PLM-eXplain: divide and conquer the protein embedding space.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein language models (PLMs) offer advanced sequence representations for computational biology tasks.
• The 'black-box' nature of PLMs hinders biological interpretation and the translation of predictions into actionable insights.
• There is a need for methods that provide interpretable explanations for PLM behavior while maintaining predictive accuracy.

Purpose of the Study:

• To develop an explainable adapter layer, PLM-eXplain (PLM-X), for protein language models.
• To bridge the gap between PLM predictive power and biological interpretability.
• To enable actionable insights from PLM predictions in computational biology.

Main Methods:

• PLM-eXplain (PLM-X) is an adapter layer that decomposes PLM embeddings into interpretable and residual subspaces.
• The interpretable subspace integrates established biochemical features like secondary structure and hydropathy.
• Embeddings from established PLMs (ESM2, ProtBert) are utilized within the PLM-X framework.

Main Results:

• PLM-X successfully incorporates biochemical features into PLM embeddings.
• High predictive performance was maintained across multiple classification tasks, including extracellular vesicle association, transmembrane helix prediction, and aggregation propensity prediction.
• PLM-X demonstrated the ability to provide biological interpretations for model decisions without compromising accuracy.

Conclusions:

• PLM-eXplain (PLM-X) offers a generalizable solution for enhancing the interpretability of protein language models.
• This approach facilitates biological interpretation of PLM predictions across diverse downstream applications.
• PLM-X enables actionable biological insights derived from complex protein sequence models.