Merizo: a rapid and accurate protein domain segmentation method using invariant point attention

Area of Science:

• Structural Biology
• Bioinformatics
• Computational Biology

Background:

• The AlphaFold Protein Structure Database offers predictions for over 200 million proteins, significantly impacting structural biology research.
• Efficient tools are needed to navigate, discover, and document the vast contents of this database.
• Identifying protein domain regions is crucial for understanding structure-function relationships, drug discovery, and comparative genomics.

Purpose of the Study:

• To develop a deep learning method for efficient protein domain segmentation.
• To enable better exploration and analysis of protein structures within large databases like AlphaFold.
• To facilitate downstream applications in drug discovery and genomics.

Main Methods:

• A deep learning approach named Merizo was developed for bottom-up residue clustering into domains.
• Merizo was trained on CATH (Common Architecture Tool Kit) domains.
• The model was fine-tuned on AlphaFold2 models using self-distillation for broad applicability.

Main Results:

• Merizo successfully segments protein domains in both experimental and AlphaFold2 models.
• Application to the human proteome identified 40,818 putative protein domains.
• These identified domains show potential for matching with known CATH representative domains.

Conclusions:

• Merizo provides an effective deep learning solution for protein domain segmentation.
• The method enhances the utility of large-scale protein structure databases.
• This facilitates advancements in structural biology, drug discovery, and comparative genomics.