Scientists are calling for more explainable AI in protein language models, a technology with immense potential to address global challenges. These models can engineer proteins with novel structures, from carbon-absorbing enzymes to energy-efficient catalysts. However, a major issue arises: protein language models often operate as black boxes, making it difficult to understand their decision-making processes and assess their reliability and safety. This lack of transparency is a significant concern as these models increasingly influence biotechnology decisions.
In a recent perspective paper published in Nature Machine Intelligence, researchers from the Centre for Genomic Regulation (CRG) delve into the application of explainable AI in protein language models. They highlight the need for better understanding and interpretation of these models' decisions, especially as they become integral to real-world applications.
Dr. Noelia Ferruz, Group Leader at CRG and corresponding author, emphasizes the urgency of the situation: "Protein language models are advancing rapidly, but our understanding of fundamental biological processes has not kept pace. We risk building powerful tools that we cannot fully trust if we don't improve our ability to explain their decision-making processes."
The paper identifies four key areas for explaining protein language models' decisions: the training data, the specific protein sequence, the model's architecture, and input-output behavior. By examining these aspects, researchers can identify biases, understand the most influential amino acids, and even 'nudge' the model to observe its responses to changes.
The current use of explainable AI in protein research is primarily as an evaluator, checking if the model learns known patterns. However, the authors argue that explainability should be more than a verification tool. They envision a future where explainable AI acts as a teacher, revealing new biological principles and guiding the design of proteins with specific traits.
This 'teacher' role is akin to AlphaZero's discovery of novel chess strategies or AI's role in deciphering ancient texts. In protein science, it would mean uncovering new rules of protein folding, catalysis, or molecular interaction, revolutionizing medicine, materials, and sustainable technologies. Dr. Ferruz envisions a future where we can instruct a model to design a protein with specific characteristics and receive a clear explanation of why it works and why alternatives fail.
However, achieving this 'teacher' status is not automatic. The authors stress the need for robust benchmarks, open-source tools, and laboratory validation to ensure the reliability and accuracy of AI-derived insights. They call for a shift in the field's focus, from using explainability as a support tool to leveraging it as a driver of discovery, ultimately leading to more trustworthy and powerful protein language models.
