Designing Odorants for a Nose That Already Smelled

In conversation, an isolated "No" carries almost no information about whether the speaker is agreeing reluctantly, refusing angrily, or joking — its meaning exists only in relation to what came before (EAA: Emotion-Aware Audio Large Language Models with Dual Cross-Attention and Context-Aware Instruction Tuning). A decade of work on conversational emotion recognition has built exactly the machinery for resolving this kind of context dependence: recurrent state that carries history forward, dual-view architectures that separate "what was said" from "how the emotional state shifted," and models of interpersonal influence. Molecular inverse design — searching the astronomically large space of possible molecules for ones matching a target property profile — has an identical blind spot, and almost nobody has noticed. Current generative chemistry systems design each molecule against a static, fixed target code, as if the molecule were smelled in a vacuum. But smell perception is deeply sequential: receptor neurons undergo competitive antagonism and masking, so that the same molecule is perceived differently depending on which odorants preceded it (Antagonism in olfactory receptor neurons and its implications for the perception of odor mixtures).

This proposal transfers the core insight of context-dependent perception — the isolated-utterance fallacy — into chemical space exploration. I propose a generative system that designs sequences of odorants conditioned on the perceptual state the previous molecules left behind, using a dual-view architecture that decouples the static molecular representation from the predicted perceptual shift each new molecule will induce. The closest existing work, De Novo Generation of Odorant Molecules with Targeted Olfactory Receptor Activation Patterns, generates molecules matching static activation codes across a panel of 20 human olfactory receptors, achieving selectivity scores of 0.95 to 1.00 — but it evaluates every molecule as an isolated ligand and explicitly leaves mixture and temporal dynamics open. No reported result found on Lacuna for sequential or mixture-aware perceptual inverse design; that is the hole this proposal fills.

Core Claim

A molecule generator that conditions on the current perceptual (receptor-occupancy) state and predicts the perceptual shift its candidate will cause will design sequences whose perceived effect matches a target trajectory far better than a context-independent generator that re-optimizes each molecule against the same static code. The context-independent generator is not a strawman: it is the strongest published system on this exact receptor panel, and its selectivity will be excellent in the static sense while its perceived selectivity — what a nose in a sequence actually experiences — degrades, because antagonistic normalization silently rewrites the effective code of every later molecule in the sequence.

Research Program

Three components, each reusing published techniques in a new role.

First, a sequential perceptual oracle. Zhang and Guan's dual-stream binder (De Novo Generation of Odorant Molecules with Targeted Olfactory Receptor Activation Patterns) — a ProtBERT branch for receptor amino-acid sequences and a message-passing network for molecular graphs, trained on M2OR — predicts static binding across the 20-receptor panel. I will couple its predicted activation vector $a_t$ for the molecule presented at step $t$ to the antagonistic-normalization dynamics of Antagonism in olfactory receptor neurons and its implications for the perception of odor mixtures, in which competitive antagonism at the receptor level acts as a normalization that keeps the fraction of active channels roughly invariant to how many ligands are present. Concretely, the oracle maintains a decayed occupancy state and computes the perceived profile as

$$\hat{a}_t = N(a_t, s_{t-1}), \qquad s_t = \rho\, s_{t-1} + a_t,$$

where $s_t$ is the residual receptor occupancy after step $t$, $\rho \in [0,1)$ a recovery/decay constant fitted to the antagonism model's overshadowing and masking behavior, and $N$ the antagonistic normalization. This gives a cheap, fully computational simulator for perceived activation under sequential presentation — the same move Zhang and Guan made in trusting a surrogate panel, with the same risk of surrogate exploitation, which the evaluation section addresses.

Second, a state-conditioned dual-view generator. The architecture mirrors DVL-CER (Dual-View Learning for Conversational Emotion Recognition Through Context and Emotion-Shift Modeling), which decouples an Utterance-View (what is said) from an Emotion-View (how the emotional state transitions) and couples them with projection learning. Here, a Molecule-View is an RNN SMILES generator of the kind Zhang and Guan fine-tuned with reinforcement learning, producing the candidate structure; a Shift-View is a small transformer that, given the target perceptual trajectory and the current occupancy state $s_{t-1}$, predicts the distribution over perceptual transitions the next molecule should cause. Following the Interaction and Transition model's trick of maintaining separate recurrent states per speaker (Interaction and Transition Model for Speech Emotion Recognition in Dialogue), the state tracker keeps a separate channel per receptor rather than one pooled vector. The two views are coupled by projection learning as in DVL-CER, and the generator is trained with reinforcement learning against a sequence-level version of the structured combinational code loss (pairwise ranking plus distribution-shape matching), since Zhang and Guan found plain regression too noisy to optimize combinatorial profiles.

Third, tasks that make context dependence unavoidable. Two headline tasks. The hold-percept task: design a sequence of $K$ molecules whose perceived profile stays within tolerance of a fixed target code at every step, despite accumulating adaptation and masking — the perfume problem of keeping a perceived character alive as the stimulus changes. The endpoint task: given a fixed prefix sequence, design a final molecule that is perceived as target code $T$, which requires counteracting whatever masking the prefix induced. In both, the static target code alone is insufficient information, by construction.

Proposed Experiments

All experiments run on the 20-receptor M2OR panel with the sequential oracle above. The primary metric is trajectory selectivity: the mean, over steps, of the profile-match score between perceived activation $\hat{a}_t$ and the target code (using Zhang and Guan's distribution-matching formulation), plus end-state selectivity for the endpoint task and a drift measure — how far perceived profiles wander from target as $K$ grows. Every generator gets the same oracle budget of 18,000 oracle calls, matching Zhang and Guan's reported budget, which ran on a single GPU in about 15 minutes, so the entire study is computationally modest. I will report curves of perceived selectivity versus sequence length $K \in \{2, 4, 8, 16\}$; the DVL-CER finding that a fixed 4-utterance shift span was optimal in dialogue motivates testing whether an analogous shift-span parameter exists here and whether it must be learned rather than fixed. Chemical validity is handled with SELFIES encodings, following TrustMol: Trustworthy Inverse Molecular Design via Alignment with Molecular Dynamics, which reports a 100% stability rate with SELFIES versus 83% for diffusion alternatives.

A secondary experiment tests surrogate exploitation, the known failure mode in which RL generators game surrogate blind spots (De Novo Generation of Odorant Molecules with Targeted Olfactory Receptor Activation Patterns). I will check that designed sequences reproduce the psychophysical phenomena the antagonism model was built to explain — overshadowing and masking asymmetries — under perturbed model parameters, and discard candidates whose effect flips sign under small changes to $\rho$ or the normalization constants.

A third, deliberately small experiment extends to subjective perception. Human receptor repertoires vary (specific anosmias, receptor polymorphisms), so a consensus-trained designer is a "one-size-fits-all" model. Meta-PerSER showed that framing listener adaptation as few-shot meta-learning lifted macro-F1 from 24.8% to 30.7% on unseen evaluators, saturating at roughly 32 calibration samples (Meta-PerSER: Few-Shot Listener Personalized Speech Emotion Recognition via Meta-learning). I will simulate individual noses by knocking out or down-weighting receptor channels, meta-train the designer across such simulated individuals, and test whether ≤32 calibration molecules on a new individual recovers the target perceived trajectory. A positive result at that sample scale would make personalization practical, not decorative.

Optionally, perceptual grounding can be checked in language: the ODIEU benchmark (Grounding olfactory perception in language: Benchmarks and models for generating natural language odor descriptions) provides over 10,000 molecules with free-form smell descriptions and a similarity metric; CIRANO generates such descriptions from single structures. Generating the intended perceptual description of a sequence position and comparing it with the oracle-implied description offers a human-readable sanity check, though the reported neural-to-text cosine similarity of 0.36–0.43 against human–human concordance of 0.55–0.58 suggests this channel is still noisy and should not carry the main evaluation.

Baselines

The naive reference point is random screening from the M2OR ligand pool, evaluated step by step under the sequential oracle. Zhang and Guan report that screening reached 0.80 selectivity on their hardest static profile (Profile H) versus 0.95 for their RL generator, so screening is not trivially bad and calibrates task difficulty; a sequential designer that cannot beat random screening under perceived-selectivity deserves to die.

The strongest competitor is Zhang and Guan's context-independent RL SMILES generator with the structured combinational code loss, reused verbatim per step: at each position in the sequence it generates the best static match to the target code, ignoring history. It reported selectivity of 0.95 to 1.00 across nine static profiles, including 0.95 on the sparse Profile H, within 18,000 oracle calls and roughly 15 minutes on a single GPU. I will run it with the identical receptor panel, identical oracle budget, and identical tuning effort as the proposed model. Beating it means higher trajectory selectivity and end-state selectivity at matched budget, with the expected signature being that its perceived selectivity decays as $K$ grows while the context-conditioned generator's does not; a decay-versus-$K$ plot is the decisive figure.

The second real competitor is Gaussian-process Bayesian optimization over the sequence composition directly (choosing each next molecule from a large enumerated pool to maximize trajectory match). With a cheap computational oracle, standard Bayesian optimization is fully applicable and must be run, not dismissed — same simulator, same 18,000-call budget, tuned acquisition function. A third run is the key ablation: the full dual-view architecture versus a pooled-history variant in which the occupancy state is collapsed into one averaged vector fed to the generator, isolating whether the per-receptor shift view, rather than conditioning per se, does the work.

Why not just…?

Why not just design the whole mixture jointly with the antagonism model, skipping sequential modeling? For simultaneous mixtures, joint design against a mixture-aware oracle is a genuine threat, and the antagonism paper's own analysis says additive design fails there. But the tasks here are inherently temporal: the hold-percept and endpoint tasks involve stimuli separated in time, where recovery dynamics $\rho$ and residual occupancy matter and no single static mixture can express them. Moreover, joint mixture design explodes combinatorially in composition space, the same difficulty that motivated factorized synthon search in Navigating the Combinatorial Complexity of Chemical Space and the agentic formulation work in Agentic Systems for Sample-Efficient Drug Formulation Design. I will nevertheless run a joint-mixture baseline on the hold-percept task where feasible; if it matches the sequential generator there, the claim should be narrowed to temporal tasks — and I would report that honestly.

Why not just pool the history into one averaged context vector? This is the pooled-history ablation, and prior evidence predicts it loses. DVL-CER directly compared simple feature aggregation against dual-view projection learning and found aggregation inferior on IEMOCAP (70.11% versus lower weighted-F1 for pooled variants) and EmoryNLP; SynerGPT likewise found that unoptimized or random context ordering underperformed genetically optimized context by up to 2.3% ROC-AUC (SynerGPT: In-Context Learning for Personalized Drug Synergy Prediction and Drug Design). The ablation will settle it in this domain rather than by analogy.

Why not just use Gaussian-process Bayesian optimization over the sequence? It has calibrated uncertainty and convergence guarantees that LLM-style surrogates lack, a limitation BoMolLLM's own authors acknowledge (Closed-Loop Bayesian Molecular Inverse Design with Semantic LLM Surrogates). With a cheap simulator, it is a fair fight, so it is included as a baseline at matched budget. Two things may defeat it: the objective is non-stationary (the value of a molecule depends on the state it will be smelled in), which violates the standard stationarity assumption of Gaussian-process surrogates; and published evidence shows BO-style optimizers clustering in a narrow, conservative chemical family in a real formulation campaign while in-context reasoning explored diverse families (Agentic Systems for Sample-Efficient Drug Formulation Design). If BO wins at matched budget, the honest conclusion is that structured conditioning only pays when feedback is expensive or semantic, and the proposal would be trimmed accordingly.

Decisive Checks

The single decisive figure is perceived selectivity versus sequence length: if the context-independent generator's static selectivity of 0.95–1.00 holds while its perceived selectivity collapses with $K$, and the dual-view state-conditioned generator stays flat, the isolated-molecule fallacy is demonstrated in chemical space and the transfer is real. If perceived selectivity does not degrade for the static generator — that is, if antagonistic normalization is too weak at these receptor panels to matter — the core premise fails and the proposal should not proceed past the first experiment. The pooled-history ablation and the Bayesian optimization baseline then decide whether the dual-view machinery specifically, or merely any conditioning, carries the gain.