acceptodds
Under review as a conference paper at ICLR 2027

A Unified 2D-3D Encoding and Joint Geometric Metric Framework for Olfactory Mixture Perception

Abstract

Existing olfactory perception methods often fail to capture critical three-dimensional spatial information and the distinction between causal odorants and environmental noise, limiting their ability to generalize across complex mixture scenarios. In this paper, we propose a unified molecular encoding framework that synergizes Causal Attention Learning with a joint geometric metric space to accurately map molecular features to perceptual outcomes. First, we integrate 2D topology with 3D spatial configurations, constructing a model that effectively addresses the representation deficiencies inherent in traditional topological graphs. Second, we introduce a causal aggregation module reinforced by orthogonal constraints to explicitly disentangle true odorants from background solvents, ensuring robustness in data-scarce environments. Crucially, we propose a joint hyperbolic-cosine metric space for mixture similarity prediction; the hyperbolic component naturally encodes the hierarchical generative ancestry of odorants, while the cosine component captures the holistic, permutation-invariant nature of complex mixtures. Extensive experiments demonstrate that integrating these diverse geometric priors with causal representations yields state-of-the-art performance on complex mixture benchmarks, while providing profound biological interpretability within the latent olfactory space. The code is available at https://anonymous.4open.science/r/3D-2D-Geo-metric-mix-A2CE.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Olfactory Mixture Perception, Molecular Encoding, Joint Geometric Metric Space

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_XT1qOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on predicting the similarity of olfactory mixtures. The authors propose combining 2D and 3D molecular representations to enhance model expressivity. To decouple dominant odorants from environmental noise, a multi-head causal attention mechanism with orthogonality constraints is introduced. Furthermore, the paper explores different methods for measuring similarity between mixtures. The method is evaluated on representative olfactory datasets alongside ablation studies.

Critical strengths

None

Critical weaknesses

This paper is poorly written and lacks a clear research focus. Its overall quality is low: the proposed method is a disjointed collection of components without a coherent logical flow. Several claims are factually incorrect or misleading, and the empirical evaluation fails to demonstrate any clear advantage over baselines.

Questions for authors

None

Reviewer_svJxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, authors propose a pipeline where a molecular encoder is pretrained with 2D-to-3D and 3D-to-2D score matching objectives. Mixture embeddings come from an attention module with a causal and a suppressive branch. Similarity is described as a learned weighted combination of cosine similarity and hyperbolic distance. Authors compared the proposed method with the relevant works such as POMMix and claiming to show state of art performance.

Critical strengths
  1. Adding hyperbolic distance to head is the most distinct idea provided in the paper.
  2. Table 3 gives all three metrics per fold for each head variant, and using 5-fold cross-validation rather than single test set strengthens the findings.
Critical weaknesses
  1. Table 2 results, doesn't match with the baseline paper POMMix (arXiv:2501.16271) published results for POMMix while matching the rest of the baselines (Snitz, XGBoost + (RDKit|POM)). Furthermore, the authors discarded one baseline (ChemMix+POM) which seems like have better results than authors on RMSE and Kendall metrics. Also it is very strange that POMMix on (arXiv:2501.16271) has Pearson of 0.779, but in this paper authors method exactly scored same 0.779 while reducing POMMix score to 0.727.
  2. There is no experiments for isolated contributions/ablations it is very hard to tell each components impact on the results.
Questions for authors
  1. Please see W1 and make clarifications regarding that, also at the time of writing the review the code is not accessible which makes it impossible to verify the claims.
  2. How exactly is S(M_A, M_B) computed?
  3. Does replacing Section 4.2 with mean pooling or single branch attention, or setting lambda to zero?
  4. Were Table 1's baselines re-run or copied, on which GS-LF split, as single models?
Reviewer_RMnuOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript presents a framework for predicting human olfactory perception of mixtures. It combines three components: 1. a SDE for fusing 2D molecular topology with 3D conformations; 2. multihead attention for separating primary and background components from a mixture; and 3. hyperbolic/cosine distance similarity. The authors report improvements over olfactory benchmarks, including POMMix.

Critical strengths
  • The paper addresses an important problem. Olfactory mixtures are difficult to model, and extending from single-molecule to mixtures is important.
  • The multihead attention approach (Section 4.2) appears unexplored for this problem.
Critical weaknesses
  • Section 4.1 shares substantial structure with MoleculeSDE (Liu et al. 2023, ICML), which is not cited in that section. The authors should state, equation by equation, what in Section 4.1 is adopted from Liu et al. and what is new.
  • Several sections in the paper read as a general knowledge dump from research in olfaction, and in these sections, what is a new idea vs. a common assumption that is being tested is unclear. The clearest example is Section 4.3, where the motivation for the hyperbolic-cosine metric is disconnected from what the paper demonstrates. The hyperbolic component is motivated through metabolic pathways (several prior works have explored this already) but no metabolic data is used, and the learned embeddings are not shown to inhabit hyperbolic geometry.
  • Section 4.3 and Table 3 do not specify the form of the combined similarity S(M_A, M_B), making the results in Table 3 uninterpretable. Section 4.2 points to Section 4.3 for the definition, but Section 4.3 only gives equations for the components. Furthermore, showing a combination of two signals improves on either alone is close to tautological, since the combination can be set to recover either single metric case. Substituting a different distance pair—hyperbolic with Euclidean, or cosine with Mahalanobis—is needed.
  • The pattern of improvements on the mixture benchmarks is uneven. Pearson correlation improves relative to POMMix (0.727 to 0.779), while RMSE is unchanged (0.131 to 0.131) and Kendall tau decreases (0.542 to 0.534). A method that improves perceptual ranking should move all three in the same direction, or this should be discussed and justified.
  • The paper does not include an AI Use Statement or a reproducibility statement.
Questions for authors
  • Could the authors state, equation by equation, what in Section 4.1 is novel relative to Liu et al. 2023?
  • Could the authors include (i) how d_H and sim_cos are scaled and signed before combination, and (ii) report the learned parameters this combination contains, per fold? If the learned parameters land at extreme values, or if they vary across folds, the “two complementary priors” framing in Section 4.3 is harder to defend.
  • Could the authors run the ablation with each of the two components replaced by a different distance (e.g., cosine with Mahalanobis; hyperbolic with Euclidean)? If a substituted-distance control also beats pure components, the choice of hyperbolic and cosine is not what drives the improvement.
  • Please report the learned mixing coefficients per fold.
  • Could the authors provide fold-level results for Table 2, together with a paired statistical test against POMMix per fold (see next point)? Given the severe fold-to-fold variance in Table 3, this is needed to determine whether the improvement (0.779 vs. POMMix's 0.727 Pearson) is consistent across folds and datasets or if it is driven by a subset of the data.
  • There is no significance testing for the paper's claims (Table 1, 2, and 3). Please include the following tests:
    • For Table 1, paired tests (or Friedman with Benjamini-Hochberg FDR correction) comparing the authors model against each baseline across multiple training seeds, with the number of seeds reported. Seed-to-seed variance is not reported anywhere in the paper.
    • For Table 2, paired Wilcoxon signed-rank tests of the authors model against POMMix per fold on each of Pearson, RMSE, and Kendall tau, with FDR correction.
    • For Table 3, paired tests comparing the combined metric against pure hyperbolic and against pure cosine on each of PLCC, KLCC, and RMSE across the five folds, with Benjamini-Hochberg FDR correction across the resulting six comparisons.
  • Could the authors analyze the discrepancy between the three mixture metrics in Table 2?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.