BioInteract combines residue-level cross-attention with protein-language and physicochemical priors for interpretable drug–target interaction prediction

Accurate prediction of drug–target interactions (DTIs) can support the prioritisation of candidate pairs, yet published models are usually compared through numbers quoted across papers whose protocols differ, and their attribution outputs are rarely accompanied by an explicit statement of where the model may be trusted. We present BioInteract, a DTI framework combining a chemistry-aware molecular graph encoder, an ESM-2/physicochemical protein residue encoder with a configured domain-label channel, and bidirectional cross-attention, and we evaluate it inside a matched-protocol benchmark on the Davis kinase inhibitor dataset in which every model is trained with one shared partition, seed, optimiser, schedule and threshold rule. The comparison is a three-model ladder that changes one component at a time, so that adding the frozen ESM-2 representation and then replacing concatenation with cross-attention can be read separately. Both steps help, but by protocol-dependent amounts: the protein language model adds 0.120 and 0.063 AUPRC under the random and Target-ID-held-out protocols, whereas cross-attention adds 0.067 and 0.201 AUPRC under the same two protocols. A breadth check across two further architecture families places these differences in context: on the random split the five architectures tested fall within a narrow band, so the controlled Target-ID-held-out ladder carries the informative contrast. We audit chemical and sequence similarity across all partitions, evaluate exact-sequence-grouped cold-start protocols, and report leakage-aware diagnostic controls. On a strict exact-entity-novel BindingDB cohort, whose construction is documented as a complete row-level filtering waterfall for both the curated-articles snapshot and the full BindingDB release, the frozen Davis checkpoint achieved an AUROC of 0.560 and an AUPRC of 0.340; an applicability-domain analysis attributes this limited transfer principally to ligand and target novelty relative to Davis. The attribution analyses generate hypotheses about model-relevant atoms and residues, but, in line with evidence from models trained with explicit non-covalent interaction supervision, unsupervised attention is not treated here as a recovered interaction mechanism. BioInteract therefore contributes an audited, reproducible and domain-of-applicability-aware DTI reference rather than a new architectural principle.