Fig. 1: AF2 can be adapted to accurately model many peptide–protein interactions.

a Concept of peptide–protein docking with a poly-glycine linker, successfully identified as unstructured region and modeled as a circle, placing the peptide at its correct position (left; example PDB ID 1ssh⁶³), or out in space (right; example PDB ID 2orz⁶⁴). The native peptide is shown in black. b Cumulative performance of AF2 and PFPD (dashed lines) on motif (12 complexes, blue), non-motif (14 complexes, red), and LNR (96 complexes, AF2 only, yellow) sets, as measured over the interface residues of the peptide (Left - Peptide interface: after aligning the receptor) or the full interface (Right - CAPRI Irms: after aligning the whole interface. See also Supplementary Fig. 2). c Correlation between performance of AF2 and PFPD for the motif and non-motif PFPD sets. Triangles indicate values over 15.0 Å RMSD. Left: PDB ID 2b9h³⁸, interaction between MAPK Fus3 and a peptide derived from MAPKK Ste7, where PFPD positions the peptide within the pocket, but in a flipped orientation (N-termini are indicated by spheres). Right: PDB ID 1awr⁶⁵, interaction between CypA and a HIV-1 Gag polyprotein derived peptide, where AF2 models the peptide within the pocket but in a wrong position. Shown are the peptide structures generated by the linker model (cyan, blue) or PFPD (magenta), and the crystal structure (black). d Overall assessment of performance measured for the individual partners (after aligning each separately). Source data are provided as a Source Data file.