InterPLM: discovering interpretable features in protein language models via sparse autoencoders.

Area of Science:

• Computational biology
• Machine learning in bioinformatics
• Protein informatics

Background:

• Protein language models (PLMs) excel at protein modeling and design but their internal workings remain unclear.
• Understanding PLM mechanisms is crucial for advancing protein science and engineering.
• Interpreting complex models like PLMs is a significant challenge in bioinformatics.

Purpose of the Study:

• To develop a systematic framework for extracting and analyzing interpretable features from PLMs.
• To investigate how biological concepts are represented within PLM embeddings.
• To demonstrate the practical applications of interpretable features in protein annotation and generation.

Main Methods:

• Utilized sparse autoencoders to analyze embeddings from the ESM-2 protein language model.
• Trained sparse autoencoders to identify and extract interpretable features.
• Employed large language models for automated feature description and validation.
• Investigated feature representation across different model scales.

Main Results:

• Identified thousands of interpretable features corresponding to biological concepts (e.g., binding sites, motifs, domains).
• Discovered that concepts are stored in superposition across neurons within PLMs.
• Observed that larger PLMs capture a greater number of interpretable concepts.
• Found that ESM-2 learns patterns across diverse protein families beyond known annotations.
• Demonstrated feature utility in identifying missing database annotations and guiding sequence generation.

Conclusions:

• PLM representations can be decomposed into meaningful, interpretable components.
• The developed framework provides a feasible and useful method for mechanistically interpreting PLMs.
• This work enhances our understanding of PLMs and opens new avenues for protein design and discovery.