Fig. 3: Recovery of peptide–protein interface characteristics by AF2 models.
From: Harnessing protein folding neural networks for peptide–protein docking
a AF2 models identify a significant fraction of the receptor binding pocket residues, even though coverage is reduced compared to PFPD models. Box-and-whisker plots are shown for motif (n = 12, blue) and non-motif (n = 14, red) sets, with median as center line, quartiles as box limits, and lower/upper whiskers extending to the maximum/minimum data points within the interquartile range, respectively. b Computational alanine scanning recovers well-predicted interface hotspots, mostly for accurate models (within 2.5 Å RMSD): Comparison of computational Rosetta alanine scanning results applied to the best model vs. the native structure. The vertical and horizontal lines represent the threshold ΔΔG = 1.5 kcal/mol used to define interface hotspots. The different colors represent distinct bins of RMSD (of the model used for alanine scanning). The number of residues in each quadrant is indicated (counts from models below 2.5 Å are shown in parentheses). c Performance is significantly reduced if the peptide sequence is changed to poly-alanine, both for motif and non-motif sets. Source data are provided as a Source Data file.
