*4.2. Protein–Protein Docking*

The 527 representative conformations of N-WASP domain V were docked into two crystallographic structures of actin (PDB ID: 2VCP [27] and 3M3N [26]), using the molecular modeling library PTools [36]. This toolbox performs rigid-body docking of coarse-grained proteins by multiple energy minimizations, starting from regularly distributed initial positions and orientations of the ligand around the receptor surface. It should be emphasized that no conformational change was allowed during these docking calculations for both protein partners, notably the intrinsically disordered domain V. The energy function minimized here is the physics-based pairwise protein–protein interaction energy SCORPION [37,38]. Then, to better discriminate the near-native interface between actin and domain V, the complexes previously generated with PTools were rescored using a knowledge-based scoring function which additionally takes into account three-body interactions. We used in this study the 2/3B*best* InterEvScore, without any evolutionary information from the actin or N-WASP domain V sequences [30].

The performance of 2/3B*best* InterEvScore was positively evaluated on an ensemble of 131 protein–protein complexes which, as far as we know, did not include IDP case [30]. Thus, to assess the validity of our approach to study the actin–domain V recognition, we performed the redocking of the folded segment 433–451 of N-WASP domain V into actin structure 2VCP [27] and checked if the X-ray structure of the complex can be retrieved. The results of this test are reported in Figure S4, which displays the actin–ligand interaction 2/3B*best* score as a function of the RMSD relative to the peptide conformation in the crystallographic structure. It can be seen that the coarse-grained protein–protein redocking is able to retrieve the experimental structure with a RMSD calculated over the C*α* atoms of only 0.5 Å. In this particular case, the modeled complex structure, which is the closest

to the experimental one is ranked first (the higher is the score, the more native-like is the interface). This benchmark led us to adopt this two-step approach consisting in generating complex structures with PTools and rescoring them with InterEvScore.
