The Chemistry of Smell: Learning Generalizable

Perceptual Representations of Small Molecules

Benjamin Sanchez-Lengeling1,* , Jennifer N Wei1,* , Brian K Lee1 , Richard C Gerkin2 , Alán

Aspuru-Guzik3 , and Alexander B Wiltschko1
1
Google Research, Brain Team
2
Arizona State University
3
Department of Chemistry, University of Toronto
3
Department of Computer Science, University of Toronto
3
Vector Institute for Artificial Intelligence, Toronto, Ontario, Canada
3
Canadian Institute for Advanced Research, Toronto, Ontario, Canada
*
Contributed equally

Abstract
In machine learning for domains in physics and chemistry, an important prob-
lem is learning meaningful and useful representations of physical phenomena that
are both predictive and generalize to new tasks. One such problem is predicting
quantitative structure-odor relationships (QSOR). A solution would impact hu-
man nutrition, manufacture of synthetic fragrance, the environment, and sensory
neuroscience. In this paper, we propose the use of graph neural networks for this
decades-old task, and show results that significantly out-perform prior methods on
a novel data set labeled by olfactory experts. Additional analysis shows that the
learned embeddings from graph neural networks capture a meaningful represen-
tation of the underlying relationship between structure and odor, as demonstrated
by strong performance on two challenging transfer learning tasks. Based on these
early results with graph neural networks for molecular properties, we hope the
field can build towards doing for olfaction what machine learning has already
done for vision.

1 Introduction
Predicting analytical properties of molecules is an area of growing research in machine learning, par-
ticularly as models for learning from graph-valued inputs improve in sophistication and robustness.
A molecular property prediction problem that has received comparatively little attention during this
surge in research activity is the prediction of Quantitative Structure-Odor Relationships (QSOR)1 .
This is a 70+ year old problem straddling chemistry, physics, sensory neuroscience and machine
learning. This has remained an open problem for so long due to its difficulty—very small changes
in a molecule’s structure can have dramatic effects on its odor, a phenomenon known in medicinal
chemistry as an “activity cliff” (1; 2). A classic example is Lyral, which is a commercially suc-
cessful muguet molecule (a floral scent often used in dryer sheets). Its structural neighbors are not
always perceptual neighbors, and vice versa (Figure 1).
Advances in deep learning for vision and audition suggest that we might be able to directly pre-
dict the end sensory result of an input stimulus, even without detailed knowledge of the systems
and circuits linking the physical world and our internal sensations. Advances in deep learning for
olfaction would aid in the discovery of new synthetic odorants, thereby reducing ecological im-
pact of harvesting natural products, and would advance our understanding of sensory perception in
1
As opposed to Quantitative Structure-Activity Relationships (QSAR), a term from medicinal chemistry

Second Workshop on Machine Learning and the Physical Sciences (NeurIPS 2019), Vancouver, Canada.
Figure 1: Structurally similar molecules do not necessarily have similar odor descriptors. A.
Lyral, the reference molecule. B. Molecules with similar structure can share similar odor descriptors.
C. However, a small structural change can render the molecule odorless. D. Further, large structural
changes can leave the odor of the molecule largely unchanged. Example from Ohloff, Pickenhagen
and Kraft (5).

the brain by offering new ways of representing and processing olfactory data. Here, we curated a
dataset of molecules associated with expert-labeled odor descriptors2 , and train Graph Neural Net-
works (GNN) to predict these odor descriptors using a molecule’s graph structure alone. Further,
we show that the embeddings learned to solve this task generalize well to downstream tasks, a rare
occurrence in machine learning applications in chemistry (3; 4). This indicates we have captured a
general-purpose representation of a molecule’s odor properties, which might be useful in the future
for rational molecular design or screening.

2 Olfactory Dataset

We assembled an expert-labeled set of n = 5030 molecules from two separate sources: GoodScents
perfume materials database, n = 3786 (6) and Leffingwell PMP 2001 database, n = 3561 (7). The
datasets share 2317 overlapping molecules. Molecules are labeled with one or more odor descrip-
tors by olfactory experts (usually a practicing perfumer), creating a multi-label prediction problem.
GoodScents describes a list of 1–15 odor descriptors for each molecule (Figure 2A), whereas Leff-
ingwell uses free-form text. Odor descriptors were canonicalized using GoodScents’ ontology, and
overlapping molecules inherited the union of both datasets’ odor descriptors. After filtering for odor
descriptors with at least 30 representative molecules, 138 odor descriptors remained, including an
odorless descriptor (Figure 2B). Some odor descriptors are extremely common, like fruity or green,
while others are rare, like radish or bready. This bias may be due to our dataset being a collection of
materials for perfumery and on the degree of specificity of the odor descriptors. Note that there is an
extremely strong co-occurrence structure among odor descriptors that reflects a common-sense intu-
ition of which odor descriptors are similar and dissimilar (Figure 2C). For example, there is a dairy
cluster that includes the dairy, yogurt, milk, and cheese descriptors. There is also a fruity cluster
with apple, pear, pineapple, pear, etc., and a bakery cluster that includes toasted, nutty, and cocoa.
Previous approaches in QSOR often train one model per odor descriptor. Here, we apply GNNs to
all odor descriptor tasks at once, allowing us to take advantage of this correlation structure.

3 Results

3.1 Classification Performance Comparison

We consider two types of GNNs: Message Passing Neural Networks (MPNN) (8) and Graph Con-
volution Networks (GCN) (9), and compare against baselines of random forests (RF) and k-nearest
neighbors (kNN). For baseline featurizations, we used bit-based RDKitFingerprints (bFP), count-
based Morgan fingerprints (cFP), and Mordred features (10). We report several metrics (Table 1),
as each metric can highlight different performance characteristics, however we primarily compare
models on mean AUROC (averaged across odor descriptors). We found that MPNNs and GCNs
perform similarly. Both MPNNs and GCNs significantly outperform all baseline models.
2
In QSAR terminology, descriptors are used for input features in a model. The term odor descriptors is
used in QSOR for odorant properties to predict.

2
Figure 2: Dataset overview. A. Distribution of odor descriptor frequencies. B. Distribution of label
density. C. Co-ocurrence matrix for odor descriptors. The 10 most frequent descriptors are removed
for visual clarity, and remaining descriptors re-ordered using spectral clustering. Main odor groups
with examples are highlighted.

AUROC Precision Recall F1

MPNN 0.890 [0.882, 0.898] 0.379 [0.352, 0.399] 0.387 [0.366, 0.408] 0.362 [0.335, 0.375]
GCN 0.894 [0.888, 0.902] 0.379 [0.351, 0.398] 0.390 [0.365, 0.412] 0.360 [0.337, 0.372]
RF-Mordred 0.850 [0.838, 0.860] 0.311 [0.288, 0.333] 0.393 [0.372, 0.417] 0.306 [0.283, 0.319]
RF-bFP 0.832 [0.821, 0.842] 0.321 [0.293, 0.339] 0.371 [0.350, 0.390] 0.295 [0.272, 0.308]
RF-cFP 0.845 [0.835, 0.854] 0.315 [0.280, 0.332] 0.375 [0.354, 0.398] 0.295 [0.272, 0.311]
KNN-bFP 0.791 [0.778, 0.803] 0.328 [0.305, 0.347] 0.390 [0.366, 0.411] 0.323 [0.299, 0.335]
KNN-cFP 0.796 [0.785, 0.809] 0.333 [0.307, 0.351] 0.365 [0.342, 0.389] 0.316 [0.292, 0.327]

Table 1: Odor descriptor prediction results. mean, 95% CI [lower, upper] bootstrap bounds
reported. Numbers reported are an unweighted mean across all 138 odor descriptors. Precision/recall
decision thresholds are optimized for F1 score. Best values for each metric are in bold; recall had
no clear winner.

3.2 Evaluating Embedding Performance in a Lookup Task

We wished to evaluate whether embeddings extracted from our trained GCN space reflected odor per-
ceptual space. Specifically, we compare whether molecules with small cosine distances in our GCN
embeddings were perceptually similar, as compared to using Tanimoto distances on bFP features.
While the latter strategy is commonly used in cheminformatics for molecule retrieval, molecules
with similar structural features do not always smell the same (Figure 1), so we anticipated that it
would not perform as well at retrieving similar-smelling molecules.
To test this idea, we trained a k-nearest neighbors classifier (k = 20) using cosine distance on GCN
embeddings, and using Tanimoto distance on bit-based Morgan fingerprints. GCN embeddings
(AUROC = 0.818, 95% CI [0.806, 0.830] ) outperformed bFP (AUROC = 0.782, 95% CI [0.773,
0.797]). By inspecting the nearest neighbors found by each method (Figure 3), we can see that
both methods yield molecules with similar structural features, but retrieval using GCN embeddings
produces molecules that are more perceptually similar to the source molecule. We conclude that
relative to bit-based fingerprints, GCN embeddings emphasize similarity of smell over similarity in
structure.

3
Figure 3: Nearest neighbor retrieval. k=5 nearest neighbors are shown for cosine similarity on
GCN embeddings and for Tanimoto distance on bit-based Morgan fingerprints. Reported AUROCs
are averaged over all odor descriptors, with k = 20.

4 Evaluating Generalization Performance of Odor Embeddings

We have shown that GNNs significantly outperform existing methods on predicting odor descriptors,
and that our embedding space has useful local structure that can be used to search for similarly-
smelling molecules. We now explore whether these odor embeddings generalize to other odor pre-
diction tasks.

4.1 Transfer Learning to Previously-Unseen Odor Descriptors

An odor descriptor may be newly invented or refined (e.g., molecules with the pear descriptor might
be later attributed a more specific pear skin, pear stem, pear flesh, pear core descriptor). A useful
odor embedding would be able to perform transfer learning (11) to this new descriptor, using only
limited data. To approximate this scenario, we ablated one odor descriptor at a time from our dataset.
Using the embeddings trained from (N − 1) odor descriptors as a featurization, we trained a random
forest to predict the previously held-out odor descriptor. We used cFP and Mordred features as
a baseline for comparison. The results are shown in Figure 4. GNN embeddings significantly
outperform Morgan fingerprints and Mordred features on this task, but as expected, still perform
slightly worse than a GNN trained on the target odor. This indicates that GNN-based embeddings
may generalize to predict new, but related, odors.

Figure 4: Mean AUROC on a previously held-out odor. Average AUROC scores across all labels
on the single label ablation task. The error bars denote 95% confidence intervals. The top bar denotes
the performance of the model trained on all of the labels. The middle bar denotes the performance
of a random forest model trained using the GNN embeddings from a model trained on (N − 1) odor
labels. The bottom bar denotes a random forest trained on counting Morgan fingerprints.

4.2 Transfer Learning to the DREAM Olfactory Prediction Challenge

The DREAM Olfaction Prediction Challenge (12) was an open competition to build QSOR models
on a dataset collected from untrained panelists. The DREAM dataset has several differences from

4
our own. First, it was a regression problem – panelists rated the amount that a molecule smelled of
a particular odor descriptor on a scale from 1 to 100. Second, it had 476 molecules compared to
our ∼ 5k (although our dataset contains nearly all of the DREAM molecules). Third, the ratings
were provided by a large panel of untrained individuals over a short period of time, whereas ours
were gleaned from a small set of experts over many years. The DREAM challenge measured model
performance as the Pearson’s r correlation of model predictions with the mean reported intensity of
each odor descriptor.
The winning DREAM model used random forest models with a combination of several sources of
features, primarily Dragon and Morgan fingerprints, among other sources of information (12). Using
only our embedding with a random forest model, we achieve a mean Pearson’s r = 0.55, and the
state-of-the-art model described above achieved a mean Pearson’s r = 0.54. While we can have
better average performance in 13 tasks, after taking into account confidence intervals, we find the
performance is indistinguishable between the two models for both r and R2 regression scores (data
not shown).
Overall, this indicates that our QSOR modeling approach can generalize to adjacent perceptual
tasks, and captures meaningful and useful structure about human olfactory perception, even when
measured in different contexts, with different methodologies.

5 Conclusion
We assembled a novel and large dataset of expertly-labeled single-molecule odorants, and trained
a graph neural network to predict the relationship between a molecules structure and its smell. We
demonstrated state-of-the-art results on this QSOR task with respect to field-recognized baselines.
Further, we showed that the embeddings capture meaningful structure on both a local and global
scale. Finally, we showed that the embeddings learned by our model are useful in downstream tasks,
which is currently a rare property of modern machine learning models and data in chemistry. Thus,
we believe our model and its learned embeddings might be generally useful in the rational design of
new odorants. For an expanded version of this preprint, please see (13).

6 Acknowledgments
We thank D. Sculley, Steven Kearnes, Jasper Snoek, Emily Reif, Carey Radebaugh, David Belanger,
Joel Mainland and Emily Mayhew for support, suggestions and useful discussions on the manuscript.
We thank Max Bileschi and Yoni Halpern for their technical guidance. We thank Aniket Zinzuwadia
for discussion and who did preliminary work on the DREAM and GoodScents datasets for his senior
thesis. We thank Bill Luebke of GoodScents and John Leffingwell of Leffingwell and Associates for
their generosity in sharing their data for research use. We thank all of our colleagues in the Google
Brain Cambridge office for creating and maintaining such a supportive and stimulating environment.

References
[1] C S Sell. On the unpredictability of odor. Angewandte Chemie International Edition,
45(38):6254–6261, 2006.
[2] Dagmar Stumpfe and Jürgen Bajorath. Exploring activity cliffs in medicinal chemistry. J. Med.
Chem., 55(7):2932–2942, April 2012.
[3] Han Altae-Tran, Bharath Ramsundar, Aneesh S Pappu, and Vijay Pande. Low data drug dis-
covery with One-Shot learning. ACS Cent Sci, 3(4):283–293, April 2017.
[4] Clyde Fare, Lukas Turcani, and Edward O Pyzer-Knapp. Powerful, transferable representations
for molecules through intelligent task selection in deep multitask networks. September 2018,
doi:1809.06334.
[5] Günther Ohloff, Wilhelm Pickenhagen, and Philip Kraft. Scent and Chemistry. Wiley, January
2012.
[6] The good scents company - flavor, fragrance, food and cosmetics ingredients information.
http://www.thegoodscentscompany.com/. Accessed: 2019-9-4.

5
 [7] John C Leffingwell. Leffingwell & associates, 2005.
[8] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neu-
ral message passing for quantum chemistry. In Doina Precup and Yee Whye Teh, editors,
Proceedings of the 34th International Conference on Machine Learning, volume 70 of Pro-
ceedings of Machine Learning Research, pages 1263–1272, International Convention Centre,
Sydney, Australia, 2017. PMLR.
[9] David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Gómez-Bombarell, Tim-
othy Hirzel, Alán Aspuru-Guzik, and Ryan P Adams. Convolutional Networks on Graphs
for Learning Molecular Fingerprints. In Advances in Neural Information Processing Systems,
pages 2215–2223, 2015.
[10] Hirotomo Moriwaki, Yu-Shi Tian, Norihito Kawashita, and Tatsuya Takagi. Mordred: a molec-
ular descriptor calculator. J. Cheminform., 10(1):4, February 2018.
[11] S J Pan and Q Yang. A survey on transfer learning. IEEE Trans. Knowl. Data Eng.,
22(10):1345–1359, October 2010.
[12] Andreas Keller, Richard C Gerkin, Yuanfang Guan, Amit Dhurandhar, Gabor Turu, Bence
Szalai, Joel D Mainland, Yusuke Ihara, Chung Wen Yu, Russ Wolfinger, Celine Vens, Leander
Schietgat, Kurt De Grave, Raquel Norel, DREAM Olfaction Prediction Consortium, Gustavo
Stolovitzky, Guillermo A Cecchi, Leslie B Vosshall, and Pablo Meyer. Predicting human
olfactory perception from chemical features of odor molecules. Science, 355(6327):820–826,
February 2017.
[13] Benjamin Sanchez-Lengeling, Jennifer N Wei, Brian K Lee, Richard C Gerkin, Alán Aspuru-
Guzik, and Alexander B Wiltschko. Machine learning for scent: Learning generalizable per-
ceptual representations of small molecules. October 2019, doi:1910.10685.