AFToolkit: a framework for molecular modeling of proteins with AlphaFold-derived representations.

Area of Science:

• Computational Biology
• Protein Engineering
• Structural Bioinformatics

Background:

• Understanding mutation effects on protein fitness and stability is crucial for protein engineering.
• Current models often fine-tune or pretrain large models, which can be resource-intensive.
• AlphaFold2 (AF2), while powerful for structure prediction, has limitations in assessing mutation impacts.

Purpose of the Study:

• To develop a modified AlphaFold2 inference method for improved prediction of mutation effects on protein structure.
• To introduce AFToolkit, a framework for protein engineering tasks using modified AF2 embeddings.
• To provide a computationally efficient alternative to extensive model retraining.

Main Methods:

• Modified AlphaFold2 inference by discarding multiple sequence alignment and masking templates during recycling.
• Developed AFToolkit framework using modified AF2 embeddings and adapter models.
• Directly modified input protein sequences to handle multiple mutations, insertions, and deletions.

Main Results:

• Achieved strong performance on protein engineering benchmarks.
• Demonstrated high Spearman correlation coefficients: 0.68 on PTMul, 0.60 on cDNA-indel, and 0.57 on C380.
• Showcased the utility of AFToolkit for various protein engineering tasks.

Conclusions:

• The proposed modification of AlphaFold2 inference effectively captures the structural impacts of amino acid mutations.
• AFToolkit offers a practical and efficient solution for protein engineering challenges.
• This approach avoids the need for extensive fine-tuning or pretraining of large models.