GeoFlowMAT: Interface-Repair Flow Matching and Triangle-Consensus Reasoning for Transferable Biomolecular Affinity Prediction

Binding-affinity models are typically specialized by interaction type. We introduce GeoFlowMAT, an all-atom model that learns a transferable interface representation through conditional flow-matching pretraining. Residue centers are displaced while residue-local geometry remains coherent under controlled contraction, and the model restores native cross-partner structure without affinity labels. Its atom-, token-, and pair-level backbone uses triangle-consensus gating to capture shared geometric context without the activation cost of full triangle attention. Across five retrospective benchmarks spanning four interaction modalities, GeoFlowMAT achieves the strongest results in the displayed complex-disjoint SKEMPI comparison, the lowest MSE in the displayed DAVIS comparison while remaining competitive in $r_m^2$, improves over ADiT in a matched PPB-Affinity comparison, and remains competitive on zero-shot HER2 and LBA. Paired-bootstrap gains over ADiT in SKEMPI and auxiliary DAVIS correlation analyses exclude zero (all $p < 0.001$). Controlled ablations identify structured interface repair as a key contributor to transfer; pair gates localize held-out contacts, and triangle consensus retains near-full-attention accuracy while reducing pair-block peak activation memory by $8.2\times$. These results support cross-regime transfer on curated benchmarks and motivate prospective evaluation of GeoFlowMAT as an affinity oracle.