Blind Assessment of Monomeric AlphaFold2 Protein Structure Models with Experimental NMR Data

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Recent advances in AI-driven molecular modeling are transforming structural biology.
• AlphaFold-2 (AF2) predicts protein structures with high accuracy using deep learning.
• The performance of AF2 on solution Nuclear Magnetic Resonance (NMR) data, particularly for proteins absent from its training set, requires thorough assessment.

Approach:

• Evaluated AF2's ability to model small, monomeric proteins using NMR data not included in AF2's training set.
• Utilized nine "blind" protein NMR datasets (chemical shift, FID, NOESY, RDC) for proteins lacking structural homologs in the Protein Data Bank at the time of AF2 training.
• Assessed AF2 model fit to experimental NMR data using established validation tools.

Key Points:

• AF2 models were generated for nine small, monomeric proteins (70-108 residues) with unique NMR data.
• AF2 models demonstrated a near-equivalent or superior fit to experimental NMR data compared to existing NMR structure models.
• Software such as RPF-DP, PSVS, and PDBStat were employed for structure quality and Residual Dipolar Coupling (RDC) assessment.

Conclusions:

• AF2 shows significant potential for modeling protein structures, even in data-scarce scenarios.
• The findings provide benchmark NMR data for evaluating new structure prediction and NMR analysis methods.
• AF2 can serve as a valuable tool for guiding NMR data analysis and generating hypotheses in structural biology.