Foldclass and Merizo-search: scalable structural similarity search for single- and multi-domain proteins using...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Large-scale protein structure data from computational methods present storage and comparison challenges.
• Multi-domain protein structures, abundant in databases like AlphaFold, require advanced comparison techniques.

Purpose of the Study:

• To develop a fast and accurate method for comparing protein structures, particularly multi-domain ones.
• To enable efficient searching and relationship detection within vast protein structure datasets.

Main Methods:

• Utilized Foldclass, an embedding-based method for detecting structural similarity between protein domains.
• Combined Foldclass with Merizo, a deep learning tool for automatic domain segmentation.
• Developed Merizo-search to segment query structures and search a Foldclass embedding database for domain similarities.

Main Results:

• Foldclass embeddings demonstrate accuracy in homology detection.
• Merizo-search rapidly segments multi-domain structures and identifies top domain matches.
• Searched 365 million domains from the Encyclopedia of Domains in as little as 2 minutes.

Conclusions:

• Merizo-search effectively addresses challenges in comparing large numbers of protein structures, especially multi-domain ones.
• The tools facilitate rapid per-domain similarity detection for complete protein chains.
• These advancements will support analyses of the growing wealth of predicted protein structural data.