SCENT: Aligning Mass Spectra with Molecular Structure for Olfactory Perception

Abstract: Predicting human olfactory perception from molecular structure has seen remarkable progress, yet these approaches require explicit chemical structure at inference, which is not available in practical sensing settings. We address this gap by exploring direct electron ionization mass spectrometry (EI-MS), a sensing technique that acquires chemically informative fragmentation fingerprints in seconds, as an alternative input modality for olfactory prediction. We contribute 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. On the multi-label odor descriptor prediction task, SCENT significantly outperforms MS-only baselines and achieves performance comparable to structure-based models, despite requiring no explicit molecular structure at test time. The learned representations also better approximate continuous human perceptual ratings and generalize to real-world lab-measured spectra, suggesting that cross-modal alignment is an effective strategy for grounding analytical spectra in chemical semantics.

• The paper introduces SCENT, a two-stage contrastive framework that aligns EI mass-spectral embeddings with pretrained molecular-structure representations, enabling odor prediction without molecular structures at inference time.
• SCENT reaches up to 89.99 micro-AUC and 85.71 weighted-AUC on GS-LF, nearly matching structure-based models while significantly outperforming unaligned spectral baselines, especially in low-data settings.
• The paper shows that SCENT transfers to human perceptual ratings and real laboratory spectra, but identifies mixture handling, concentration effects, domain adaptation, and larger external validation sets as key deployment challenges.

Motivation and problem setting

Structure-based models for quantitative structure–odor relationship (QSOR) prediction have advanced considerably, but they all presuppose access to the molecular identity or an explicit chemical structure at inference time. In practical sensing scenarios—field monitoring, quality control, environmental analysis—such structural information requires specialized instrumentation, expert interpretation, and acquisition times on the order of hours. The paper identifies this as the central gap between computational olfaction research and deployable chemical sensing, and proposes direct electron ionization mass spectrometry (EI-MS) as an alternative input modality: a single-molecule fragmentation fingerprint acquirable in seconds that retains chemically interpretable content, unlike the non-specific signals of electronic noses.

Prior end-to-end approaches that predict odor directly from mass spectra treat the spectrum as a flat signal and do not explicitly exploit its latent chemical content. The paper asks whether a model can be trained such that chemical structure information is distilled into mass-spectral representations during training, making spectra alone sufficient for olfactory prediction at inference.

The SCENT framework

SCENT (Spectrum-to-Chemical Embedding aligNmenT) is a two-stage, CLIP-style contrastive framework. In the alignment stage, a learnable MS encoder is trained against frozen pretrained chemical structure encoders over a shared latent space using a symmetric temperature-scaled cosine-similarity contrastive loss. The MS encoder builds on EIMS2Vec-style peak co-occurrence embeddings with power-scaled intensities ($p=0.5$), but replaces sum pooling with a two-layer, four-head Transformer over the peak tokens, followed by a projection MLP on the CLS token. Two frozen structure teachers are considered: Open-POM (a message-passing GNN retrained from scratch on leakage-filtered odor labels, 256-dim) and MolFormer (a foundation model pretrained on over 1.1 billion molecules, 768-dim, used off-the-shelf), yielding two variants, SCENT (Open-POM) and SCENT (MolFormer).

Alignment training uses the NIST EI-MS library (2023), filtered to molecular weights of 50–300 Da, yielding 148,212 compound–spectrum pairs with PubChem-retrieved, RDKit-canonicalized SMILES. Strict leakage controls ensure no downstream test molecule appears in alignment or teacher-retraining data.

Downstream evaluation uses a filtered GS-LF benchmark (138 odor descriptors, 2,588 single-molecule samples) with a fixed 10% iterative-stratified test set and 5-fold cross-validation on the remainder, training an MLP classification head on frozen embeddings. Perceptual rating regression uses the DREAM dataset (385 molecules with valid spectra; 21 descriptors; ridge regression over 100 repeated 80/20 splits, evaluated by Pearson correlation).

Main results

On the filtered GS-LF benchmark, both SCENT variants substantially outperform the unaligned EIMS2Vec baseline and are essentially on par with structure-based models, despite requiring no structure at inference:

Model	Input	Micro-AUC	Weighted-AUC	Adj. P@5
Open-POM	SMILES	89.94 ± 0.05	86.70 ± 0.15	52.22 ± 0.28
MolFormer	SMILES	89.29 ± 0.20	84.14 ± 0.44	50.23 ± 0.16
EIMS2Vec	MS	87.98 ± 0.15	82.29 ± 0.61	46.49 ± 0.53
SCENT (Open-POM)	MS	89.80 ± 0.09	85.45 ± 0.30	50.60 ± 0.53
SCENT (MolFormer)	MS	89.99 ± 0.18	85.71 ± 0.26	50.62 ± 0.75

The residual weighted-AUC gap to structure-based models is small and, per frequency-binned per-label analysis, reflects modest differences accumulated across label groups rather than failure in any particular frequency regime. SCENT's gains over the baseline are statistically significant (Wilcoxon signed-rank, $p < .05$) across all frequency bins, with the largest absolute improvements in low- and medium-frequency categories—an important property given the long-tailed GS-LF label distribution. Analysis of intra-label spectral and structural similarity confirms that SCENT's advantage holds across descriptors regardless of similarity position or label frequency.

Ablation and data efficiency

The ablation isolates the source of improvement decisively. Replacing the contrastive objective with BERT-style masked reconstruction drops micro-AUC to 76.34 and weighted-AUC to 55.39; replacing it with end-to-end supervised training on odor labels yields 77.90 / 64.60. Both are far below the aligned models, demonstrating that neither encoder capacity nor label supervision alone explains the gains—it is specifically the alignment with pretrained chemical structure embeddings. Removing the Transformer head while retaining alignment produces a smaller, consistent degradation (e.g., 88.40 vs. 89.99 micro-AUC for the MolFormer variant), indicating the Transformer contributes a secondary benefit in modeling peak co-occurrence.

Data-scaling experiments (10%–100% training fractions) show SCENT consistently outperforming EIMS2Vec at every fraction, with the gap most pronounced in the 10–20% regime. This indicates the alignment pretraining instills a chemical prior that cannot be recovered from the spectral signal alone and improves low-data generalization.

Human perceptual rating regression

On DREAM, both SCENT variants significantly outperform the unaligned baseline on per-attribute Pearson correlation (Wilcoxon signed-rank, $n=21$ attributes: $p=0.035$ for Open-POM-based, $p=0.002$ for MolFormer-based comparisons). Notably, SCENT (MolFormer) is statistically indistinguishable from its MolFormer teacher ($p=1.000$)—a strong result, since MolFormer was never trained with olfactory supervision, implying that general chemical structure representations already encode perceptually relevant information that SCENT successfully transfers into spectral space. In contrast, SCENT (Open-POM) remains below Open-POM ($p<0.001$); the authors attribute this to Open-POM's odor-supervised embeddings containing task-specific structure that cannot be fully recovered from MS signals.

Real-world validation

To test robustness to the library-to-lab domain gap (varying peak intensities, background contamination, no chromatographic separation), the authors collected direct EI-MS measurements of 30 single-molecule odorants in laboratory air (~60 s per measurement), applying a sliding-slope segmentation of the total ion current and per-channel baseline correction at $\mu_\text{bg} + 5\sigma_\text{bg}$, with the m/z range restricted to [50, 180] to avoid air ions and low-volatility fragments.

The headline finding of this section is that SCENT evaluated on real-world spectra outperforms EIMS2Vec evaluated on clean library spectra across all metrics—e.g., micro-AUC of 88.67 ± 0.49 (95% CI [86.94, 90.36]) versus 84.80 ± 0.52 (CI [82.63, 86.96]) for the baseline, and Adj. P@5 of 51.18 versus 40.67 with non-overlapping confidence intervals. Weighted-AUC intervals overlap, reflecting sensitivity to the 30-sample external test set. Additionally, when both models are retrained under the truncated [50, 180] m/z range, the unaligned baseline suffers a notably larger degradation than SCENT (e.g., weighted-AUC drop of 4.23 points for EIMS2Vec vs. 2.72 for SCENT (Open-POM)), suggesting that contrastive alignment yields representations more robust to spectral truncation and domain shift.

Limitations and open questions

The authors are explicit about several constraints. Alignment and downstream training rely on library spectra, and while the case study is encouraging, the library-to-real-world domain gap remains unresolved for deployment, and the 30-molecule external test set limits statistical strength—weighted-AUC differences there are not significant. The framework assumes single-molecule inputs; extending alignment to mixed-compound spectra, which are typical in direct sampling without chromatographic separation, is left open. Concentration is not modeled: the representation reflects relative fragmentation structure rather than abundance, so concentration-dependent, potentially non-linear perceptual effects are outside the model's scope. Finally, the residual gap to odor-supervised structure teachers (Open-POM) suggests that task-specific perceptual information is only partially recoverable through spectra, and the paper leaves open how much of this ceiling is informational versus architectural.

Conclusion

SCENT establishes cross-modal contrastive alignment as an effective mechanism for grounding rapid analytical spectra in chemical semantics: it closes most of the gap between MS-only and structure-based odor prediction, transfers human-perceptual structure into spectral embeddings, and retains its advantage under realistic measurement conditions without domain-specific fine-tuning. The work positions direct EI-MS as a practical input modality for deployable computational olfaction, while leaving domain adaptation, mixture spectra, and concentration modeling as clearly stated open problems.