design of medicines, materials, and sustainable technologies. Achieving the 'Teacher' status for protein language models requires robust benchmarks, open-source tooling, and laboratory validation to ensure that AI-derived insights are experimentally confirmed biological knowledge. This shift from pattern recognition to true understanding is crucial for the responsible and effective use of protein language models in addressing global challenges.
