acceptodds
Under review as a conference paper at ICLR 2027

SCENT: Aligning Mass Spectra with Molecular Structure for Olfactory Perception

Abstract

Predicting human olfactory perception from molecular structure has seen remarkable progress in recent years. However, existing approaches typically require access to the explicit chemical structure of an unknown substance at inference time, which is not directly observed. We address this gap by exploring electron ionization mass spectrometry (EI-MS) as a direct input modality for olfactory prediction, leveraging chemically informative fragmentation patterns acquired within seconds. We introduce Spectrum-to-Chemical Embedding alignmeNT (SCENT), a multi-modal contrastive learning framework that aligns EI-MS representations with pretrained chemical structure embeddings while requiring only mass spectra at inference, focusing here on single-compound odorants. On the multi-label odor descriptor prediction task, SCENT outperforms MS-only baselines and approaches the performance of structure-based models without access to explicit molecular structure at test time. On a prospective perceptual panel of structurally or perceptually distinct odorants, SCENT maintains its advantage over MS-only baselines. Its learned representations also provide more informative features for predicting continuous human perceptual ratings and generalize to independently acquired laboratory spectra. Together, these results suggest that cross-modal alignment can effectively transfer structure-derived chemical information into EI-MS representations for olfactory prediction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Representation Learning, Olfactory Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_quLwOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper is about predicting odor perception from EI-MS spectra. The main idea is a form of contrastive distillation: an MS encoder (that is pretrained on the structure/odor information) is trained with a CLIP-like loss to align with frozen structure embeddings (Open-POM or MolFormer) on ~148k NIST spectrum–SMILES pairs, so that only the spectrum is needed at test time. Experiments on GS-LF, a prospective panel, DREAM and 50 lab-measured spectra show consistent gains over EIMS2Vec, and the authors claim the model approaches structure-based models without access to structure at inference.

Critical strengths

1 Olfaction is an important but underexplored problem in ML, and work that pushes this direction forward is valuable to the community. 2 Framing the problem as contrastive distillation from structure to spectra is a neat idea. The authors note that the contrastive objective itself is not new, but applying it between EI-MS and molecular structure makes a lot of sense. Odor modeling has long lacked a good, shared representation, and this approach makes use of ~148k spectrum–structure pairs, while labeled odor data covers only a few thousand molecules. 3 The experiments are extensive, covering four evaluation settings (GS-LF, a prospective panel, DREAM and real lab spectra) along with ablations and significance tests. The results consistently show that the learned representation is useful for odor prediction.

Critical weaknesses

My major concerns:

  1. The paper motivates spectrum-only prediction by saying molecular identity is unavailable at inference. But for conventional EI-MS, library search is the standard way to get identity and it is fast. The paper’s own data seemingly shows high library coverage of odorants (385/476 DREAM molecules, 178/208 panel molecules). The obvious baseline is therefore search + Open-POM. The only retrieval comparison (App. E.2) is open-set with top-1 retrieval, which is the worst case for search. I would like to see closed-set retrieval, at least on the 50 lab spectra searched against NIST, plus a top-k / kNN variant, with results split by match score. Without this it is unclear when SCENT is actually useful.

  2. This is closely related to point 1: single pure compounds are exactly the setting where library search works best. The cases where search fails (mixtures, background interference, compounds not in the library) are not evaluated. Hasebe et al. (2022), which the paper cites, already works with mixture spectra, but this is not discussed. Even a simple test, such as adding a small fraction of a second compound to a spectrum and measuring how the predictions change, would help. Also, a natural competing pipeline that also uses only spectra at inference is MS → fingerprint → odor model. Ji et al. (2020), cited in the introduction, predicts fingerprints from EI-MS, but it is not used as a baseline or discussed thoroughly.

  3. Open-POM is trained with odor labels and is a much stronger teacher than MolFormer on DREAM (Table F2: 0.37–0.42 vs. 0.28–0.34). Yet its student is weaker (Table F3: 0.301 vs. 0.325), and on GS-LF the two students are on par. A stronger, odor-aware teacher giving a weaker student is surprising and deserves analysis.

Minors:

Gains over EIMS2Vec are moderate (+1.8–2.0 micro-AUC, +3.2–3.4 macro-AUC on GS-LF), and SCENT (Open-POM) stays below Open-POM on GS-LF and DREAM. Since both the teacher and the downstream encoder are frozen, have the authors tried fine-tuning the encoder on odor labels after alignment, or training the alignment and odor losses jointly?

The masked reconstruction baseline (76.34 micro-AUC) is far below even EIMS2Vec. Please give its details: masking ratio, reconstruction target, and pooling. If only [CLS] is used downstream, it may receive no training signal under a masked objective.

Related work. CLIP-style alignment has also been used to learn olfactory representations from sensor data, e.g., aligning e-nose signals with images in the wild (New York Smells, Ozguroglu et al., 2025), and similarly on language-guided olfactory representation learning (Tsonis et al., 2026). A short discussion of this line of work would help position the contribution.

Questions for authors

See my major concerns. I am currently slightly leaning towards borderline rejection. I would re-evaluate the rating if the authors could address my concerns.

3 comments
AuthorsOpenReview ↗

Author's Response to Reviewer quLW (1/2)

We thank the reviewer for the time spent reviewing our paper. Below, we address the reviewers questions and concerns:


C1: When is SCENT preferable to library search followed by Open-POM? How does it compare with closed-set retrieval and top-k/kNN variants on the 50 laboratory spectra, including performance across different match-score ranges?

A1: We agree that library search is the standard first step, and we have run the requested comparison on the 50 laboratory spectra. Retrieval uses the official NIST MSPepSearch (Identity search, NIST MS Search 3.0 algorithm, pre-search off, m/z 50–180) against 148,212 NIST23 spectra. The top-k retrieved molecules (k = 1, 3, 5, 10) are passed through Open-POM to an odor head trained on the same split as SCENT's head, so neither head has seen the odor labels of these molecules.

For a closed-set (every query in the library) setting, NIST identifies 36/50 measured spectra correctly. Library search → Open-POM performs on par with SCENT: top-1 Macro-AUC is 83.05 ± 0.40 vs. 82.09 ± 0.61 (Δ +1.0 [−1.8, +3.1]), and top-3/5/10 give 82.32–83.41. No match-factor tertile is significantly above SCENT. For an open-set (query removed from the library)​​ setting,top-1 drops to 79.05 Macro-AUC (Δ −3.0 [−7.3, −0.5]), and every top-k variant trails SCENT on Adj.P@5 by 6.9-12.1 points.

We therefore treat SCENT as a useful complement to library search, especially when the query is not in the library. We added these results to the Appendix E.2 and modified the discussion in Section 4.1 in the updated paper.


C2-1: How robust is SCENT when spectra contain mixtures or background interference, or when compounds are absent from the library? Can a controlled second-compound contamination experiment clarify its robustness?

A2-1: We thank the reviewer for raising this point. To answer the reviewer question, we have conducted an experiment where each GS-LF test spectrum was mixed with one spectrum from the training/validation pool, chosen with low (spectra cosine < 0.05) or high (spectra 0.6–0.9) spectral overlap. Both spectra TIC(x)=mxm, giving a=a/TIC(a) and c=c/TIC(c), and mixed as x=a+c. Hence SNR=1010[TIC(a)/TIC(c)]=-1010, swept from 40 dB (=10-4) to 0 dB (=1, equal ion current) in 2 dB steps. We test the robustness of both SCENT variants and EIMS2Vec. The results show that the performance degrades gracefully as the contaminant grows. SCENT stays above EIMS2Vec down to about 15 dB (contaminant ≈ 3% of the target's ion current), the margin closes only below about 10 dB (contaminant ≥ 10%), and a spectrally similar contaminant costs less than a dissimilar one. We added those results with a figure to Appendix E in the updated version paper.


C2-2: How does SCENT compare with an MS → fingerprint → odor prediction pipeline? How does it relate to Ji et al. (2020)

A2-2: Thank you for the question. Ji et al. (2020) predict a binary molecular fingerprint from an EI spectrum and use it to rank candidate structures for compound identification. The difference between this pipeline with SCENT is mainly in this intermediate representation. As Lee et al. (2023) demonstrated that POM captures odor-relevant information more effectively than conventional molecular fingerprints, we chose to focus on this representation rather than explicitly predicting fingerprints. We discussed the differences between our work and the work of Ji et al. (2020) in Section 2 of the updated paper.


C3: Why does the stronger, odor-supervised Open-POM teacher produce a weaker student than MolFormer on DREAM, while their students perform similarly on GS-LF

A3: We thank the reviewer for this question. The student primarily inherits its teacher's dominant, high-variance principal components (R2 0.65 vs. 0.33 for tail components). While Open-POM relies on its low-variance tail directions to predict out-of-vocabulary DREAM descriptors (such as Acid and Chemical), its student fails to recover these weak signals during alignment, whereas MolFormer's broader variance distribution allows its student to capture them. Consequently, the two students tie on GS-LF descriptors (which rely on dominant components), while SCENT (Open-POM) lags specifically on out-of-vocabulary DREAM terms.

AuthorsOpenReview ↗

Author's Response to Reviewer quLw (2/2)

C4: Since both the teacher and the downstream encoder are frozen, have the authors tried fine-tuning the encoder on odor labels after alignment, or training the alignment and odor losses jointly?

A4: Compared with structure-based models, MS-based models have less direct access to structure information. Our goal is therefore to transfer this information to the MS encoder. For this reason, we choose to keep the teacher model frozen. We also tested the fine tuning during the downstream task, but the model performance didn’t benefit from it. We hypothesize that this may be due to the limited amount of downstream training data relative to the parameters in the encoder, which prevents the data from effectively refining the upstream representations. We also observed that a randomly initialized classification head can disrupt aligned features during the early stages of training, while also increasing the total training time. Consequently, we opted to freeze the encoder and train only the classification head. We thank the reviewer for raising this question and we clarified our rationale in the Appendix A of the updated paper.


C5: The masked reconstruction baseline (76.34 micro-AUC) is far below even EIMS2Vec. Please give its details: masking ratio, reconstruction target, and pooling. If only [CLS] is used downstream, it may receive no training signal under a masked objective.

A5: Thank you for pointing out this question. We checked the ablation and confirmed that the original MR baseline masked 15% of observed peaks (Bernoulli per peak, replaced by a zero vector), regressed their relative intensities with a per-token decoder, and used the frozen [CLS] output downstream. We therefore re-implemented it two ways with SCENT's architecture and data. The ablation uses BERT-style masking (mask token plus m/z, per-token decoder) and a variant predicting masked intensities from [CLS], training it directly. We report both [CLS] and mean pooling. The results show that the mean pooling is better than [CLS] in this setting, since the [CLS] doesn’t have a task of summarizing the full molecule information. The ablation also proves that the gain of SCENT comes from alignment, not simply from the architecture. We updated Table 2 in the updated version of paper with these results.

Model VariantTransformerObjectiveMicro-AUC↑Macro-AUC↑Adj.Precision@k↑
w/o CLIP 1a, [CLS]✓MR78.06 ± 0.0858.53 ± 0.4129.65 ± 0.10
w/o CLIP 1b, mean pool✓MR86.10 ± 0.1078.68 ± 0.1940.94 ± 0.35
EIMS2Vec×PC87.98 ± 0.1582.29 ± 0.6146.49 ± 0.53
SCENT (Open-POM)✓C89.80 ±0.0985.45 ±0.3050.60 ±0.53
SCENT (MolFormer)✓C89.99 ±0.1885.71 ±0.2650.62 ±0.75

C6: CLIP-style alignment has also been used to learn olfactory representations from sensor data, e.g., aligning e-nose signals with images in the wild (New York Smells, Ozguroglu et al., 2025), and similarly on language-guided olfactory representation learning (Tsonis et al., 2026). A short discussion of this line of work would help position the contribution.

A6: Thank you for highlighting these related studies. We agree that these works help position SCENT within the broader context of multimodal olfactory representation learning. Existing sensor / language guided alignment shows the potential of CLIP-style contrastive learning in bridging olfactory sensory modality. SCENT, although applying a similar approach, grounds the representation in chemistry by mapping the smell signal with exact chemical structure information, and offers a complementary perspective to the community. We added this discussion in Section 2 of our updated paper.

I thank the authors for the additional information and experiments provided in the rebuttal.

Most of my concerns have been addressed. The one that remains is the in-library search issue, which I consider important: the authors should better position their work with respect to the search for in-library compounds. Otherwise, I like the paper. I believe olfaction deserves more attention from the community, as it will matter a great deal for the future of physical AI.

I am therefore raising my score to Weak Accept and this will be the final score.

Reviewer_azBHOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce SCENT, a contrastive learning framework that trains an electron-ionization mass spectrometry (EI-MS) encoder to match molecular structure embeddings for ~150K paired examples. They use their model to predict olfactory properties from mass spectra alone (enabling the bypassing of molecular identification). The authors evaluate on four tasks: multi-label odor descriptors, human perceptual regression, a 178-molecule prospective panel, and transfer to independently acquired lab-measured spectra.

Critical strengths
  • The introduction clearly sets up both the problem and the solution. This is an important and frontier problem in scientific ML for chemical and materials science.
  • The four evaluations are complementary and well-chosen.
  • The ablation study (Table 2) is thorough.
  • The authors are honest about where the method's advantage ends. SCENT is statistically indistinguishable from MolFormer (the expected ceiling given the alignment target) and the paper states this plainly.
  • The statistics are solid.
  • Section 4.3 is a rare and valuable form of evaluation in the ML literature. Testing on spectra measured in a separate lab with different instrumentation is a real-world test that most ML papers don’t (or can’t) perform.
Critical weaknesses
  • Innovation along the pure ML methodology angle is minimal (and the authors note this). I simply note this but don’t perceive this as a weakness in terms of score.
Questions for authors

The following are scientific and clarity questions, they are not demands for additional experiments or a precondition for acceptance.

  • What does the aligned MS encoder learn relative to the structure target? A CKA or linear-mapping analysis between the aligned MS embedding and the structure embedding would distinguish if SCENT reconstructs the structure embedding or extracts complementary spectrum-specific information.
  • Can the framework extend to mixtures? I encourage the authors to discuss this possibility.
  • Why freeze the structural encoder rather than co-train it with the MS encoder? Does co-training help either representation?
  • The term "MS-only baseline(s)" (also "unaligned MS baselines") is used in the abstract, introduction, and body, but it's not always clear what model that is referring to. Table 1 implies both EIMS2Vec and the autoencoder, Section 4.2 refers to the "unaligned EIMS2Vec baseline", and Table 3 reports only EIMS2Vec. It is therefore not clear in each case whether both models or only EIMS2Vec is being referred to. Or at other points I thought the authors were referring to the pretrained model before contrastive alignment. Could the authors name the specific baseline(s) at each mention?
  • In Section 4.1, could the authors frame Table 1 against the structure-based ceiling more explicitly? Open-POM represents what is achievable when molecular identity is available and SCENT approaches but remains below this ceiling. An explicit statement that SCENT closes most of the gap to the structure-based ceiling without needing molecular identification would make the contribution clearer to those not following closely.
  • Could the authors give a concrete example of where this method would be used — a research lab doing rapid odorant characterization, an industrial quality-control setting, or a field deployment with portable mass spectrometry — to anchor the method to a realistic use case?
3 comments
AuthorsOpenReview ↗

Author's Response to Reviewer azBH (1/2)

Thank you for the careful review. Below we respond to your questions and concerns.


Q1: What does the aligned MS encoder learn relative to the structure target? A CKA or linear-mapping analysis between the aligned MS embedding and the structure embedding would distinguish if SCENT reconstructs the structure embedding or extracts complementary spectrum-specific information.

A1: We compared SCENT and EIMS2Vec with the structure targets on the GS-LF test molecules held out of alignment, using debiased linear CKA. Alignment moves SCENT toward Open-POM (0.53 vs. 0.21 for EIMS2Vec). A ridge map recovers 61% of Open-POM's variance from SCENT (40% before alignment). Under a linear probe, concatenating SCENT to Open-POM raises test Macro-AUC from 85.93 to 87.25 (+1.32 ± 0.22), while adding 500 row-shuffled SCENT features gives no gain. Alignment therefore preserves this spectral information while adding the structural part. We added these experiments to the Appendix E of the updated paper.

EIMS2VecSCENT (Open-POM)SCENT (MolFormer)Open-POM
SCENT (Open-POM)0.468 ± 0.016
SCENT (MolFormer)0.454 ± 0.0150.869 ± 0.003
Open-POM0.205 ± 0.0120.528 ± 0.0080.509 ± 0.010
MolFormer0.177 ± 0.0160.358 ± 0.0100.411 ± 0.0100.287 ± 0.010

Q2: Can the framework extend to mixtures? I encourage the authors to discuss this possibility.

A2: Thank you for highlighting this direction. To explore transfer beyond single compounds, we conducted a pilot evaluation on measured mixture spectra from 40 University of Pennsylvania Smell Identification Test (UPSIT) odor samples. For 24 samples, the odor name directly matches one of our 138 output descriptors. For the remaining 16, an LLM (GPT 5.6 Sol high) mapped each name to three descriptors from this vocabulary. These are semantic reference labels, not independently measured perceptual annotations. Hit@k measures whether at least one reference label appears among the top-k predictions; Recall@k measures the fraction of reference labels recovered. Results are reported by percentage number.

Objects whose name is a label (24/40):

Hit rateEIMS2VecSCENT (MOL)SCENT (POM)Random
Top@10.00.00.00.7
Top@50.012.512.53.6
Top@108.316.716.77.2
Top@2016.737.545.814.5

Objects with LLM-mapped labels (16/40):

ModelHit@1Hit@5Hit@10Hit@20
SCENT (POM)18.843.856.268.8
SCENT (MOL)25.050.056.268.8
EIMS2Vec6.218.831.250.0
Random2.210.620.337.7

These preliminary results suggest that chemical alignment improves descriptor ranking for measured mixture spectra. Given the small sample size and name-derived reference labels, we interpret this pilot as exploratory evidence rather than a validated evaluation of mixture perception. As discussed in the Limitations section, further validation requires paired mixture spectra and perceptual annotations.


Q3: Why freeze the structural encoder rather than co-train it with the MS encoder? Does co-training help either representation?

A3: We freeze the structural encoder because our goal is to transfer its pretrained chemical representation to the MS encoder while preserving a stable alignment target and a fixed reference for comparison. We have not evaluated joint training in this study, so whether it improves either representation remains an open question. We agree that this is a worthwhile extension and have clarified the rationale for freezing the structural encoder in Appendix A.


Q4: The term "MS-only baseline(s)" (also "unaligned MS baselines") is used in the abstract, introduction, and body, but it's not always clear what model that is referring to. Table 1 implies both EIMS2Vec and the autoencoder, Section 4.2 refers to the "unaligned EIMS2Vec baseline", and Table 3 reports only EIMS2Vec. It is therefore not clear in each case whether both models or only EIMS2Vec is being referred to. Or at other points I thought the authors were referring to the pretrained model before contrastive alignment. Could the authors name the specific baseline(s) at each mention?

A4: Thank you for pointing out this. In Table 1, ‘MS-only baseline’ refers to both EIMS2Vec and the autoencoder. In the prospective panel, DREAM, and laboratory-spectrum, the baseline is EIMS2Vec alone. These references do not denote the SCENT encoder (EIMS2Vec + Transformer) before contrastive alignment. We have clarified those expressions in our updated version paper.

AuthorsOpenReview ↗

Author's Response to Reviewer azBH (2/2)

Q5: In Section 4.1, could the authors frame Table 1 against the structure-based ceiling more explicitly? Open-POM represents what is achievable when molecular identity is available and SCENT approaches but remains below this ceiling. An explicit statement that SCENT closes most of the gap to the structure-based ceiling without needing molecular identification would make the contribution clearer to those not following closely.

A5: Thank you for this suggestion. We will add this comparison explicitly to Section 4.1 our updated version paper. ‘Without access to structural information at inference, SCENT (Open-POM) closes 72% of the Macro-AUC gap between EIMS2Vec and Open-POM (93% of the Micro-AUC gap), and SCENT (MolFormer) exceeds MolFormer on all three metrics.’


Q6: Could the authors give a concrete example of where this method would be used — a research lab doing rapid odorant characterization, an industrial quality-control setting, or a field deployment with portable mass spectrometry — to anchor the method to a realistic use case?

A6: Thank you for this suggestion. A concrete use case is preliminary odor characterization of an unidentified single-compound sample in a research laboratory. Given its EI-MS spectrum, SCENT predicts descriptive odor terms without first requiring explicit molecular identification. These predictions translate an instrumental measurement into an interpretable odor profile, allowing researchers to communicate an approximate ‘concept’ of the sample’s smell even before its molecular identity has been established.

Thanks for you through reply. This address all of my concerns, and my score was already strong. My only remaining comment is regarding Q6. I would suggest the authors include this in the paper itself -- most ML papers sorely lack a demonstration of their real world need, and this is a nice case.

Reviewer_cWqUOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper is presenting a work on predicting odors from electron ionization mass spectrometry (EI-MS). Structure-based odor models need the molecule’s structure at inference but SCENT instead uses structure only at training, via CLIP-style alignment between an MS encoder and frozen pretrained structure encoders (Open-POM, MolFormer). At inference it predicts only based on the spectra. It is evaluated on GS-LF descriptor prediction, DREAM rating regression, a prospective panel, and 50 real-world lab spectra.

Critical strengths

The problem and the idea is novel and very needed. Olfaction is a very underexplored modality in general. Predicting odor from a fast measurement, with structure needed only at training, is a useful new framing. The framework is very well described and easy to follow.

The retrieval baseline was a nice addition and the natural competitor in this kind of comparison.

The 50-spectrum real-world dataset is very useful (although the size is quite small).

The chemical-novelty analysis (Appendix E.4, Table E2) directly tests whether the gains come from close analogues of the training molecules, and shows the advantage over EIMS2Vec is largest on the least similar molecules. It deserves to be in the main text.

Critical weaknesses

The contribution is evaluation/application, not a new method; an EIMS2Vec-style MS encoder, a standard Transformer encoder, a standard CLIP/InfoNCE loss, and frozen MolFormer/OpenPOM embeddings. Every component is off-the-shelf; the value is the problem and the validation.

Under GS-LF’s heavy class imbalance, the reported AUROC (especially micro) might be optimistic; AUPRC / mean average precision (and macro-F1), reported by comparable GS-LF work, are missing.

The main GS-LF table and the ablation report mean ± std but no significance tests. Many per-label / per-descriptor Wilcoxon tests at α=0.05, with no multiple-comparison correction reported.

The 2-layer/4-head transformer is justified only as "empirically found".

Alignment and the transformer seem to work together, not separately. Alignment alone adds little over EIMS2Vec (about +0.4 / +0.8 / +0.9), while the full model adds about +1.8 / +3.2 / +4.1. The claim that alignment helps on its own rests on that small, untested gap. A transformer trained with EIMS2Vec’s objective and no alignment is missing as a control.

The DREAM evaluation is on a reduced, non-standard set (476 molecules filtered to the 385 with valid library EI-MS spectra, with ratings averaged across panelists and replicates), which makes the reported correlations hard to compare against previously published DREAM results.

Questions for authors
  1. Given the class imbalance, can you add AUPRC / mean average precision and a macro-averaged threshold metric?

  2. Can you add paired significance tests to Tables 1 and 2, especially for alignment without the Transformer vs EIMS2Vec? Can you add a Transformer trained with EIMS2Vec’s objective and no alignment?

  3. Can you correct for multiple comparisons? Also, “statistically indistinguishable” on DREAM rests on a non-significant test, which does not show equivalence. Please use an equivalence test (e.g. TOST) or say “no significant difference detected".

  4. Is the retrained Open-POM a single model or an ensemble? Which MolFormer checkpoint is used, and how is its embedding pooled?

  5. Appendix E.4 calls Q4 the most dissimilar subset, but Table E2 gives Q4 a Tanimoto of 0.90 and Q1 0.41. Which subset do p = 0.006 and 0.009 refer to? Can you add a trend test across strata, with a bootstrap over molecules?

  6. The 2-layer, 4-head Transformer is “empirically found.” Can you report the search and confirm it was chosen on validation data?

  7. The provided link for the code is expired. Could you please provide an anonymized link that works?

3 comments
AuthorsOpenReview ↗

Author‘s Response to Reviewer cWqU (1/2)

We thank the reviewer for the detailed and constructive feedback and for the careful consideration of our work. Below, we answer the questions you raised.


W2 & Q1: Given the class imbalance, can you add AUPRC / mean average precision and a macro-averaged threshold metric?

A1: Thank you for raising this. We now report Macro-AUPRC, mAP and Macro-F1 as a supplement to Table 1. Macro-F1 uses one threshold per fold, chosen on that fold's validation set. Because the labels are 4% positives on average, the chance levels of these metrics are low. Relative to these references, both SCENT variants perform on par with the structure-based models from the spectrum alone and improve over EIMS2Vec on all three metrics, consistent with the ROC-AUC results. We have updated these results to Appendix E in our updated version paper.

ModelMacro-AUPRCmAPMacro-F1
Open-POM32.78 ± 0.4448.55 ± 0.2426.43 ± 0.48
MolFormer33.42 ± 0.6547.36 ± 0.6729.24 ± 0.58
EIMS2Vec28.26 ± 0.4543.16 ± 0.3925.02 ± 0.49
SCENT (Open-POM)33.16 ± 0.6447.33 ± 0.2828.73 ± 0.86
SCENT (MolFormer)32.87 ± 0.8447.22 ± 0.3729.51 ± 0.92
Random baseline5.787.127.13 ± 0.03
Frequency prior4.0826.097.40 ± 0.02

W5 & Q2: Can you add paired significance tests to Tables 1 and 2, especially for alignment without the Transformer vs EIMS2Vec? Can you add a Transformer trained with EIMS2Vec’s objective and no alignment?

A2: Thank you for the suggestion. We trained SCENT's Transformer with EIMS2Vec's negative-sampling objective and no alignment: the [CLS] output serves as the spectrum context and must score the spectrum's own peaks above sampled negatives. The frozen encoder is evaluated with SCENT's stage-2 protocol. Because this objective trains [CLS] on a peak-level target rather than on the whole spectrum, we report the stronger readout, mean pooling over the token outputs (the [CLS] readout reaches a Macro-AUC of 62.07). The results show that the gain of SCENT therefore comes from alignment, and the Transformer adds to it once the encoder is aligned.

ModelMicro-AUCMacro-AUCAdj.Precision@k
Transformer, EIMS2Vec objective, no alignment (mean pool)86.09 ± 0.1178.69 ± 0.2941.61 ± 0.74
EIMS2Vec87.98 ± 0.1582.29 ± 0.6146.49 ± 0.53
SCENT (Open-POM)89.80 ± 0.0985.45 ± 0.3050.60 ± 0.53
SCENT (MolFormer)89.99 ± 0.1885.71 ± 0.2650.62 ± 0.75

Here we add the significance test Holm correction table.

Model VariantTransformerObjectiveMicro-AUCMacro-AUCAdj.Precision@k
w/o CLIP 1✓MR86.10 ± 0.10 * †78.68 ± 0.19 * †40.94 ± 0.35 * †
w/o CLIP 2✓S86.16 ± 0.54 * †78.98 ± 0.62 * †41.48 ± 0.86 * †
w/o CLIP 3✓PC86.09 ± 0.11 * †78.69 ± 0.29 * †41.61 ± 0.74 * †
w/o Transformer (Open-POM)×C88.39 ± 0.12 * †83.04 ± 0.60 * †47.40 ± 0.59 †
w/o Transformer (MolFormer)×C88.40 ± 0.08 * †83.15 ± 0.34 * †47.42 ± 0.59 †
Regression (Open-POM)✓MSE88.25 ± 0.10 †83.24 ± 0.25 * †46.28 ± 0.73 †
Regression (MolFormer)✓MSE87.69 ± 0.08 †82.06 ± 0.16 †44.94 ± 0.38 * †
EIMS2Vec×PC87.98 ± 0.15 †82.29 ± 0.61 †46.49 ± 0.53 †
SCENT (Open-POM)✓C89.80 ± 0.09 *85.45 ± 0.30 *50.60 ± 0.53 *
SCENT (MolFormer)✓C89.99 ± 0.18 *85.71 ± 0.26 *50.62 ± 0.75 *

Mean ± Std over the 5 folds. Paired t-test over the 5 shared CV folds, p-values Holm-corrected within each comparison family. * = p < 0.05 vs. EIMS2Vec, † = same thresholds vs. SCENT with the same teacher (for EIMS2Vec and w/o CLIP 2, the more conservative p of the two SCENT teachers).

We added those results to Table 2 in the updated version paper.


Q3: Can you correct for multiple comparisons? Also, “statistically indistinguishable” on DREAM rests on a non-significant test, which does not show equivalence. Please use an equivalence test (e.g. TOST) or say “no significant difference detected".

A3: Thanks for mentioning this. For Figure 3 and Figure E1, we now control the family-wise error rate with the Holm–Bonferroni procedure within each pre-specified family of tests. SCENT achieves a higher mean AUC than EIMS2Vec in every frequency bin. After Holm correction, the difference is significant in the 20–35 and ≥100 bins for both variants, and additionally in the 35–50 bin for SCENT (MolFormer). The largest gain is in the 20–35 bin (+5.9 and +6.0 AUC points). We will change the Figure 3 and E1 as well as the captions in the updated version paper. We have said ‘no significant difference detected’ for other cases this might apply in the updated version paper.

AuthorsOpenReview ↗

Author‘s Response to Reviewer cWqU (2/2)

Q4: Is the retrained Open-POM a single model or an ensemble? Which MolFormer checkpoint is used, and how is its embedding pooled?

A4: The Open-POM in our paper is a single model. We used the MolFormer called ‘N-Step-Checkpoint_3_30000.ckpt’ published by MolFormer paper. We take the last-layer representation of the first token. The first token is the classification token defined by the MolFormer tokenizer. After the embedding is extracted, they are frozen in our paper’s setting.We added those details of setting to the Appendix A in the updated version paper.


Q5: Appendix E.4 calls Q4 the most dissimilar subset, but Table E2 gives Q4 a Tanimoto of 0.90 and Q1 0.41. Which subset do p = 0.006 and 0.009 refer to? Can you add a trend test across strata, with a bootstrap over molecules?

A5: Thank you for pointing this out. We agree that the Q4 is the most similar subset while Q1 is the most dissimilar subset. We will fix Appendix E.4 in the updated version paper. The p-values 0.006 and 0.009 refer to the last row of Table E2: the 30 test molecules with NN Tanimoto < 0.4, a subset of Q1. On this subset, SCENT (Open-POM) and SCENT (MolFormer) improve the mean per-molecule ROC-AUC over EIMS2Vec by 4.63 and 6.07 percentage points, respectively, with two-sided paired Wilcoxon p-values of 0.006 and 0.009.

We also add a trend test across strata which bootstrap over the molecules. For each molecule, we compute the paired gap to EIMS2Vec and estimate the linear trend of the quartile-mean gap from Q1 to Q4. Confidence intervals come from 10,000 bootstrap replicates that resample molecules within strata. SCENT’s performance advantage over EIMS2Vec is largest on the most novel molecules and shrinks toward familiar ones. This confirms that structure alignment provides the benefit where training similarity is lowest. We added these results to the Appendix E.4 of the updated version paper.

Gap vs EIMS2VecΔ by quartile (Q1→Q4)Slope per stratum [95% CI]Bootstrap pPermutation pQ1–Q4 Δ [95% CI]
SCENT (Open-POM)3.19 / 2.56 / 1.93 / 0.56−0.85 [−1.48, −0.25]0.0060.011+2.63 [+0.73, +4.60]
SCENT (MolFormer)4.19 / 2.43 / 2.45 / 0.75−1.03 [−1.74, −0.39]0.0010.002+3.44 [+1.38, +5.69]

Q6: The 2-layer, 4-head Transformer is “empirically found.” Can you report the search and confirm it was chosen on validation data?

A6: We add a supplementary for the Transformer configuration. We noticed that a 2-layer structure achieves the best trade-off between performance gains and model parameter scale. Additionally, as the number of attention heads does not significantly affect alignment validation AUC, we opted for a standard 4-head configuration commonly used in practice. We added these results with parameter search figures to the Appendix C of the updated version paper.


Q7: The provided link for the code is expired. Could you please provide an anonymised link that works?

A7: Thank you for your reminder. The link is now updated.


W6: The DREAM evaluation is on a reduced, non-standard set (476 molecules filtered to the 385 with valid library EI-MS spectra, with ratings averaged across panelists and replicates), which makes the reported correlations hard to compare against previously published DREAM results.

A8: We thank the reviewer for this point. On DREAM, for this response we re-ran the analysis end to end, extracting every spectrum embedding with exactly the input pipeline and checkpoints used for the GS-LF experiments. This updates the spectrum-based rows. We have updated those results to Section 4.2 and Appendix F. To make DREAM results comparable with published work, we ran the published Keller et al. (2017) on our subset(the same 385 molecules, highest available concentration) and 100 train/test splits. The forest uses 5,473 features computed from the known structure (Dragon molecular descriptors and Morgan fingerprints), with the original code and settings. The dilution input is omitted so that every model sees only the molecule. SCENT is instead evaluated as a frozen representation that is not trained on DREAM, with only a linear probe fitted to the ratings. Without the structure at inference, SCENT(Open-POM) recovers 58% of the gap between the unaligned spectrum encoder (EIMS2Vec) and its structure teacher Open-POM, showing that structural odor information transfers from the spectrum to a different dataset with the one in training.

ModelInputReadoutMean r
Keller et al. (2017)5,473 featuresrandom forest0.478 ± 0.038
Open-POMstructurelinear probe0.468 ± 0.039
SCENT(MOL)MSlinear probe0.441 ± 0.034
SCENT(POM)MSlinear probe0.411 ± 0.036
MolFormerstructurelinear probe0.382 ± 0.038
EIMS2VecMSlinear probe0.335 ± 0.038

Thank you for the thorough response and additional experiments, which address my concerns. I would encourage moving the trend analysis into the main text. Although the scope remains that of an application paper rather than a new method, this is thorough and sound work in a largely unexplored modality (olfaction). I am therefore raising my overall rating to 4.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.