Heterogeneous Molecular Signatures of Human Odor Perception

A. Fazzio; E. V. C. Lopes; F. Crasto de Lima; G. R. Schleder; L. N. Lemos; P. Zanineli

arxiv: 2604.09758 · v1 · submitted 2026-04-10 · ❄️ cond-mat.mtrl-sci · physics.bio-ph· physics.chem-ph

Heterogeneous Molecular Signatures of Human Odor Perception

P. Zanineli , E. V. C. Lopes , G. R. Schleder , L. N. Lemos , F. Crasto de Lima , A. Fazzio This is my paper

Pith reviewed 2026-05-10 17:31 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.bio-phphysics.chem-ph

keywords odor perceptionmolecular descriptorsfeature importancemachine learningstructure-odor relationshipsfirst-principles calculationsolfactory receptorsheterogeneous signatures

0 comments

The pith

Different odors depend on distinct molecular properties rather than any single universal set of features.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines whether a common set of molecular properties governs all odor perception or whether different smells arise from separate physicochemical regimes. It trains interpretable machine-learning models on first-principles molecular descriptors that cover electronic, vibrational, and structural properties, then tracks which descriptors drive predictions for specific odor categories and receptors. No single descriptor class dominates across odors. Feature importance instead varies sharply by odor, and this pattern holds across models. The findings imply that odor perception follows receptor- and odor-specific structure-odor relationships rather than any shared encoding scheme, supplying statistical limits on theories that favor shape, vibration, or one fixed mechanism.

Core claim

Using interpretable machine-learning models trained on molecular descriptors derived from first-principles calculations that span electronic, vibrational, and structural properties, the analysis shows that no single descriptor class universally dominates odor prediction. Different odors exhibit strongly odor-specific patterns of feature importance, with substantial variability across physicochemical domains. This heterogeneity is consistent across different models, suggesting that a universal encoding scheme does not capture odor perception but reflects receptor- and odor-dependent structure-odor relationships. The results provide statistical constraints on competing olfactory theories and a

What carries the argument

Interpretable machine-learning models that extract and compare feature-importance rankings from first-principles molecular descriptors across electronic, vibrational, and structural domains for different odor categories and receptors.

If this is right

Odor space can be organized according to data-driven signatures of which molecular properties matter for each smell.
Competing theories of olfaction are limited to those consistent with receptor- and odor-dependent feature use.
Universal models that assume one encoding scheme for all odors are unlikely to succeed.
Structure-odor relationships must be treated as context-specific rather than fixed across perception.
Predictions for new molecules require accounting for the particular odor and receptor involved.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Models of olfaction may need receptor-expression profiles to improve accuracy for individuals.
Applications such as fragrance design or treatment of smell disorders could target specific feature classes per odor.
The observed heterogeneity offers a route to test how different receptor subtypes select distinct molecular cues.

Load-bearing premise

Feature-importance rankings from the machine-learning models accurately reflect the biologically relevant contributions of those molecular properties at actual olfactory receptors.

What would settle it

A dataset or biological assay in which one descriptor class predicts human odor ratings or receptor activation with comparable accuracy and dominance across many chemically diverse odorants.

Figures

Figures reproduced from arXiv: 2604.09758 by A. Fazzio, E. V. C. Lopes, F. Crasto de Lima, G. R. Schleder, L. N. Lemos, P. Zanineli.

**Figure 2.** Figure 2: FIG. 2. ( [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. ( [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. ( [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

read the original abstract

Understanding how molecular structure gives rise to odor perception remains a long-standing challenge, with ongoing debate over whether olfaction is primarily governed by molecular shape, vibrational properties, or their interplay at the level of olfactory receptors. Here, we ask whether different odors rely on common molecular determinants or instead emerge from distinct physicochemical regimes. Using interpretable machine-learning models trained on molecular descriptors derived from first-principles calculations that span electronic, vibrational, and structural properties, we analyze feature contributions for odor categories and their associated receptors. We find that no single descriptor class universally dominates odor prediction; instead, different odors exhibit strongly odor-specific patterns of feature importance, with substantial variability across physicochemical domains. This heterogeneity is consistent across different models, suggesting that a universal encoding scheme does not capture odor perception but reflects receptor- and odor-dependent structure-odor relationships. Our results provide statistical constraints on competing olfactory theories and offer a data-driven framework for organizing odor space.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper finds that odor perception draws on odor-specific mixes of electronic, vibrational, and structural molecular properties rather than any single universal class.

read the letter

The paper's central finding is that different odors rely on different mixes of molecular properties for their perception. Using machine learning on descriptors from first-principles calculations, they show that feature importance varies strongly by odor type rather than following one common scheme. This heterogeneity is the new part. Earlier work has looked at shape or vibration theories separately, but here they combine them in one framework and let the data show the variation across categories. The consistency across multiple models adds some reliability to that observation. They do a decent job with the setup. The descriptors come from quantum calculations covering electronic structure, vibrations, and geometry. That range lets them test the competing ideas in one go. Interpretable models let them pull out which properties matter for each odor. Checking that the pattern holds in different algorithms is a good check against overfitting to one method. The softer part is the biological interpretation. The results constrain the theories statistically, but the step from model feature ranks to what receptors actually use is still indirect. Feature importance can highlight correlations without proving mechanism, and the paper treats it as suggestive rather than definitive, which is fair. Data quality and labeling of odors could be a minor concern if the categories are broad or based on limited human ratings, but nothing flags a clear problem there. This paper is for people in computational chemistry or sensory science who want a data-driven look at structure-odor links. It gives a framework for organizing odor space that others could build on. I would send it to peer review. The methods are standard enough to evaluate, and the claim is clear enough to test against new data.

Referee Report

2 major / 2 minor

Summary. The paper uses interpretable machine-learning models trained on first-principles molecular descriptors (electronic, vibrational, and structural) to analyze feature contributions across odor categories. It reports that no single descriptor class universally dominates odor prediction; instead, odors exhibit strongly odor-specific patterns of feature importance with substantial variability across physicochemical domains. This heterogeneity persists across model variants, supporting the conclusion that structure-odor relationships are receptor- and odor-dependent rather than governed by a universal encoding scheme, thereby providing statistical constraints on olfactory theories.

Significance. If the central observation holds, the work offers a data-driven framework for organizing odor space and statistical constraints on competing theories of olfaction (shape vs. vibration vs. interplay). Credit is due for employing first-principles descriptors rather than empirical fingerprints and for demonstrating consistency of heterogeneity across multiple models, which strengthens the claim against a single universal scheme.

major comments (2)

[Methods] Methods: The manuscript does not provide sufficient detail on cross-validation procedures, data exclusion rules, or error bars associated with feature-importance rankings (e.g., permutation or SHAP values). These omissions make it difficult to evaluate the robustness of the reported odor-specific patterns and their consistency across models.
[Discussion] Results/Discussion: The interpretive step linking feature-importance heterogeneity to receptor-dependent biology is presented as a suggestion, but the weakest assumption—that these rankings accurately reflect biologically relevant contributions at olfactory receptors—lacks any direct comparison to receptor-binding or activation data, which is load-bearing for the claim that the patterns reflect receptor- and odor-dependent relationships.

minor comments (2)

[Figures] Figure captions and axis labels should explicitly state the number of odors or samples per category and the exact importance metric used (permutation vs. SHAP).
[Methods] Notation for descriptor classes (electronic, vibrational, structural) is introduced without a clear table summarizing their definitions and computation methods from first principles.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive review and recommendation for minor revision. We address each major comment point by point below, indicating revisions where they have been made to strengthen the manuscript.

read point-by-point responses

Referee: [Methods] Methods: The manuscript does not provide sufficient detail on cross-validation procedures, data exclusion rules, or error bars associated with feature-importance rankings (e.g., permutation or SHAP values). These omissions make it difficult to evaluate the robustness of the reported odor-specific patterns and their consistency across models.

Authors: We agree that additional methodological transparency is warranted. In the revised manuscript we have added a new subsection to Methods that specifies: (1) the stratified 5-fold cross-validation scheme with odor-category balancing to prevent leakage; (2) explicit exclusion criteria (molecules with missing first-principles descriptors or unphysical values); and (3) bootstrap (n=1000) and permutation-based error bars on SHAP and permutation-importance rankings. These additions confirm that the reported odor-specific heterogeneity remains stable across resamples and model families. revision: yes
Referee: [Discussion] Results/Discussion: The interpretive step linking feature-importance heterogeneity to receptor-dependent biology is presented as a suggestion, but the weakest assumption—that these rankings accurately reflect biologically relevant contributions at olfactory receptors—lacks any direct comparison to receptor-binding or activation data, which is load-bearing for the claim that the patterns reflect receptor- and odor-dependent relationships.

Authors: We appreciate the referee’s distinction. The original text already frames the receptor link as an interpretation rather than a mechanistic claim, noting consistency with known receptor diversity. We acknowledge that direct receptor-activation or binding data would constitute stronger validation; such comprehensive, feature-resolved datasets do not yet exist at the scale of our odor panel. Our primary contribution remains the statistical demonstration that no single descriptor class suffices across odors, thereby constraining universal-encoding theories. We have revised the Discussion to sharpen this interpretive boundary and to state explicitly that the work supplies statistical constraints rather than direct biological proof. revision: partial

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper derives its central claim—that odor perception exhibits heterogeneous, odor-specific patterns of molecular feature importance rather than a universal encoding—from standard application of interpretable ML models (permutation importance or SHAP-style attribution) to independently generated first-principles descriptors spanning electronic, vibrational, and structural properties. This heterogeneity is quantified directly from the trained models' outputs on odor-category labels and persists across model variants, without any equation or derivation reducing the reported patterns back to fitted parameters, self-defined quantities, or self-citation chains. Descriptor computation, model training, and importance extraction are described as external to the interpretive conclusion, which is framed as a data-driven statistical constraint rather than a deductive necessity. No load-bearing step collapses the result to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that ML-derived feature importances map onto biological relevance and that the chosen first-principles descriptors are sufficient to capture the relevant physics.

free parameters (2)

machine-learning hyperparameters
Standard training choices (regularization, tree depth, etc.) that affect which features are ranked as important.
feature-importance threshold
Any cutoff used to declare a descriptor class as dominant or variable across odors.

axioms (1)

domain assumption Feature importance rankings in the trained models reflect biologically causal contributions to odor perception at receptors
Invoked when the authors interpret the heterogeneous patterns as evidence against a universal encoding scheme.

pith-pipeline@v0.9.0 · 5490 in / 1355 out tokens · 67667 ms · 2026-05-10T17:31:11.663575+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · 1 internal anchor

[1]

Heterogeneous Molecular Signatures of Human Odor Perception

Heterogeneous Molecular Signatures of Human Odor Perception P. Zanineli ,1, 2 E. V. C. Lopes ,3 G. R. Schleder ,1, 2 L. N. Lemos ,3 F. Crasto de Lima ,3,∗ and A. Fazzio 3, 2,† 1Brazilian Nanotechnology National Laboratory (LNNano), Brazilian Center for Research in Energy and Materials (CNPEM), Campinas, SP, Brazil. 2Center for Natural and Human Sciences, ...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

Olfactory perception: receptors, cells, and circuits,

descriptors. These patterns reinforce the view that subsets of receptors may exhibit preferential sensitivity to distinct physicochemical regimes rather than sharing a uniform selectivity profile. Although these clusters resemble those obtained in our feature-importance analyses, it remains uncertain to what extent receptor selectivity truly aligns with t...

work page doi:10.1063/1.4905377 2024

[1] [1]

Heterogeneous Molecular Signatures of Human Odor Perception

Heterogeneous Molecular Signatures of Human Odor Perception P. Zanineli ,1, 2 E. V. C. Lopes ,3 G. R. Schleder ,1, 2 L. N. Lemos ,3 F. Crasto de Lima ,3,∗ and A. Fazzio 3, 2,† 1Brazilian Nanotechnology National Laboratory (LNNano), Brazilian Center for Research in Energy and Materials (CNPEM), Campinas, SP, Brazil. 2Center for Natural and Human Sciences, ...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[2] [2]

Olfactory perception: receptors, cells, and circuits,

descriptors. These patterns reinforce the view that subsets of receptors may exhibit preferential sensitivity to distinct physicochemical regimes rather than sharing a uniform selectivity profile. Although these clusters resemble those obtained in our feature-importance analyses, it remains uncertain to what extent receptor selectivity truly aligns with t...

work page doi:10.1063/1.4905377 2024