Summary
Training Structure prediction neural networks on both interacting and non-interacting pairs of proteins improves precision and recall when discriminating of positive and negative protein-protein interactions (1). This suggests limited generalization on the part of vanilla structure prediction models. This is comparable to how protein folding neural network outputs are not predictive of monomeric stability.
1.
Zhang J, Humphreys I, Pei J, Cong Q. Computing the human interactome. Structural Dynamics. 2025;12(5_Supplement):A176–A176. Available from: https://doi.org/10.1063/4.0000966