DualBAN

A compound-protein interaction predictor built around a bilinear attention network learns something genuinely useful: which atoms of a ligand line up against which stretch of a target's sequence, read straight off an attention map. But the encoders feeding that map — a graph network over the molecule, a convolutional stack over the residue string — are trained from the interaction labels and nothing else. Everything such a model knows about chemistry and protein biology has to be recovered from a few hundred thousand binary labels, which is why these predictors degrade on compounds and targets unlike the ones they were fit on.

Pretrained sequence models carry exactly the knowledge that is missing, but they cannot simply replace the interaction module: a language model emits one pooled vector per molecule and one per protein, with no notion of which part of one meets which part of the other. DualBAN runs both paths at once. Atom-level and residue-level features still go through a bilinear attention network to model the inter-molecular contact pattern, while pooled representations from a chemical language model and a protein language model are fused by cross-attention to carry the intra-molecular context. The two joint vectors are concatenated before the classifier, so neither view has to stand alone.

The model comes from a group at Quzhou People's Hospital — the Quzhou affiliated hospital of Wenzhou Medical University — with a collaborator at Xiamen University, published in IEEE Journal of Biomedical and Health Informatics in September 2026. Its bilinear attention formulation and its evaluation protocol both follow DrugBAN, the 2023 model that introduced that component to drug-target prediction; DualBAN's contribution is the second, sequence-level pathway grafted onto it.

#Key Features

• Two interaction views in one classifier: A bilinear attention map over atoms and residues captures local contacts, while cross-attention over pooled sequence embeddings captures molecule-wide and protein-wide context. Both feed the same decision.
• Frozen pretrained backbones: Protein features come from ESM-2 at the 3B scale and compound features from ChemBERTa, precomputed once per unique sequence and SMILES string rather than fine-tuned.
• Interpretable contact map: The bilinear attention weights are recoverable at evaluation time, so a prediction can be traced back to specific atoms and residue windows.
• Cross-domain and cold-start evaluation: Beyond random splits, the model is tested on clustering-based source/target splits and on a cold-start split where test compounds and proteins are absent from training.

#Technical Details

The intra-molecular encoders follow DrugBAN closely. Molecules become 2D graphs with 75-dimensional canonical atom features, padded to 290 nodes, embedded to 128 dimensions and passed through three 128-unit graph convolution layers. Proteins are integer-encoded over a 26-symbol residue alphabet, embedded to 128 dimensions, and passed through three 1D convolutions with kernel widths 3, 6 and 9. A two-head bilinear attention layer pools these into a 256-dimensional joint representation.

The sequence pathway uses the CLS embedding from ChemBERTa-77M-MTR (384 dimensions) for each SMILES string and the mean-pooled residue representations from esm2_t36_3B_UR50D (2,560 dimensions) for each protein, truncated at 1,022 residues. Each is projected to 256 dimensions by a two-layer GELU network, then passed through paired four-head cross-attention blocks — compound attending to protein and protein attending to compound — whose outputs are summed. Concatenating the two 256-dimensional vectors gives a 512-dimensional input to a four-layer MLP with a single sigmoid output, trained with binary cross-entropy at a learning rate of 5e-5, batch size 64, for up to 100 epochs. Benchmarks are BindingDB, BioSNAP, Human, C. elegans and KIBA, scored by AUROC, AUPRC, F1, sensitivity, specificity and accuracy; the cross-domain setting reuses DrugBAN's conditional domain-adversarial module.

#Applications

The target use is hit prioritization: scoring a library of candidate compounds against a protein target, or a panel of targets against a compound of interest, before anything is docked or assayed. Because the two backbones are frozen and their embeddings cached per unique molecule and per unique sequence, the marginal cost of an additional pair is only the small trainable network, which suits repeat screening over a fixed library. The recoverable attention map gives medicinal chemists a substructure-level rationale to check against known pharmacophores.

#Impact

DualBAN is a peer-reviewed demonstration that pretrained chemical and protein language models complement, rather than replace, explicit pairwise interaction modelling — the two signals are additive enough to justify carrying both. Its practical reach is limited by what has been released: the training and evaluation code is public on GitHub, but the repository carries no license, no README and no trained checkpoints, so reproducing the results requires re-running the pipeline from the bundled benchmark splits. Evaluation is retrospective, on curated datasets with a binary label, and the output is an interaction probability rather than a binding affinity, so the model ranks candidates rather than estimating potency.