Docking-Aware Attention

Docking-Aware Attention (DAA) is a protein representation architecture from the Technion that makes an enzyme's embedding depend on the molecule it is acting upon. It was published at CIKM 2025.

The motivation is a property of enzymes that static representations cannot express. The same protein routinely catalyzes different transformations depending on its substrate, because different molecular partners engage different pockets and residues. Existing approaches to computational reaction prediction either ignore the protein entirely, encode it as a discrete EC-number token, or attach a single fixed embedding from a protein language model. All three give an enzyme one representation regardless of what it is reacting with, which throws away exactly the context that determines the outcome.

DAA instead computes a fresh representation for each protein-molecule pair. It runs molecular docking to obtain physical interaction scores between the small molecule and each residue, and folds those scores into an attention mechanism over per-residue embeddings from a frozen protein language model. The pooled protein vector therefore concentrates on the regions the specific substrate actually contacts. Because the underlying encoders stay fixed, only the attention module is trained.

#Key Features

• Physics-informed attention: Docking-derived interaction scores are combined with learned attention weights, so the model is guided by predicted binding geometry rather than having to infer it from sequence alone.
• Ensemble of poses: Docking is sampled to produce multiple plausible binding configurations per pair, which captures proteins with more than one viable binding site instead of committing to a single pose.
• Frozen backbones: Both the protein encoder and the docking model are used off the shelf, keeping the trainable footprint to the attention module.
• Interpretable by construction: The attention weights are per-residue and shift with the molecular partner, so ablations and visualizations can show how the same enzyme redistributes its focus across different substrates.
• Transfers beyond its training task: Representations learned on enzymatic reaction prediction carry over to drug-target interaction without retraining the encoders.

#Technical Details

The base protein encoder is ESM3-6B (see ESM-3), which produces one embedding per residue; binding poses and interaction scores come from DiffDock, a diffusion-based docking model that samples multiple configurations per protein-ligand pair. Docking scores modulate attention over the residue embeddings, and the resulting context-dependent protein vector is injected as a special token at the start of a sequence-to-sequence reaction predictor. Training uses ECREACT, the largest public biocatalysis dataset, aggregating roughly 60,000 enzymatic reactions from Rhea, BRENDA, PathBank, and MetaNetX, augmented as is standard with roughly one million non-enzymatic reactions from USPTO. Evaluation is by top-k accuracy. Against baselines that use static protein embeddings, DAA reaches 62.2% accuracy on complex molecules versus 56.79%, and 55.54% on innovative reactions versus 49.45%; overall it reports 71.48% top-5 accuracy against 65.62% for the ESM3 baseline, and outperforms both the chemistry-only baseline and the EC-token approach. On drug-target interaction, the learned representations reach 76.12% AUC on BioSNAP without retraining the underlying encoders.

#Applications

The immediate use is enzymatic synthesis planning: given a starting material and a candidate enzyme, predict the products, which is the inner loop of retrosynthesis for biocatalytic routes in green chemistry and pharmaceutical manufacturing. The gains are largest exactly where such planning is hardest — structurally complex substrates and reaction types underrepresented in the training data. Beyond biocatalysis, the substrate-dependent representations are a drop-in protein featurization for any task where a protein is scored against a specific small molecule, which is why the drug-target interaction transfer matters. The interpretable attention maps also give a medicinal or protein chemist a readable hypothesis about which residues a given substrate engages.

#Impact

DAA reframes protein representation for interaction tasks as a conditional problem rather than a lookup, and shows that a modest attention module over frozen encoders is enough to capture the conditioning — a cheaper path than training a joint protein-molecule model from scratch. The dependence on a docking step is the principal constraint, since the quality of the representation is bounded by the quality of the predicted poses and docking adds inference cost for every pair scored. Evaluation is computational throughout, on retrospective reaction datasets rather than assayed transformations. The implementation — model code, preprocessing, and training scripts — is on GitHub; the README states MIT terms, though the repository carries no license file and no trained checkpoints.