Most computational models of how odor mixtures smell work the same way. They embed each molecule as a vector, then average or attention-pool those vectors into a single mixture vector. The nose does something else. Molecules first compete for binding sites on a few hundred receptor types, so a weak binder can silence a strong one. A recurrent loop between the olfactory bulb and cortex is then thought to "explain away" components it has already identified. We propose a mixture-perception model that copies this order of operations. A pretrained structure-to-odor network predicts, for each molecule and each of a few hundred learned "receptors," how tightly the molecule binds and how strongly it activates. A closed-form competitive-binding layer combines these into a mixture-level receptor response. A short recurrent loop with shared weights across steps then performs inhibitory explaining-away, and the perceived distance between two mixtures is read from its output.
The opportunity matters because the mixture benchmarks are tiny. Fewer than 1,000 human-rated mixture pairs exist in total. The best models also share one diagnosed failure: they regress to the mean. They call distinct blends too similar and near-identical blends too different (High-Fidelity Tuning of Olfactory Mixture Distances…). Averaging in embedding space produces exactly this compression, because every mean of many vectors drifts toward the same centroid. Competitive binding is the mechanism the periphery uses to avoid it. It keeps the fraction of active channels stable as mixtures grow (Antagonism in olfactory receptor neurons…). Building that mechanism in as a low-parameter, structure-tied layer is a direct way to buy back accuracy without more labels.
Core Claim
We claim two things. First, aggregating molecules inside a learned receptor space with competitive binding, instead of inside an embedding space with averaging or attention, reduces regression to the mean on human mixture-distance data. Second, a few steps of learned recurrent inhibition on top of that layer give an additional, separately measurable gain that is concentrated on mixtures with many components and heavy overlap.
The model is nested, so these two claims can be tested independently. With zero recurrent steps it reduces to the competitive front end alone. With the competition replaced by a plain sum, it reduces to a linear-additive receptor model.
How It Differs From What Exists
The closest work is POMMix (Does this smell the same?). It pre-trains a graph neural network, the Principal Odor Map (POM), on single molecules and pools their embeddings with permutation-invariant attention. It reports Pearson $r = 0.779 \pm 0.028$ on pooled Snitz, Ravia and Bushdid data. Its pooling is a learned weighted average over static embeddings, with no receptor stage, no competition and no dynamics. Our proposal keeps POM as the structure encoder and replaces the pooling.
DeepNose (DeepNose) is a 3D rotation-equivariant convolutional network. It learns 96 receptor-like features and predicts mixture distance zero-shot by arithmetically averaging those features, reaching $R = 0.40$ on DREAM. That result motivates our proposal: receptor-like features plus naive averaging is weak. MoLFormer probing (Can Foundation Models Smell Like Humans?) concatenates SMILES strings, the text encodings of molecules, and finds a linear probe adds nothing. This constrains our design: more readout capacity on a fixed embedding is not the missing ingredient.
The mechanism comes from three biophysical papers. Nallamala et al. separate binding affinity from activation efficacy and show that competition normalizes mixture responses (Antagonism in olfactory receptor neurons…). The Q-space simulator shows the same effect delays collapse into "olfactory white," the uniform percept that very large mixtures converge to (A physicochemical model of odor sampling). Reddy et al. confirm in freely breathing mice that sub-linear, antagonistic interactions dominate (Antagonistic odor interactions…). None of these three learns from real chemical structures.
The recurrent stage comes from the bulb–cortex demixing circuits of A probabilistic approach to demixing odors and Demixing odors - fast inference in olfaction. Those circuits assume linear receptor input and random, fixed connectivity. We instead train the loop end to end, feeding it competitive, structure-derived inputs. The feedforward alternative of Mathis et al. argues that demixing is unnecessary for detecting a known target. Our recurrence ablation tests directly whether that holds for whole-mixture similarity, a task their argument does not cover.
Research Program
For each molecule $i$ and learned receptor $r$, a small linear head on the frozen or lightly tuned POM embedding outputs a log-affinity $\log K_{ir}$ and an efficacy $\eta_{ir} \in [0,1]$. Given component concentrations $c_i$, the mixture response of receptor $r$ follows standard single-site competitive binding:
$$a_r = \frac{\sum_i \eta_{ir}\, c_i / K_{ir}}{1 + \sum_j c_j / K_{jr}}.$$
The numerator is the summed activation contributed by each molecule. The denominator is shared occupancy, so a molecule with high affinity but low efficacy suppresses the others. The available human datasets were collected at roughly matched intensity. We therefore set $c_i$ to a constant by default. As an ablation, we set $c_i$ from the predicted ordinal strength of the odor-strength model, which uses that paper's finding that transport properties govern perceived strength.
The recurrent stage runs for $T$ steps with the same weights at every step. A "mitral" vector $m$, standing for the bulb's output neurons, is divisively inhibited by feedback from a nonnegative "cortical" vector $s$:
$$m^{t} = \frac{a}{1 + G\, s^{t-1}}, \qquad s^{t} = \mathrm{relu}\big(s^{t-1} + \alpha (F m^{t} - \lambda)\big).$$
Here $F$ is the feedforward matrix, $G$ is the nonnegative feedback matrix, $\alpha$ is a step size and $\lambda$ is a sparsity threshold, with $s^0 = 0$. The loop is unrolled and trained by backpropagation through time, which is cheap because $T \le 10$. The design is a deliberate choice about where to put the recurrence. It runs over inference iterations on a single receptor vector, not over a sequence of mixture components. The model therefore stays permutation-invariant without an ordering heuristic.
Mixture distance is a scaled cosine distance between the final $s^T$ vectors of the two mixtures. We include a learned offset for the non-zero distances humans report between identical presentations, following POMMix's head. The receptor heads, $F$, $G$ and the distance head together add tens of thousands of parameters. We tune on the same pair data as POMMix with matched budgets. Our optional pretraining step targets mechanism, not labels: fitting the front end to reproduce the Q-space simulator's antagonism statistics on random POM-embedded molecules.
Baselines
The naive reference is the cosine distance between averaged POM embeddings, which has no trainable mixture parameters. Its DREAM analog is DeepNose's feature averaging at $R = 0.40$ (DeepNose). We will recompute the POM-average version ourselves, since no reported result was found on Lacuna for averaged-POM cosine distance specifically.
The strongest existing method is POMMix (Does this smell the same?), at Pearson $r = 0.779 \pm 0.028$ on the pooled Snitz/Ravia/Bushdid set. It shares our encoder, inductive biases and data, so it isolates exactly the aggregation choice. We will rerun it on identical splits with identical POM weights, the same hyperparameter search budget and the same number of seeds. Beating it means a mean improvement in Pearson $r$ larger than the combined seed standard deviation, together with a smaller root-mean-squared error (RMSE). On DREAM's own splits, the reference is the top-six ensemble reported by High-Fidelity Tuning…: test RMSE 0.08, Pearson $r = 0.57$, validation RMSE 0.13, against a human test-retest ceiling of $r = 0.74$. Any single model approaching the ensemble on the held-out test and prospective validation sets would be notable. Closing a measurable part of the gap to 0.74 is the stretch target.
MoLFormer cosine similarity ($r = 0.64$ Snitz, $r = 0.66$ Ravia; Can Foundation Models Smell Like Humans?) serves as a label-free reference.
Three ablations isolate the key ingredients, each sharing every other component and the tuning budget. In the first, the recurrence is removed ($T = 0$). In the second, competition is removed by dropping the denominator, giving a linear receptor sum. In the third, the receptor layer is replaced by an unconstrained MLP of matched parameter count, which separates structure from capacity.
Why not just…?
Why not just put an LSTM or GRU over the list of mixture components? A sequence model has to impose an order on an unordered set, and on fewer than 1,000 pairs it would mostly learn that arbitrary order. Shuffling augmentation only partly fixes this and multiplies variance. Our recurrence runs over inference steps on an order-free receptor vector. We will still train a set-shuffled GRU aggregator as a baseline, because if it matches us, the mechanistic story is unnecessary.
Why not just add capacity to POMMix's attention, or fine-tune a foundation model end to end? The MoLFormer study found that a linear probe on frozen embeddings gave no gain, slipping to $r = 0.64$ on Ravia. This suggests the bottleneck is not readout capacity. The matched-capacity MLP ablation tests this claim on our data. If an unconstrained receptor-sized layer does as well as competitive binding, the biophysical constraint is not earning its keep and we will say so.
Why not just use the competitive front end without any recurrence? This is the real threat. Competition alone may supply all the normalization the data can detect. The model is nested so this is a measured quantity, not an argument. The prediction is that recurrence helps mainly on pairs with many components and high overlap, where explaining-away should matter, and is roughly neutral on small mixtures. If the gain is flat or absent across those subgroups, the contribution reduces to the receptor-space aggregation layer, and we will report it that way.
Decisive Checks
The primary experiment compares Pearson $r$ and RMSE against POMMix on the pooled Snitz/Ravia/Bushdid splits. It uses a molecule-disjoint split, holding out every pair that contains a held-out molecule, to rule out memorized components. On DREAM, it uses the official held-out test and 50-pair prospective validation sets.
Regression to the mean is measured directly. We fit the slope of predicted against observed distance and report error in the top and bottom distance quintiles. A slope closer to 1 than POMMix's and the ensemble's is the specific prediction that follows from replacing averaging with normalization.
We also stratify by number of components, up to DREAM's 43, and by component overlap. This tests whether the full-loop advantage grows with mixture size, and it tests the olfactory-white behavior: distances should shrink with component count more gradually than averaging predicts.
Two mechanistic checks keep the model honest. On binary mixtures, we compute the fraction of learned receptors whose response falls below the linear sum and below the strongest single component. Reddy et al. report about 80.5% and over 40% respectively in mouse glomeruli (Antagonistic odor interactions…). Learned receptors that reproduce these proportions without being trained on them would support the claim that the layer learned antagonism rather than a generic nonlinearity. On data generated by the Q-space simulator, we test whether the fitted affinities and efficacies recover the simulator's ground truth. This establishes identifiability before any claims are made about real chemistry.
References
- High-Fidelity Tuning of Olfactory Mixture Distances in the Perceptual Space of Smell Through a Community Effort
- Antagonism in olfactory receptor neurons and its implications for the perception of odor mixtures
- Does this smell the same? Learning representations of olfactory mixtures using inductive biases
- DeepNose: An Equivariant Convolutional Neural Network Predictive of Human Olfactory Percepts
- Can Foundation Models Smell Like Humans?
- A physicochemical model of odor sampling
- Antagonistic odor interactions in olfactory sensory neurons are widespread in freely breathing mice
- A probabilistic approach to demixing odors
- Demixing odors - fast inference in olfaction
- Reading Out Olfactory Receptors: Feedforward Circuits Detect Odors in Mixtures without Demixing
- Machine learning for smell: ordinal odor strength prediction of molecular perfumery components