Has AlphaFold 3 Solved the Protein Folding Problem for D-Peptides?

Area of Science:

• Computational biology and structural bioinformatics.
• Protein design and drug discovery.

Background:

• D-peptides offer therapeutic advantages over L-peptides due to improved stability and bioavailability.
• Accurate prediction of chirality and binding pose is crucial for computational design of D-peptides.
• AlphaFold 3 (AF3) is a deep learning model designed to predict chemical structures, including chiral centers.

Purpose of the Study:

• To evaluate the accuracy of AlphaFold 3 in predicting the fold, chirality, and binding pose of D-peptides.
• To assess AF3's performance in modeling heterochiral complexes of D-peptides and L-proteins.

Main Methods:

• Conducted 3,255 experiments using AF3 with explicit D-stereocenter inputs for D-peptides.
• Evaluated AF3's predictions for chirality, fold, and binding pose accuracy against empirical data.
• Assessed the impact of increasing prediction seeds on chirality violation rates.

Main Results:

• AF3 exhibited a high chiral violation rate of 51% for D-peptide binders, significantly deviating from the reported 4.4% for diverse chiral molecules.
• Predictions showed incorrect folds and binding poses, with D-peptides frequently misoriented in L-protein binding pockets.
• AF3 confidence metrics failed to correlate with prediction accuracy, unable to distinguish correct from incorrect models.

Conclusions:

• AlphaFold 3 is a poor predictor of D-peptide chirality, fold, and binding pose in heterochiral complexes.
• The model's performance is comparable to random chance for predicting D-peptide stereochemistry.
• Further development is required for accurate computational prediction of D-peptide structures and interactions.