Benchmarking AlphaFold3-like Methods for Protein-Peptide Complex Prediction

Protein-short peptide interactions are central mechanisms regulating signal transduction, chaperone functions, and drug targeting design. However, the precise prediction of their complex structures has long been constrained by the limited accuracy of traditional methods. The breakthrough of AlphaFold2 resolved the challenge of predicting protein monomer structures. Its derivative method, AlphaFold2-multimer, surpassed traditional docking approaches in predicting protein complexes, yet its accuracy remains suboptimal. AlphaFold3 and its replication models, such as Protenix, Chai-1, and Boltz-1, have further improved prediction accuracy for protein-protein interactions. To evaluate the performance of these methods in protein-short peptide structure prediction, this study conducted predictions and comprehensive comparisons on two benchmark datasets of protein-short peptide complexes. One dataset did not intentionally exclude structures from the training set of these methods, while the other did. For the former, the results demonstrate that the new-generation structural method significantly increased the success rate under stringent criteria (Fnat≥ 0.8) to 70-80%, compared to AlphaFold2-multimer (53%), with Protenix achieving the highest accuracy at 80.8%. For the second dataset, AlphaFold2-multimer showed consistent performance, but the success rates of the new-generation methods all dropped significantly (to 40-56%). Our analysis suggests that the main challenge for the performance decline comes from the accurate prediction of binding sites. For all datasets, A multi-method combination strategy enabled a higher-quality prediction success rate compared to the individual models, highlighting synergistic advantages among methods. Notably, inter-model consistency analysis provided an efficient metric for selecting optimal predictions.