Today, we are announcing Sense1, our first Scent Foundation Model. This is the first model to accurately predict olfactory receptor activation across the complete set of human receptors. From there, it maps those activations to perceptual qualities.
Among all the ways we experience the world, our sense of smell is the most directly linked to memory and emotion. It is our most ancient sense, and makes up the majority of flavor. Place, mood, physical embodiment: all are shaped by it.
Humans have developed a thorough understanding of vision and sound, but our understanding of scent remains incomplete. Plato observed that “the varieties of smell have no name”: too fleeting to categorize. And yet Rousseau called it “the sense of the imagination,” while Nietzsche claimed “all my genius is in my nostrils.”
So why has it remained unsolved?
The challenge begins with biological complexity. Scent, like all other senses, is built on receptors: specialized proteins that detect a stimulus and convert it into a signal the brain can interpret. Receptors are the interface between the physical world and perception. Our eyes have just 3 types of color receptors (plus a fourth type, rods, for brightness). This simplicity gave us the RGB color model; every screen, camera, and digital image is built on it. Humans can discriminate roughly 10 million colors, and we can reproduce nearly all of them with three channels of light.
Our noses have close to 400 types of olfactory receptors (Marenco et al., Database, 2016), each responding to a different set of molecular features. The number of discriminable olfactory stimuli has been estimated at over one trillion (Bushdid et al., Science, 2014), though the exact figure is debated (Meister, eLife, 2015). This difference (roughly one hundred times more receptor types than color vision, and orders of magnitude more possible stimuli) is the core reason that machine olfaction, the computational study of smell, has lagged so far behind computer vision and audition.
There is also a practical obstacle. Vision is stimulated through electromagnetic waves, and hearing is stimulated through vibrations. Smell requires actual molecules binding to actual receptor proteins. No one has figured out how to encode smells and tastes into a machine-readable form. Without a compression scheme for scent, like MP3 or RGB, there is no “frequency of rose” that can be transmitted electronically.
Olfactory receptors have not been a core area of health research, so experimental data is limited. Recent advances in computational protein modeling have begun to change this, making it possible to predict how odorant molecules interact with receptor proteins without running a physical experiment. Even with these new tools, understanding smell requires combining computational predictions with experimental receptor data and human sensory evaluation: no single approach is sufficient on its own.
Smell has also resisted systematic study because the words used to describe it (like “floral,” “woody,” or “green”) are deeply cultural. Perfumers are trained to anchor these descriptors on known reference materials, but these cover a small subset of olfactory space. Furthermore, different languages carve up the olfactory world differently: some hunter-gatherer communities in Southeast Asia have around a dozen dedicated abstract smell terms, while English speakers struggle to name even common odors (Majid & Burenhult, 2014). Because the vocabulary itself is not universally consistent, it cannot serve as a universal foundation.
Receptors can.
The receptor code works universally across humans, operating beneath culture at the level of shared biology. Where linguistic descriptors serve as sign posts, the receptor activation pattern is the high-resolution map. Individual receptors contribute identifiable qualities, and their combinations give rise to the full richness of scent.
At the elemental level, individual receptors carry recognizable perceptual identities: molecules that strongly activate the same receptor tend to share similar perceptual aspects, even when chemically unrelated (Raps et al., bioRxiv, 2025). These associations have only been demonstrated for a handful of receptors: never across the full set of ~400. We validate our results against the existing literature in the Elemental Coding Validation section below.
At the combinatorial level, each odorant activates a distinct pattern across many receptors simultaneously: a code first described by Malnic et al. (Cell, 1999). Small structural changes can produce large differences in the pattern (Maggiora, J. Chem. Inf. Model., 2006), an activity cliff such that two molecules that are nearly identical on paper can smell completely different (Sell, Angew. Chem. Int. Ed., 2006).
One finding from our work: the code is not uniform. Even among broad perceptual categories, “fruity” correlates with a small number of receptors, while “floral” involves over a hundred: no single receptor can explain it alone.
The central question driving this work is: how can we advance our understanding of scent to the level that was achieved for vision and hearing over a century ago?
We can trace the progress of machine olfaction in stages, each building on the last:
For most of the history of fragrance and flavor, the only way to characterize a scent was to have a human smell it: a process that is subjective and tells us little about what a new molecule might smell like.
The first computational breakthrough was the ability to predict perceived odor qualities (“floral,” “sweet,” “fruity”) directly from molecular structure (Keller et al., Science, 2017; Lee et al., Science, 2023). This was a significant advance, but it provides only surface-level insights into scent. These models have no concept of the biological machinery that produces perception. They cannot tell you which receptors are involved, why two similar molecules smell completely different, or how a new molecule will interact with the human olfactory system. Predicting what something smells like, without understanding why, is like translating a poem word by word from a language you do not speak: the text might look right, but the meaning is lost.
As computational protein modeling matured, it became possible to predict which of the ~400 olfactory receptors a given molecule will activate, and how strongly. Existing computational approaches, molecular docking and structure prediction models, were not designed for olfactory receptors and fail to transfer to this domain. These methods provide biological ground truth, but do not yet connect to perception.
This stage involves predicting receptor activation from molecular structure, and then mapping those activation patterns to perceptual qualities. At the core is a deep understanding of structure-odor relationships and computational modeling of olfactory receptor activity. This captures the elemental dimension of the code, linking individual receptors and their combinations to perceived smell, though the full combinatorial grammar remains ahead. This is what gives us the biological “why” behind a molecule’s smell, and it is what enables generalization to new molecules, new mixtures, and new applications. Patina is now at this stage.
Smell arises from patterns of receptor activation, but how combinations produce percepts, and how receptors suppress or enhance one another in mixtures, is not yet understood. The grand challenge ahead is a complete grammar of how receptor combinations give rise to the vast space of olfactory experience.
We compared Sense1 against an optimized docking pipeline and a structural biology foundation model. For the docking pipeline, we used AutoDock Vina, with all 389 receptor structures coming from GPCRdb’s AlphaFold predictions, docked in both active and inactive conformations. Our best configuration achieves an AUROC of 0.616; without these optimizations, the raw docking score alone achieves 0.572: barely above chance.
We also evaluated Boltz-2, a structure foundation model that jointly predicts protein-ligand structure and binding affinity. For each receptor-ligand pair, we ran Boltz-2 with pre-computed MSAs from ColabFold. Using Boltz-2’s predicted binding probability as a single-feature agonism predictor yields an overall AUROC of 0.548. Unlike our optimized docking pipeline, Boltz-2 requires no box definition, scoring function selection, or pose post-processing, but its per-prediction cost is ~100x higher than AutoDock Vina.
Sense1 outperforms a state-of-the-art deep-learning model by 65%, and a fully optimized docking baseline by 39%.
Human olfactory receptors are grouped into families (OR1, OR2, etc.) based on genetic similarity. The chart below reports AUROC per family.
* Families marked with an asterisk had only one positive example in the test set. AUROC estimates for these families are inherently high-variance and should be interpreted accordingly.
Three receptor families (OR6, OR9, and OR12) are effectively orphanized, no known molecules in our test set activate them, and are omitted from the chart.
To estimate practical hit rates, we simulated a realistic screening scenario. From our agonist dataset (positive rate 3.7%), we drew 1 million bootstrapped samples of 100 ligand-receptor pairs each and checked whether the top-ranked result in each sample was a true agonist.
With the optimized docking pipeline, the top-1 hit rate was 3.3%: statistically insignificant compared to random selection. With Sense1, the top-1 hit rate was 54%.
Across all evaluable receptor families, Sense1 achieves a significant improvement in accuracy compared to docking and existing deep-learning models. Sense1 outperforms docking on 13 of 15 families, and outperforms Boltz-2 on 12 of 15 families. Performance varies by family, reflecting significant differences in the available scientific data: some families have extensive assay data and well-characterized receptor structures, while others have very little of either. At screening time, Sense1 finds the right receptor-molecule match 54% of the time, where previous methods are no better than random guessing.
Sense1 provides the first large-scale in-silico validation of elemental coding: 18 of 19 published receptor-percept associations confirmed across the complete receptor set.
The elemental encoding theory correlates individual receptors with specific perceptual qualities: if a receptor responds to molecules that smell “musky”, its predicted activation should track the “musky” descriptor across a large set of odorants. To test this, for every molecule with known perceptual descriptors in our dataset, we predicted the full receptor activation pattern using Sense1, and computed the Pearson correlation coefficient between descriptors and activations.
The table below lists receptor–descriptor pairs that have been experimentally validated in published work, alongside the Pearson correlation coefficient and p-value for the Sense1 predictions. Rows with positive r-values indicate that Sense1’s predictions align with the known association. For hierarchical descriptors (e.g., “Green” encompasses “Green.Grass,” “Green.Hay,” etc.), we report the maximum r among the parent and its children, with the corresponding p-value.
| Receptor | Descriptor | Literature association | Citation | Sense1 r | p-value |
|---|---|---|---|---|---|
| OR5AN1 | Musky | Musk compounds | Raps et al.; Takase & Matsunami, AChemS, 2026 | +0.235 | < .05 |
| OR2A25 | Floral | Citronellol, geraniol | Emter et al., Current Biology, 2025 | +0.202 | < .05 |
| OR5A2 | Musky | Diverse musk compounds | Emter et al., Current Biology, 2025 | +0.148 | < .05 |
| OR10G4 | Warm | Guaiacol, vanillin | Mainland et al., Nature Neurosci., 2014 | +0.122 | < .05 |
| OR10G3 | Warm | Vanillin | Emter et al., Current Biology, 2025 | +0.118 | < .05 |
| OR2A25 | Rose | Citronellol | Emter et al., Current Biology, 2025 | +0.112 | < .05 |
| OR2M3 | Sulfurous | Thiol compounds | Noe et al., Chemical Senses, 2017 | +0.109 | < .05 |
| OR1G1 | Green | Broadly tuned | Charlier et al., Cell. Mol. Life Sci., 2012 | +0.105 | < .05 |
| OR1G1 | Mushroom | Broadly tuned | Charlier et al., Cell. Mol. Life Sci., 2012 | +0.088 | < .05 |
| OR51E1 | Dairy | Isovaleric acid | Saito et al., 2009 | +0.083 | < .05 |
| OR10A6 | Lactonic | gamma-decalactone | Emter et al., Current Biology, 2025 | +0.077 | < .05 |
| OR1A1 | Menthol | (R)-(-)-carvone | Geithe et al., Cell. Mol. Life Sci., 2017 | +0.068 | < .05 |
| OR2T11 | Sulfurous | Low-MW thiols | Li et al., JACS, 2016 | +0.065 | < .05 |
| OR5A1 | Violet | beta-ionone | McRae et al., Current Biology, 2013 | +0.059 | < .05 |
| OR2J3 | Green | cis-3-hexen-1-ol | McRae et al., Chemical Senses, 2012 | +0.054 | < .05 |
| OR7D4 | Musky | Androstenone | Keller et al., Nature, 2007 | +0.050 | < .05 |
| OR7A17 | Woody | Arborone | Emter et al., Current Biology, 2025 | +0.046 | < .05 |
| TAAR5 | Fishy | Trimethylamine | Wallrabenstein et al., PLOS ONE, 2013 | +0.038 | < .05 |
| OR51E2 | Violet | beta-ionone | Saito et al., 2009 | -0.038 | < .05 |
Of the 19 unique receptor–descriptor pairs with published experimental evidence, 18 show a positive correlation in Sense1’s predictions and are statistically significant with the expected sign (p < 0.05). The strongest signals (musks, florals, vanillin-associated warmth, and sulfur compounds) are clearly captured. One pair does not match: OR51E2/Violet (r = -0.038, significant but in the wrong direction).
Though correlations are modest in absolute terms (median r ~ 0.08), this is a diagnostic signal. Each receptor responds to many structurally diverse molecules, and each descriptor encompasses a range of perceptual experiences.
The grid below shows Sense1’s predictions across all ~400 human olfactory receptors. Click any receptor to see which perceptual qualities (“floral,” “musky”, “citrus”) are statistically associated with its activation. This is elemental encoding made visible.
Faithful replication of natural materials. Plants record their environment and replay it to us as complex aromas: rose oil alone contains over 300 chemical components, and no synthetic reconstruction has matched it. Understanding what makes a natural material smell the way it does at the receptor level is the key to faithful replication, especially as climate change disrupts supply chains for these materials.
On-demand scent and flavor. The receptor code opens the door to designing scents and flavors to specification: composing new sensory experiences the way a digital artist composes with color, or a musician uses a synthesizer.
Health and biology. Olfactory receptors are not only found in the nose. Ectopic olfactory receptors have been discovered in skin, gut, and tumor tissue (Maßberg & Hatt, Physiological Reviews, 2018), and their biological roles are only beginning to be understood. A receptor-level understanding of scent opens the door to fragrances and cosmetics that do more than smell good: they actively benefit skin health or interact with biological pathways in targeted ways.
Sense1 predicts olfactory receptor activation across the complete human receptor set and maps those activations to perception: achieving an AUROC of 0.854, a 39% improvement over optimized structure-based baselines, and a 16x improvement in screening enrichment. Its predictions align with experimentally validated receptor-percept associations across 18 of 19 published receptor-descriptor pairs.
The receptor code is the missing layer between molecular structure and sensory experience. Sense1 reads it for olfaction; taste will follow. The work ahead, the combinatorial grammar of how receptor patterns give rise to percepts and how receptors interact in mixtures, will require new data, new experimental partnerships, and new modeling approaches. We are making Sense1 available to select research and industry partners to accelerate that work.
Learn more at patina.earth.
Bushdid, C., Magnasco, M. O., Vosshall, L. B., & Keller, A. (2014). Humans can discriminate more than 1 trillion olfactory stimuli. Science, 343(6177), 1370–1372.
Charlier, L., Topin, J., Ronin, C., Kim, S.-K., Goddard, W. A., Efremov, R., & Golebiowski, J. (2012). How broadly tuned olfactory receptors equally recognize their agonists. Human OR1G1 as a test case. Cell. Mol. Life Sci., 69(24), 4205–4213.
Emter, R., Merillat, C., Buchli, F., Flachsmann, F., & Natsch, A. (2025). Decoding human olfaction by high heterologous expression of odorant receptors detecting signature odorants. Current Biology, 35(21), 5252–5263.
Geithe, C., Protze, J., Kreuchwig, F., Krause, G., & Krautwurst, D. (2017). Structural determinants of a conserved enantiomer-selective carvone binding pocket in the human odorant receptor OR1A1. Cell. Mol. Life Sci., 74(22), 4209–4229.
Keller, A., Zhuang, H., Chi, Q., Vosshall, L. B., & Matsunami, H. (2007). Genetic variation in a human odorant receptor alters odour perception. Nature, 449(7161), 468–472.
Keller, A., Gerkin, R. C., Guan, Y., Dhurandhar, A., Turu, G., Szalai, B., … & Meyer, P. (2017). Predicting human olfactory perception from chemical features of odor molecules. Science, 355(6327), 820–826.
Lee, B. K., Mayhew, E. J., Sanchez-Lengeling, B., Wei, J. N., Qian, W. W., Little, K. A., … & Wiltschko, A. B. (2023). A principal odor map unifies diverse tasks in olfactory perception. Science, 381(6661), 999–1006.
Li, S., Ahmed, L., Zhang, R., Pan, Y., Matsunami, H., Burger, J. L., Block, E., Batista, V. S., & Zhuang, H. (2016). Smelling sulfur: copper and silver regulate the response of human odorant receptor OR2T11 to low-molecular-weight thiols. J. Am. Chem. Soc., 138(40), 13281–13288.
Maggiora, G. M. (2006). On outliers and activity cliffs — why QSAR often disappoints. J. Chem. Inf. Model., 46(4), 1535.
Mainland, J. D., Keller, A., Li, Y. R., Zhou, T., Trimmer, C., Snyder, L. L., … & Matsunami, H. (2014). The missense of smell: functional variability in the human odorant receptor repertoire. Nature Neurosci., 17(1), 114–120.
Majid, A., & Burenhult, N. (2014). Odors are expressible in language, as long as you speak the right language. Cognition, 130(2), 266–270.
Malnic, B., Hirono, J., Sato, T., & Buck, L. B. (1999). Combinatorial receptor codes for odors. Cell, 96(5), 713–723.
Marenco, L., Wang, R., McDougal, R., Olender, T., Twik, M., Bruford, E., … & Crasto, C. (2016). ORDB, HORDE, ODORactor and other on-line knowledge resources of olfactory receptor-odorant interactions. Database, 2016, baw132.
Maßberg, D., & Hatt, H. (2018). Human olfactory receptors: novel cellular functions outside of the nose. Physiological Reviews, 98(3), 1739–1763.
McRae, J. F., Mainland, J. D., Jaeger, S. R., Adipietro, K. A., Matsunami, H., & Newcomb, R. D. (2012). Genetic variation in the odorant receptor OR2J3 is associated with the ability to detect the “grassy” smelling odor, cis-3-hexen-1-ol. Chemical Senses, 37(7), 585–593.
McRae, J. F., Jaeger, S. R., Bava, C. M., Beresford, M. K., Hunter, D., … & Newcomb, R. D. (2013). Identification of regions associated with variation in sensitivity to food-related odors in the human genome. Current Biology, 23(16), 1596–1600.
Meister, M. (2015). On the dimensionality of odor space. eLife, 4, e07865.
Noe, F., Polster, J., Geithe, C., Kotthoff, M., Schieberle, P., & Krautwurst, D. (2017). OR2M3: a highly specific and narrowly tuned human odorant receptor for the sensitive detection of onion key food odorant 3-mercapto-2-methylpentan-1-ol. Chemical Senses, 42(3), 195–210.
Raps, D. A., Pierce, G. M., Papiani, G., Wu, L., Arroyave, R., Brann, J. H., & Pfister, P. (2025). Evidence of elemental encoding at the olfactory periphery. bioRxiv, 2025.04.21.649882.
Saito, H., Chi, Q., Zhuang, H., Matsunami, H., & Mainland, J. D. (2009). Odor coding by a mammalian receptor repertoire. Science Signaling, 2(60), ra9.
Sell, C. S. (2006). On the unpredictability of odor. Angew. Chem. Int. Ed., 45(38), 6254–6261.
Takase, D., & Matsunami, H. (2026). Structural exploration of musk recognition and activation mechanisms of odorant receptors [Conference poster]. Association for Chemoreception Sciences (AChemS) XLVIII Annual Meeting, St. Pete Beach, FL.
Wallrabenstein, I., Kuklan, J., Weber, L., Zborala, S., Werner, M., … & Gisselmann, G. (2013). Human trace amine-associated receptor TAAR5 can be activated by trimethylamine. PLOS ONE, 8(2), e54950.