Performance of pyDock in 8th CAPRI: Energy-Based Scoring Applied to Docking and AlphaFold Models

The 8th CAPRI edition has shown a significant evolution in the field of protein-protein complex structure prediction. We have participated in all 11 targets proposed in this edition, involving domain-domain, protein-protein, protein-peptide, and protein-DNA interactions, including homo- and hetero-meric interfaces. Our prediction strategy has significantly evolved during this edition due to the appearance of ground-breaking AI-based predicting methodologies, like AlphaFold (AF). As predictors, while for the first targets our modeling approach was mostly based on our standard pyDock protocol, after target T187 and onwards, interacting subunits were routinely modeled by AlphaFold2. In the last round (targets T231-T234) we also generated complex models with AlphaFold-Multimer, which were scored by a combination of pyDock energy and AF model confidence. As scorers, we mostly used pyDock to directly score the provided models, applying the same restraints and additional filters as in predictors. Overall, our performance in this 8th CAPRI edition was in line with that in past editions. Considering successful targets as those with submitted models of acceptable (or better) quality within top 5 (predictors) or top 10 (scorers) for the entire complex or at least for one of their assessment units (AUs), we succeeded in 45% of cases as predictors (ranking 4th among participants) and in 64% of cases as scorers (ranking 3rd among participants). The targets where we failed both as predictors and as scorers were actually challenging for all participants. This experiment has shown that the problem of protein-protein docking is not yet solved, and has confirmed the value of energy-based scoring and other approaches in combination with AI-based predictions.