OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization

Area of Science:

• Computational Biology
• Structural Biology
• Artificial Intelligence in Biology

Background:

• AlphaFold2 achieved high accuracy in protein structure prediction.
• Existing AlphaFold2 implementations lack necessary code and data for retraining.
• Retraining is crucial for new tasks and understanding model generalization.

Purpose of the Study:

• Introduce OpenFold, a trainable and efficient implementation of AlphaFold2.
• Validate OpenFold's performance against AlphaFold2.
• Investigate OpenFold's generalization capabilities and learning process.

Main Methods:

• Developed OpenFold, a fast and memory-efficient AlphaFold2 implementation.
• Trained OpenFold from scratch on protein structure data.
• Analyzed model generalization with limited training sets and studied intermediate training structures.

Main Results:

• OpenFold achieved accuracy comparable to AlphaFold2.
• Demonstrated robust generalization even with reduced training data diversity.
• Provided insights into the hierarchical nature of protein folding learned by the model.

Conclusions:

• OpenFold is a powerful and versatile open-source tool for protein structure prediction.
• It facilitates tackling new challenges like protein-ligand complex prediction.
• OpenFold will be a valuable resource for the structural biology and protein modeling community.