CrossVOI

Deorphanizing a human olfactory receptor is slow work. The receptor has to be expressed in a heterologous cell line and challenged with odorants one at a time, and most human ORs still have no confirmed ligand. Computational screening is the obvious shortcut, but a screen is only useful for mechanism if you can ask why it made a call. CrossVOI answers both halves: it scores whether a volatile organic compound (VOC) activates a given receptor from amino acid sequence and SMILES alone, and it exposes the cross-attention weights behind each prediction as a residue-level map of what the model attended to.

The framework comes from Hui-Meng Lu's group at Northwestern Polytechnical University, with collaborators at the University of Science and Technology of China and Shaanxi Normal University. Both encoders are frozen general-purpose language models: ProtT5 from ProtTrans embeds the receptor sequence, and a ChemBERTa checkpoint trained on PubChem SMILES embeds the ligand. Self-attention stacks refine each modality, a cross-attention encoder fuses them, and a multilayer perceptron emits a binary interaction probability. Only that fusion head is trained, so the pretrained sequence and chemical semantics reach the classifier unchanged.

Attention is the scientific instrument here, not merely an architectural choice. The authors align cross-attention across all 383 human ORs and test what it corresponds to. It tracks sequence conservation — average attention correlates negatively with per-position Shannon entropy (r = -0.394, p < 0.001) and concentrates on the DRY-like motif of TM3 and the NPxxY-like motif of TM7. It does not, on average, favor the ligand-binding pocket; pocket residues surface only in per-receptor attention, once the family average is subtracted.

#Key Features

• Two frozen backbones, one trained head: embeddings are precomputed to HDF5 files outside the training graph, so neither language model is fine-tuned and the trained component stays small.
• Cross-attention as an interpretability readout: shipped scripts dump per-residue cross- and self-attention matrices, turning each prediction into a region-level hypothesis about the receptor.
• Evaluated on unseen receptors and unseen ligands: alongside the conventional pair split, OR-split and VOC-split evaluations share no receptor, respectively no compound, between training and test sets.
• Sequence and SMILES in, no structure required: nothing in the pipeline needs a resolved or predicted OR structure, which matters for a family whose experimental structures remain scarce.
• Data and a trained checkpoint ship together: the repository carries the curated interaction table, receptor sequences and VOC list alongside a fold checkpoint the prediction script loads by default.

#Technical Details

CrossVOI trains on human-30k, extracted from the M2OR database: after merging duplicate VOC-OR pairs and dropping receptors that responded to nothing, 32,201 interactions covering 383 human OR sequences and 654 VOCs, with outcomes from heterogeneous assay systems collapsed to binary responsive/non-responsive labels. Sequences are truncated or padded to 400 residues and 200 SMILES tokens; the 1024-dimensional ProtT5 embeddings and 768-dimensional ChemBERTa embeddings are projected to a shared 256-dimensional space, passed through two self-attention layers per modality and two cross-attention layers with four heads each, then a 512-256 MLP. Over 5-fold cross-validation the pair split reaches AUROC 0.918 ± 0.007 and AUPRC 0.736 ± 0.009 (F1 0.712, precision 0.730, recall 0.696), beating OdoriFy, DrugBAN, a random forest and AutoDock Vina docking on the same protocol. Performance holds above chance but drops sharply under the harder splits — AUROC 0.761 ± 0.030 and AUPRC 0.342 ± 0.022 for unseen receptors, AUROC 0.783 ± 0.046 and AUPRC 0.426 ± 0.068 for unseen compounds.

#Applications

The immediate use is triage: ranking odorant libraries against uncharacterized ORs so the heterologous expression assays that would otherwise be run blind are pointed somewhere. The attention output supports a second use, nominating regions for mutagenesis — residues carrying high attention but low cross-receptor embedding similarity cluster in the near-intracellular segments of TM1 and TM5 and in ECL2, regions implicated in G-protein coupling rather than ligand contact. Groups building biosensors or bionic electronic noses around specific receptor-odorant pairs are the likely beneficiaries.

#Impact

CrossVOI's contribution is less a leaderboard position than a bounded answer to whether attention over frozen protein language model embeddings recovers biology the embeddings do not already encode. It does, partially: the cross-attention adds interaction-relevant signal in G-protein binding and putatively allosteric regions beyond the conservation information ProtT5 carries alone. The authors are careful about the limits — known key residues concentrate in the upper half of attention rankings but rarely at the very top, so attention is a region-level indicator of importance trends rather than a residue-level locator, and using it alone would produce many false positives. Scope is narrow by construction: human ORs only, one binary output at a 0.5 threshold, no hosted inference. The repository is explicitly research code, carries no license file, and ships one of the five cross-validation fold checkpoints.