Nearest neighbor search on embeddings rapidly identifies distant protein relations

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• State-of-the-art homology detection relies on protein sequences and PSSMs.
• Protein Language Models (pLMs) offer a generalized approach using embeddings.

Purpose of the Study:

• To explore pLM embeddings for remote homology detection (<20% PIDE).
• To assess the efficacy of nearest neighbor searches in protein embedding space.

Main Methods:

• Utilized pLM embeddings for nearest neighbor searches.
• Applied methods to protein pairs with diverged sequences.
• Employed challenging dataset separations for robust evaluation.

Main Results:

• pLM embeddings excelled in detecting remote homology for single-domain proteins.
• Current limitations were identified for multi-domain proteins.
• Crucial role of dataset separation in evaluating embedding space behavior was observed.

Conclusions:

• pLM embeddings represent a promising avenue for homology detection, particularly for single-domain proteins.
• Further research is needed to address challenges in multi-domain protein analysis.
• Developed methods are publicly available to facilitate further research.