Inferring Local Protein Structural Similarity from Sequence Alone.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Identifying local protein structural similarity is crucial for understanding protein function and evolution.
• Current methods often rely on 3D structural models, limiting scalability and accessibility.
• Protein language models (pLMs) have shown promise in capturing biological information from sequences.

Purpose of the Study:

• To investigate if protein language models (pLMs) can implicitly capture fine-grained structural signals from sequence data alone.
• To develop a novel framework for detecting locally aligned structural regions directly from protein sequences.
• To provide a scalable, sequence-based alternative to traditional structure-based methods for protein analysis.

Main Methods:

• Utilizing mean-pooling of residue embeddings from pLMs over sliding windows.
• Employing cosine similarity to compare these pooled embeddings across different proteins.
• Developing a framework to identify diagonal patterns indicative of local structural alignment.

Main Results:

• Diagonal patterns in embedding comparisons reveal locally aligned protein regions, even without sequence identity.
• The proposed framework successfully detects locally aligned structural regions directly from sequences.
• A case study on SRC homology 3 (SH3) domains demonstrated the recovery of structurally conserved motifs across diverse sequence contexts.

Conclusions:

• Protein language models implicitly capture structural information, enabling sequence-based detection of local structural similarities.
• The developed framework offers a lightweight and scalable alternative to structure-based methods.
• This approach facilitates high-throughput structural discovery and annotation using only protein sequence data.