Multi-scale structural similarity embedding search across entire proteomes.

Area of Science:

• Structural Biology
• Bioinformatics
• Computational Biology

Background:

• The exponential growth of 3D biomolecular structure data, fueled by AI/deep learning (DL) predictions, necessitates advanced similarity search tools.
• Conventional structure comparison methods are computationally intensive and struggle with the scale of modern structural databases.

Purpose of the Study:

• To develop a scalable and efficient method for searching large repositories of 3D biomolecular structures.
• To enable rapid identification of similar structures, including both experimentally determined and AI/DL-predicted models.

Main Methods:

• Utilized protein language models and a deep neural network to generate fixed-length vector embeddings for 3D structures.
• Integrated these embeddings with vector databases for efficient large-scale similarity retrieval.
• Developed a method capable of comparing single-domain, full-length, and multimeric protein structures.

Main Results:

• The developed method efficiently navigates vast collections of 3D structure data.
• The vector embedding approach allows for rapid large-scale structure comparisons.
• The model demonstrates generalization beyond its training scope, accurately assessing similarity for diverse structural types.

Conclusions:

• The new strategy provides a scalable solution for structure similarity searching in the era of AI-driven structural biology.
• This approach addresses the computational challenges posed by the increasing volume of structural data.
• Facilitates efficient retrieval and analysis of both experimental and predicted 3D biomolecular structures.