GeoFlowMAT: Interface-Repair Flow Matching with Triangle-Consensus Reasoning for Transferable Bio Molecular Binding Affinity Prediction

Predicting biomolecular binding affinity remains fragmented across interaction types: models tuned for protein--protein mutations, protein--ligand complexes, or drug--target pairs rarely share a transferable representation. We introduce GeoFlowMAT, an all-atom model that learns such a representation directly from interface geometry and can be adapted across affinity regimes. GeoFlowMAT uses conditional flow matching for interface-repair pretraining: residues are displaced as coherent local frames, and the model learns to restore native cross-partner geometry without affinity labels. The backbone couples atom-, token-, and pair-level states through interface-aware hierarchical blocks and triangle-aware pair reasoning; triangle-consensus gating captures shared geometric context while avoiding the activation bottleneck of full triangle attention. Across five benchmarks covering mutational effects, absolute structural affinity, and drug--target binding, the same pretrained trunk improves over the matched ADiT baseline and transfers zero-shot to HER2 antibody mutations. The gains are not only predictive: matched ablations show that structured interface repair is a substantial source of transfer, learned pair gates recover held-out cross-chain contacts, and triangle consensus retains near-attention accuracy with 8.2 times lower peak memory. Together, these results show that interface geometry can serve as a reusable all-atom representation for broad biomolecular affinity prediction.