How well do contextual protein encodings learn structure, function, and evolutionary context?

Area of Science:

• Computational Biology
• Bioinformatics
• Protein Engineering

Background:

• Protein function is dictated by amino acid sequence, influenced by structural, evolutionary, and functional contexts.
• Understanding these contextual relationships is crucial for predicting protein behavior and designing novel proteins.

Purpose of the Study:

• To train masked label prediction models for learning residue representations in diverse protein contexts.
• To investigate how pretraining and fine-tuning contextual encodings improve specialized protein representations.
• To explore the utility of learned representations in predicting protein structure, flexibility, and interactions.

Main Methods:

• Trained masked label prediction models on protein sequences to learn contextual residue representations.
• Sampled sequences from learned representations to assess their ability to fold into template structures.
• Evaluated generated sequences for evolutionary conservation, structural plasticity, and binding energies at protein-protein interfaces.

Main Results:

• Learned representations successfully generated sequences that fold into template structures and reflect evolutionary variations.
• For flexible proteins, sampled sequences explored the full conformational space, indicating encoded plasticity.
• Generated sequences accurately replicated wild-type binding energies at protein-protein interfaces in silico.
• Fine-tuning captured conserved patterns at antibody-antigen interfaces, while pretraining enhanced sequence recovery for the H3 loop.

Conclusions:

• Contextual encodings derived from masked label prediction models provide powerful representations of amino acid residues.
• These representations effectively capture structural, evolutionary, and functional properties of proteins.
• The approach holds promise for protein design, understanding protein dynamics, and analyzing protein interactions.