Severe deviation in protein fold prediction by advanced AI: a case study

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Artificial intelligence (AI) and deep learning have advanced protein structure prediction.
• AlphaFold demonstrates high accuracy in modeling protein folds from amino acid sequences.
• Experimental structure determination is essential despite AI advancements.

Purpose of the Study:

• To compare AI-predicted protein structures with experimental data.
• To highlight discrepancies in domain orientation and residue positioning.
• To identify limitations in current AI-driven protein structure prediction.

Main Methods:

• Utilized AI (AlphaFold) for protein structure prediction.
• Determined experimental structure of a two-domain protein.
• Compared AI predictions against experimental data, analyzing domain relative orientation and positional divergence.

Main Results:

• Observed severe deviations between experimental and AI-predicted structures.
• Found positional divergence in equivalent residues exceeding 30 Å.
• Calculated an overall Root Mean Square Deviation (RMSD) of 7.7 Å between models.

Conclusions:

• AI predictions can significantly deviate from experimental structures.
• Deviations may stem from unusual conformations, limited training data, or protein folding complexity.
• Experimental validation is indispensable for accurate protein structure determination.