Deep-learning structure elucidation from single-mutant deep mutational scanning.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Deep learning, exemplified by AlphaFold2, has greatly advanced protein structure prediction.
• However, current methods face limitations in accurately predicting numerous protein systems.
• Accurate protein structure prediction is crucial for understanding biological function and disease.

Purpose of the Study:

• To enhance AlphaFold2's protein structure prediction accuracy using deep mutational scanning (DMS) data.
• To develop a novel computational method, DMS-Fold, integrating sparse residue burial restraints from DMS.
• To validate the efficacy of DMS-Fold against AlphaFold2 using simulated and experimental data.

Main Methods:

• Extracted sparse residue burial information from deep mutational scanning (DMS) datasets.
• Integrated DMS-derived burial restraints to guide residue placement during protein structure generation.
• Developed and validated the DMS-Fold computational tool, comparing its performance against AlphaFold2.

Main Results:

• DMS-Fold significantly enhanced AlphaFold2 predictions for 88% of tested protein targets.
• A notable improvement in TM-Score (greater than 0.1) was observed for 252 proteins.
• The method demonstrated superior performance in refining protein structures using experimental and simulated DMS data.

Conclusions:

• Integrating sparse residue burial restraints from DMS effectively refines deep learning-based protein structure prediction.
• DMS-Fold offers a substantial improvement over standard AlphaFold2, particularly for challenging protein systems.
• DMS-Fold is freely available, providing a powerful, accessible tool to advance structural biology research.