MotifAE Reveals Functional Sequence Patterns from Protein Language Model: Unsupervised Discovery and Interpretability...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in protein science

Background:

• Protein language models (pLMs) capture evolutionary sequence patterns but function as black boxes.
• Interpreting learned patterns is crucial for understanding protein function and engineering.

Purpose of the Study:

• To develop an unsupervised framework, MotifAE, for interpreting patterns learned by pLMs.
• To enable the discovery and analysis of functional motifs and structural domains within protein sequences.

Main Methods:

• Developed MotifAE, an unsupervised framework using a sparse autoencoder (SAE) architecture.
• Incorporated a smoothness loss to improve feature coherence and motif identification.
• Projected pLM embeddings into an interpretable, sparse latent space.

Main Results:

• MotifAE successfully identified known functional motifs and diverse sequence patterns.
• The framework captured structural domains, correlating feature activation with residue importance and domain function.
• Identified features associated with domain folding stability, enabling improved stability prediction and engineering.

Conclusions:

• MotifAE provides a general framework for systematic sequence pattern discovery and interpretation.
• The approach advances protein function analysis, mutation effect interpretation, and rational protein engineering.
• MotifAE facilitates the engineering of proteins with enhanced stability.