Explaining how mutations affect AlphaFold predictions.

Area of Science:

• Artificial Intelligence
• Structural Biology
• Computational Biology

Background:

• Transformer models are advanced neural networks driving AI progress.
• Understanding the internal mechanisms of these models, especially in protein structure prediction, remains challenging.

Purpose of the Study:

• To investigate how transformer models within AlphaFold select protein conformations.
• To develop a method for identifying critical amino acid patterns influencing these predictions.

Main Methods:

• Developed the Conformational Attention Analysis Tool (CAAT) algorithm.
• CAAT identifies amino acid positions significantly impacting AlphaFold predictions upon modification.
• Validated CAAT findings through experimental modifications.

Main Results:

• CAAT successfully identified sparse amino acid patterns crucial for AlphaFold's conformational selection.
• Experimental modifications at CAAT-identified positions substantially altered predictions.
• Modifications at non-CAAT-identified positions had minimal impact on predictions.

Conclusions:

• CAAT effectively pinpoints amino acids vital for protein structure prediction.
• This tool narrows the search space for impactful mutations in protein engineering.
• The framework offers potential applications for interpreting other transformer-based neural networks.