MolOR

Two odorant molecules that differ by almost nothing can smell nothing alike, while a third sharing no obvious structural features with either smells the same as one of them. These Sell's triplets are the standard demonstration that fingerprints, functional-group counts and physicochemical descriptors do not map cleanly onto odor — and why a model reading structure alone has such a noisy signal to learn from.

Biology routes around the discontinuity with an intermediate representation. Odorants act on G-protein-coupled olfactory receptors (ORs), each olfactory sensory neuron expressing a single receptor type, and it is the combination of receptors an odorant activates that the brain reads as a percept. MolOR, developed at Microsoft Research with UC Berkeley and first posted in September 2024, makes that step explicit: it predicts an odorant's receptor activation profile from the molecule and the receptor's amino acid sequence, and a second stage feeds those activations into percept prediction.

Olfactory receptors have essentially no experimental structures, so MolOR borrows its structural prior from protein language modelling. Per-residue embeddings of each OR sequence are precomputed with a frozen ESM-2 650M checkpoint, atom-wise embeddings of the odorant come from a graph convolutional network trained in the paper, and a bidirectional cross-attention block learns how much to weight each residue and atom before pooling both sides into fixed vectors. The model is never told where the binding pocket is.

#Key Features

• Learned cross-attention pooling: Two cross-attention blocks produce per-atom and per-residue importance scores that act as weights when aggregating the molecule and receptor tensors, replacing the mean-pooling of earlier compound–protein interaction models.
• Frozen protein language model prior: Receptor sequences enter as precomputed ESM-2 embeddings rather than one-hot encodings, and an ablation against a randomly initialised backbone shows the pretrained weights carry most of the benefit.
• Applies to receptors it never trained on: The trained checkpoint runs over all 845 OR sequences in the HORDE catalogue of the human olfactory subgenome, and a leave-one-subfamily-out test withholding OR2J still reaches 75.05 ± 0.52 ROC-AUC on those held-out receptors.
• Receptor activations as percept features: Binarised predictions are concatenated with molecular graph features to predict 152 odor descriptors, so the percept model sees the receptor code rather than structure alone.
• Interpretable receptor–percept associations: After filtering promiscuous receptors, the predictions yield per-percept enrichments — OR5K2 activates for 74.1% of the "nutty, roasted, meaty" family against 5.2% of non-family molecules.

#Technical Details

The binding model trains on M2OR — 46,700 molecule–OR pairs covering 596 odorants and 1,237 receptor sequence variants across 11 mammalian species — as a binary activation task. The molecule encoder is a two-layer GCN with hidden size 256; receptor embeddings come from esm2_t33_650M_UR50D at 1,280 dimensions; the fused representation passes through a two-layer MLP with hidden dimension 128. Training runs 100 epochs on a single A100 with early stopping, batch size 32 and 5% dropout, under either plain binary cross-entropy or a weighted scheme correcting for screening quality, class imbalance and receptor sampling. The best configuration reaches 88.97 ± 0.30 test AUROC across five seeds, against 85.49 ± 0.93 with a randomly initialised ESM-2, 78.92 ± 0.37 for mean-pooled PLM embeddings, 85.47 ± 0.42 for PerceiverCPI and 77.73 ± 0.46 for the BiLSTM model OdoriFy; swapping in ESM-2 3B does not help (88.10 ± 0.38). The percept model trains on the merged GoodScents–Leffingwell set of 5,862 molecules and 152 descriptors, which overlaps the binding data at only 426 exact SMILES matches. Adding all 845 HORDE activations lifts percept AUROC from 87.04 ± 0.48 to 87.62 ± 0.33 (paired t-test p = 0.004; Jonckheere–Terpstra trend p = 0.010).

#Applications

The receptor stage is the part most directly useful at the bench: it scores an arbitrary odorant against an arbitrary OR sequence, which suits screening candidate ligands for the many orphan receptors with no known agonist and prioritising pairs for heterologous expression assays. Fragrance and flavour groups can use the percept stage to rank molecules by descriptor, with the receptor code as an interpretable intermediate. Because predictions run from sequence and SMILES alone, the same checkpoint extends to non-human OR repertoires.

#Impact

MolOR demonstrates that routing odor prediction through its biological intermediate beats mapping structure to percept directly, and its cross-attention pooling block is a reusable answer to compound–protein interaction problems where the binding site is unknown. The gains are honest about their size: the percept improvement is statistically robust but modest, and the authors attribute the ceiling to the binding data, which mixes noisy single-concentration screens with dose-response measurements across heterogeneous protocols and barely overlaps the perceptual datasets chemically. The model predicts detection above threshold only — it handles neither concentration and perceived intensity nor mixtures, where most real-world percepts come from. The work remains a preprint, revised in May 2026, and has not been peer reviewed.