A Structure-Aware Generative AI Framework for Revealing Functional Relationships in Proteins Families.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Protein Science

Background:

• Protein sequence databases grow faster than experimental structure determination, leaving many sequences unannotated.
• Protein folds are more conserved and informative than sequences, offering a powerful analysis tool.

Purpose of the Study:

• To develop a generative, structure-aware framework for mapping, clustering, and designing protein sequences.
• To leverage geometric encoding and coevolutionary constraints for enhanced protein analysis.

Main Methods:

• Employing the 3D interaction (3Di) alphabet for discrete geometric representation of protein local structures.
• Utilizing ProstT5 for bidirectional translation between amino acid sequences and 3Di representations.
• Integrating 3Di alignments with Direct Coupling Analysis (DCA) and Variational Autoencoders (VAE) within a latent generative landscape.

Main Results:

• Achieved sensitive homology detection and structure-guided protein sequence generation.
• Enhanced detection of coevolutionary signals and rational sampling of structural variants.
• Demonstrated improved contact prediction, homology inference, and sequence generation across diverse protein families.

Conclusions:

• The integrative framework provides a quantitative, generative view of protein structure space.
• Advances protein evolution and design studies by incorporating structural information.
• Enables sensitive analysis and design even for remote homologs with low sequence identity.