From memorization to generalization: Why physics will improve machine learning -based prediction of protein complexes

AlphaFold-like models have transformed monomer structure prediction yet reliable, generalizable prediction of protein-protein interactions (PPIs) remains challenging, particularly for antibody-antigen docking. These limitations often stem from reliance on pattern memorization under severe data scarcity. Here, we review the emerging transition toward physics-integrated machine learning to address these gaps. We categorize recent efforts to improve generalization into three complementary approaches: (I) enriching inputs with physics-based sampling (e.g., molecular dynamics/fast Fourier transform ensembles); (II) designing architectures with strict geometric inductive biases (e.g., SE(3)-equivariance); and (III) constraining the generative process using physical energy functions or potentials. While standard models often struggle on out-of-distribution targets, these hybrid strategies aim to enforce physical plausibility at different stages of the prediction pipeline, offering a path from memorization to true physical generalization.