Organizations: CSIR- Central Scientific Instruments Organisation, Sector 30-C, Chandigarh-160030, India · Academy of Scientific and Innovative Research (AcSIR), Ghaziabad-201002, India · Department of Higher Education,Himachal Pradesh,Shimla -171001,India · Indian Institute of Science Education and Research Bhopal(IISERB), Madhya Pradesh-462066,India · Monell Chemical Senses Center, Philadelphia, Pennsylvania, USA 19104 · Department of Otolaryngology, Head and Neck Surgery, University of Pennsylvania, Philadelphia, Pennsylvania, USA 19104
Predicting olfactory qualities from molecular structure is an open problem in chemoinformatics. Although linear models can link molecular features to odor descriptors, they often fail when extrapolating to novel chemical scaffolds, extreme molecular weights, or complex odor mixtures. To address this, we introduce GraphNOSE, an open-source graph transformer framework that predicts multi-label odor descriptors from simplified molecular-input line-entry system (SMILES) strings for single molecules and binary mixtures. By integrating positional and structural encodings within a transformer-based graph architecture, GraphNOSE achieves strong performance with six times fewer parameters than standard graph neural network (GNN) baseline while consistently outperforming linear models, molecular language model embeddings, molecular fingerprints, and baseline GNNs by an average area under the ROC curve (AUROC) margin of 4.52% (p < 0.01). GraphNOSE achieves an AUROC of 84% on out-of-distribution compounds (OODs). This exceeds the current state-of-the-art GNN for OOD in olfaction (Open-POM: 81%, p < 0.001), and identifies conditions under which linear models empirically fail. Finally, we apply XAI (explainable AI) methods to identify which substructures and molecular features drive odor predictions, yielding insights consistent with chemical intuition and grounded in the model's learned representations. Together, these results establish GraphNOSE as a scalable and interpretable architecture for olfactory prediction that generalizes to structurally distinct compounds underrepresented in current perceptual databases.
Figures & tables
Figure 1: Overview of the GraphNOSE framework architecture for multi-label molecular odor prediction. (A) Raw- input SMILES strings are processed through a custom featurizer using RDKit Landrum et al. (2025) and DeepChem Ramsundar et al. (2019) , constructing undirected molecular graphs G(V,E,Xn,Xe) , where atoms and bonds represent nodes and edges, respectively. (B) The structural graph and node features are augmented using a graph positional and structural encoder (GPSE) pre-trained on diverse molecular and graph datasets (e.g., ZINC, ChEMBL, GEOM, PCQM4Mv2, MolPCBA) or standard baseline (RWSE, LapPE) to generate comprehensive spatial representations ( Xn′ ). (C) The enriched graph representations are fed into a hybrid GraphGPS Rampášek et al. (2022) backbone that integrates local message-passing neural networks (MPNNs) for near-neighbor functional group interactions with global self-attention mechanisms for long-range molecular topology. (D) Finally, the model predicts labels for the input molecule from the 138 odor descriptors (e.g., fruity, sweet, musky, garlic).
Feature
Categories
Atomic number
1 (H) to 54 (I), UNK
Valence
0, 1, 2, 3, 4, 5, 6, UNK
Degree
0, 1, 2, 3, 4, 5, UNK
Formal charge
−2 , −1 , 0, 1, 2, UNK
Number of hydrogens
0, 1, 2, 3, 4, 5, 6, 7, 8, UNK
Hybridization
sp, sp 2 , sp 3 , sp 3 d, sp 3 d 2 , UNK
Table 1: Node (atomic) and edge (bond) features used in molecular graphs. All features are one-hot encoded and concatenated into a single bit vector. UNK denotes unknown values and serves as a catch-all category.
PSE
GraphNOSElin
GraphNOSEtrans
RWSE
87.356±0.15∗∗∗
84.474±1.04∗∗∗
LapPE
87.796±0.13∗∗∗
88.378±0.06∗∗∗
GPSEChEMBL
88.864±0.08
89.084±0.07
GPSEGEOM
88.916±0.05
89.124±0.13
GPSEPCQM4Mv2
88.930±0.12
88.990±0.09
GPSEZINC
88.958±0.18
89.026±0.10
Table 2: Different PSE augmentations. These macro AUROC scores are averaged over five random seeds. Positional and structural encodings (PSEs) are explicitly augmented to the node features to provide topological context to the global attention mechanism. GraphNOSElin uses a linear transformer Choromanski et al. (2020) for global attention, while GraphNOSEtrans uses standard self-attention Vaswani et al. (2017) with quadratic complexity. Significance levels ( ∗ denote paired t -test comparisons relative to the GPSEMolPCBA ): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 .
Class
Model
ROC (%)
p -value
Cohen’s d
Linear [logistic regression (LR)]
LR (TopTorsion) Nilakantan et al. (1987)
80.82
–
–
LR (RDKit) Landrum et al. (2025)
82.05
–
–
LR (AtomPair) Carhart et al. (1985)
82.83
–
–
LR (Morgan_ECFP4) Rogers and Hahn (2010)
83.56
–
–
LR (MACCS) Durant et al. (2002)
84.76
–
–
LLM embedding
ChemBERTa Chithrananda et al. (2020)
84.78±0.04∗∗∗
4.93×10−7
26.40
Table 3: Baselines . Models are grouped into descriptor-based methods, LLM embedding approaches, standard GNN architectures, other GNN variants, and the proposed method. Linear models using the limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) solver are deterministic and therefore do not report standard deviations, paired p -values, or effect sizes. Significance levels and effect sizes ( ∗ denote paired t -test comparisons and Cohen’s d relative to the GraphNOSElin baseline): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 . LR, logistic regression; RF, random forest.
Model
Parameters
OpenPOM
2M
LDMDMPNN
674K
LDMGCN
671K
GraphNOSE
314K
Table 4: Parameter counts for single molecule odor prediction models , reported in thousands (K) and millions (M).
Figure 2: Impact of training set size on model performance: Mean predictive performance on the held-out test set as a function of the training data proportion. Error bars represent the standard deviation across 5 random seeds. While all architectures exhibit performance gains as the training set size increases, both GraphNOSE variants consistently outperform baseline GNNs (OpenPOM and LDM variants) across all sampling rates, demonstrating superior data efficiency even in low-data regimes.
Model
AUROC (%)
p -value
Cohen’s d
GraphNOSEtrans
84.34±0.18
–
–
GraphNOSElin
84.26±0.33
–
–
OpenPOM
80.84±0.35∗∗∗
4.39×10−7
27.19
LDMGCN
79.13±0.44∗∗∗
1.52×10−5
11.18
LDMDMPNN
79.03±0.73∗∗∗
5.63×10−5
8.04
LR (MACCS)
78.46
–
–
Table 5: Out-of-distribution (OOD) training results for molecule odor prediction across five random seeds . As linear models using the LBFGS solver are deterministic, they do not report standard deviations, paired p -values, or effect sizes. The best-performing score is highlighted in bold. LR, logistic regression. Significance levels ( ∗ denote paired t -test comparisons relative to the GraphNOSElin baseline): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 .
Figure 3: Nearest neighbor retrieval for a reference molecule.a) Reference molecule b) Top five nearest neighbors retrieved using cosine similarity on GraphNOSE continuous embeddings. c) Top five nearest neighbors retrieved using Tanimoto distance on bit-based Morgan fingerprints (serving as the established baseline in chemo-informatics Jasial et al. (2016) ).
Figure 4: Analysis of odor space: Kernel density estimation (KDE) plots for Morgan fingerprint (a) and predicted labels by OpenPOM (b) and GraphNOSE (c) Shaded and contoured areas are kernel density estimates of the distribution of labeled data. Colors denote odor categories, with shaded areas for broad classes and contours for specific descriptors (e.g., lavender, roasted).
Figure 5: Explainability through integrated gradients Sundararajan et al. (2017) of different odor descriptors. The visualizations highlight the molecular features most important for predicting each odor category. Green highlights atom-level contributions; red highlights bond-level contributions. Node and edge importance are visualized by the opacity and are proportional to the values. For more examples in each category, see Supplementary Material, Figure S4.
Model
PSE / Features
ROC (%)
p -value
Cohen’s d
GraphNOSElin
GPSEMolPCBA
77.96±0.58
–
–
GPSEChEMBL
77.84±0.54
0.778
0.14
GPSEpcqmv2
77.64±0.99
0.349
0.47
GPSEzinc
77.58±0.67
0.396
0.42
LapPE
77.54±1.24
0.601
0.25
GPSEGeom
77.24±0.63∗
0.038
1.36
Table 6: Results for binary odor mixture prediction. Scores are macro AUROC scores averaged over five random seeds. Linear models (LR, logistic regression) are deterministic and therefore do not report standard deviations, paired p -values, or effect sizes. Significance levels ( ∗ denote paired t -test comparisons relative to GPSEMolPCBA ): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 .
Figure 6: Figure S1: Structural similarity does not imply perceptual similarity. Perceptual odor similarity to a reference molecule (here, m-cresol, a) can be observed among some structurally related molecules (e.g., o-cresol, b) but not others (e.g., 3-methylcyclohexanol, c) , and structurally dissimilar molecules (e.g., a complex bicyclic ketone, d) can show odor similarity. Molecules sourced from the GF-LS dataset. Luebke (2019) ; Leffingwell (2005)
Figure 7: Figure S2: Dataset insights: (a) Frequency distribution of the top 30 odor labels among 138. (b) Number of labels per molecule in the GS-LF dataset Barsainyan et al. (2023a) (total n = 4983).
Dataset
Primary Source
License
Unique Graphs
MolPCBA
MoleculeNet Hu et al. (2020)
MIT
323,555
ZINC
ZINC Database Gómez-Bombarelli et al. (2018)
Apache 2.0 License
249,455
GEOM
Axelrod & Gomez-Bombarelli Axelrod and Gomez-Bombarelli (2022)
CC0 1.0
169,925
ChEMBL
Gaulton et al. Gaulton et al. (2012)
CC BY-SA 3.0
970,963
PCQM4Mv2
PubChemQC Hu et al. (2021)
CC BY 4.0
273,920
Table 12.1 : Table S2: Summary of molecular datasets used in pretraining of GPSE.
Hyperparameter
Search Space
Optimal Value
Architecture Parameters
Hidden dimension
{64, 128, 256, 288, 352, 512}
64
Classifier dimension
{128, 256, 512, 1024}
512
Heads
{1, 2, 4}
1
No. of GPS layers
{4, 5, 6, 7}
4
Dropout
[0.0, 0.5]
0.440
Table 13.2 : Table S3: Hyperparameters for GraphNOSE
Sampled Train
Seed 1
Seed 10
Seed 100
Seed 21
Seed 42
p -val (vs. GraphNOSElin )
p -val (vs. GraphNOSEtrans )
30%
82.32
81.88
82.62
80.55
82.40
1.51×10−3∗∗
1.43×10−3∗∗
40%
83.91
83.52
84.12
84.01
84.06
4.39×10−4∗∗∗
1.55×10−4∗∗∗
50%
84.35
84.35
84.01
84.95
84.56
2.95×10−4∗∗∗
1.06×10−4∗∗∗
60%
85.01
85.54
85.38
85.15
85.20
3.00×10−5∗∗∗
2.80×10−5∗∗∗
70%
86.72
86.96
86.58
86.90
86.51
7.59×10−4∗∗∗
1.10×10−4∗∗∗
80%
87.03
87.07
86.92
87.48
86.65
2.87×10−4∗∗∗
5.70×10−4∗∗∗
Table 14.3 : Table S4: AUROC on the held-out test set across five random seeds and paired t -test p -values comparing OpenPOM against GraphNOSElin and GraphNOSEtrans across training data proportions (30% – 90%). Significance levels ( ∗ denote paired t -test comparison): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 .
Sampled Train
Seed 1
Seed 10
Seed 100
Seed 21
Seed 42
p -val (vs. GraphNOSElin )
p -val (vs. GraphNOSEtrans )
30%
82.71
77.31
80.97
82.09
81.56
1.50×10−2∗
1.29×10−2∗
40%
83.57
83.45
82.74
83.24
83.85
5.10×10−5∗∗∗
3.60×10−5∗∗∗
50%
83.80
84.80
83.87
84.97
84.72
2.79×10−4∗∗∗
2.11×10−4∗∗∗
60%
84.68
84.91
83.83
84.43
85.10
1.38×10−4∗∗∗
1.80×10−4∗∗∗
70%
85.01
85.30
85.85
85.31
85.56
1.60×10−5∗∗∗
2.30×10−5∗∗∗
80%
86.16
86.18
86.13
86.08
86.56
1.00×10−5∗∗∗
2.31×10−7∗∗∗
Table 14.4 : Table S5: AUROC on the held-out test set across five random seeds and paired t -test p -values comparing LDMGCN against GraphNOSElin and GraphNOSEtrans across training data proportions (30% – 90%). Significance levels ( ∗ denote paired t -test comparison): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 .
Sampled Train
Seed 1
Seed 10
Seed 100
Seed 21
Seed 42
p -val (vs. GraphNOSElin )
p -val (vs. GraphNOSEtrans )
30%
81.76
82.40
82.10
81.81
81.67
3.00×10−6∗∗∗
1.66×10−6∗∗∗
40%
83.96
83.90
83.39
81.82
83.69
1.70×10−3∗∗
2.55×10−3∗∗
50%
83.87
83.30
82.95
83.58
84.03
3.71×10−4∗∗∗
1.62×10−4∗∗∗
60%
85.49
84.76
84.70
85.02
84.86
5.20×10−5∗∗∗
3.40×10−5∗∗∗
70%
85.88
85.63
85.46
85.74
83.94
1.46×10−3∗∗
1.06×10−3∗∗
80%
85.94
86.08
86.32
85.70
86.12
7.50×10−5∗∗∗
1.90×10−5∗∗∗
Table 14.5 : Table S6: AUROC on the held-out test set across five random seeds and paired t -test p -values comparing LDMDMPNN against GraphNOSElin and GraphNOSEtrans across training data proportions (30% – 90%). Significance levels ( ∗ denote paired t -test comparison): ∗∗∗p<0.001 , ∗∗p<0.01 , and ∗p<0.05 .
Figure 15.1 : Figure S3: Analysis of odor space: KDE plot for true label. Shaded and contoured areas are kernel density estimates of the distribution of labeled data. Colors denote odor categories, with shaded areas for broad classes and contours for specific descriptors (e.g., lavender, roasted).
Figure 16.2 : Figure S4: Explainability through integrated gradients Sundararajan et al. (2017) for different odor descriptors: (a) alcohol, (b) sulfurous, (c) phenolic, (d) fruity, (e) musk, and (f) ketonic. Green indicates atom-level contributions; red, bond-level contributions. Opacity indicates node and edge importance and is proportional to the values.
Foundation models have transformed molecular property prediction, yet it remains unclear whether a molecular foundation model, fine-tuned on a single canonical olfactory prediction task, can learn representations that transfer across diverse machine olfaction problems. We investigate this question by fine-tuning Uni-Mol2 on the GS-LF benchmark for multi-label odor descriptor prediction and evaluating the resulting model, without additional deep-learning training, on four complementary downstream settings: cross-dataset odor descriptor prediction, odorous-versus-odorless classification, enantiomer evaluation, and odor mixture discriminability. The fine-tuned model matches or exceeds the performance of the state-of-the-art olfaction-specific baseline on the primary GS-LF benchmark and consistently transfers across these downstream evaluations. The enantiomer analysis further shows that three-dimensional molecular representations distinguish mirror-image molecules in a way that two-dimensional graph models fundamentally cannot, although accurately predicting the perceptual consequences of stereochemistry remains an open challenge. Together, these results support a train-once, transfer-across-tasks paradigm for machine olfaction and suggest that chemically pretrained molecular representations provide a strong foundation for transferable olfactory prediction.
Yikun Han, Yi Wang, Neil Mankodi +2
Department of Statistics, University of Michigan, Ann Arbor, Michigan, USA · Computer Science and Engineering Division, University of Michigan, Ann Arbor, Michigan, USA
Olfaction lies at the intersection of chemical structure, neural encoding, and linguistic perception, yet existing representation methods fail to fully capture this pathway. Current approaches typically model only isolated segments of the olfactory pathway, overlooking the complete chain from molecule to receptors to linguistic descriptions. Such fragmentation yields learned embeddings that lack both biological grounding and semantic interpretability. We propose NOSE (Neural Olfactory-Semantic Embedding), a representation learning framework that aligns three modalities along the olfactory pathway: molecular structure, receptor sequence, and natural language description. Rather than simply fusing these signals, we decouple their contributions via orthogonal constraints, preserving the unique encoded information of each modality. To address the sparsity of olfactory language, we introduce a weak positive sample strategy to calibrate semantic similarity, preventing erroneous repulsion of similar odors in the feature space. Extensive experiments demonstrate that NOSE achieves state-of-the-art (SOTA) performance and excellent zero-shot generalization, confirming the strong alignment between its representation space and human olfactory intuition.Code and data are available at https://github.com/Xianyusyy/NOSE
Yanyi Su, Hongshuai Wang, Zhifeng Gao +1
State Key Laboratory of Physical Chemistry of Solid Surface, College of Chemistry and Chemical Engineering, Xiamen University, Xiamen, China · DP Technology · Laboratory of AI for Electrochemistry (AI4EC), Tan Kah Kee Innovation Laboratory (IKKEM), Xiamen, China +1
Here we introduce the Olfactory Perception (OP) benchmark, designed to assess the capability of large language models (LLMs) to reason about smell. The benchmark contains 1,010 questions across eight task categories spanning odor classification, odor primary descriptor identification, intensity and pleasantness judgments, multi-descriptor prediction, mixture similarity, olfactory receptor activation, and smell identification from real-world odor sources. Each question is presented in two prompt formats, compound names and isomeric SMILES, to evaluate the effect of molecular representations. Evaluating 21 model configurations across major model families, we find that compound-name prompts consistently outperform isomeric SMILES, with gains ranging from +2.4 to +18.9 percentage points (mean approx +7 points), suggesting current LLMs access olfactory knowledge primarily through lexical associations rather than structural molecular reasoning. The best-performing model reaches 64.4% overall accuracy, which highlights both emerging capabilities and substantial remaining gaps in olfactory reasoning. We further evaluate a subset of the OP across 21 languages and find that aggregating predictions across languages improves olfactory prediction, with AUROC = 0.86 for the best performing language ensemble model. LLMs should be able to handle olfactory and not just visual or aural information.
Eftychia Makri, Nikolaos Nakis, Laura Sisson +4
Department of Electrical Engineering, Yale University, New Haven, CT, USA · Human Nature Lab, Yale University, New Haven, CT, USA · Patina, New York, NY, USA +1