arxiv: 2604.10452 · v2 · submitted 2026-04-12 · 💻 cs.CL

Recognition: unknown

NOSE: Neural Olfactory-Semantic Embedding with Tri-Modal Orthogonal Contrastive Learning

Yanyi Su , Hongshuai Wang , Zhifeng Gao , Jun Cheng

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:05 UTC · model grok-4.3

classification 💻 cs.CL

keywords olfactory embeddingstri-modal contrastive learningmolecular structurereceptor sequencessemantic alignmentzero-shot generalizationweak positive sampling

0 comments

The pith

NOSE aligns molecular structures, receptor sequences, and language descriptions into embeddings that match human olfactory intuition.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to fix fragmented models that capture only isolated parts of how smells work, from chemical molecules to receptors to words. It proposes NOSE as a single framework that pulls these three modalities together through contrastive learning. Orthogonal constraints keep each modality's distinct information intact instead of mixing them, while a weak positive sample method handles the fact that smell language is sparse and avoids wrongly pushing similar odors apart. The result is state-of-the-art performance on standard tasks plus strong zero-shot results on new odors, which would matter because it supplies a more biologically grounded and semantically usable representation of smell.

Core claim

NOSE is a tri-modal representation learning method that aligns molecular structure, receptor sequence, and natural language description along the olfactory pathway. It applies orthogonal constraints to decouple the contributions of each modality and preserve their unique encoded information, paired with a weak positive sample strategy that calibrates semantic similarity to prevent erroneous repulsion of similar odors in the embedding space. This produces representations that reach state-of-the-art performance and exhibit excellent zero-shot generalization, demonstrating close alignment with human olfactory intuition.

What carries the argument

Tri-modal orthogonal contrastive learning combined with a weak positive sample strategy, which uses orthogonality to separate modality-specific signals while still aligning them across the pathway and uses weak positives to manage sparse linguistic data.

If this is right

The embeddings support zero-shot prediction of language descriptions from molecular or receptor inputs alone.
Each modality retains distinct information, allowing more interpretable use in downstream tasks such as odor classification or similarity search.
Joint modeling of the full molecule-to-perception chain outperforms methods that use only one or two modalities.
The approach scales to new odors without retraining, enabling broader coverage of the olfactory space.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar orthogonal decoupling could be tested on other sensory domains where linguistic descriptions are sparse, such as taste or texture.
The representations might improve virtual simulation of odors by providing a shared space that links chemistry to perception.
If the alignment holds, the model could serve as a foundation for predicting how molecular changes affect perceived smell without new human labeling.

Load-bearing premise

That orthogonal constraints can separate each modality's unique information without blocking meaningful cross-modal alignment, and that the weak positive sample strategy correctly groups similar odors without adding bias or wrong repulsions.

What would settle it

A direct comparison showing that NOSE places known similar odors farther apart in embedding space than dissimilar ones, or that its zero-shot accuracy on new odor descriptions falls below standard baselines, would disprove the alignment claim.

Figures

Figures reproduced from arXiv: 2604.10452 by Hongshuai Wang, Jun Cheng, Yanyi Su, Zhifeng Gao.

**Figure 1.** Figure 1: Subscripts r and d denote receptor and description modalities, respectively. (Left) Multimodal Orthogonal Pre-training: Molecular representations zmol are extracted by a frozen Uni-Mol encoder, receptor embeddings zrec are obtained from ESM-2 with a trainable projection layer, and odor semantic descriptions zdesc are extracted by LoRA-finetuned Qwen3 Embedding after LLM-based weak positive augmentation. Th… view at source ↗

**Figure 2.** Figure 2: Vector Space Visualization. For a given SMILES input, NOSE generates vectors in three orthogonal [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗

**Figure 3.** Figure 3: (a) Qwen3 Embedding without LoRA fine-tuning lack olfactory semantics. (b) After contrastive learning, [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗

read the original abstract

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

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

NOSE puts forward a tri-modal orthogonal contrastive setup for molecule-receptor-language alignment in olfaction, but the SOTA and human-intuition claims rest on experiments not visible in the abstract.

read the letter

The main new piece here is the tri-modal framework that treats molecular structure, receptor sequences, and odor descriptions as separate modalities, then uses orthogonal constraints to keep their contributions distinct while still aligning them contrastively. The weak-positive sampling step for handling sparse semantic labels is a practical addition that avoids pushing similar odors apart too aggressively. Releasing the code is also a plus for anyone who wants to test the idea directly. Those elements look like a coherent attempt to model the full olfactory pathway rather than isolated parts. The abstract does not include any numbers, baselines, ablation results, or dataset details, so the claims of state-of-the-art performance and strong zero-shot generalization cannot be checked yet. The further assertion that the space confirms alignment with human olfactory intuition is especially thin without psychophysical validation data such as human similarity ratings. The orthogonal constraint and weak-positive threshold are free parameters that could easily drive the reported behavior, and the circularity risk noted in the stress-test note still applies. This work is aimed at groups doing multimodal representation learning for sensory or biological data. A reader already working on contrastive methods for chemistry or language might pick up the architecture or the sampling trick. The paper deserves a serious referee to examine the full experimental section and see whether the quantitative support matches the abstract language.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces NOSE, a tri-modal representation learning framework that aligns molecular structure, receptor sequence, and natural language description embeddings for olfaction via orthogonal contrastive learning and a weak positive sample strategy to address sparse linguistic data. It claims state-of-the-art performance and strong zero-shot generalization that confirms alignment between the learned space and human olfactory intuition.

Significance. If the results hold, the work offers a more integrated model of the olfactory pathway than prior unimodal or loosely fused approaches, with potential implications for multimodal sensory AI. The public release of code and data at the cited GitHub repository is a clear strength that supports reproducibility and follow-on work.

major comments (3)

[Abstract] Abstract: The central claims of SOTA performance and excellent zero-shot generalization are asserted without any quantitative metrics, baselines, ablation results, or dataset details. This gap is load-bearing because the significance and the inference to human olfactory intuition rest entirely on these unshown experiments.
[Method] Method section (orthogonal constraints and weak positive strategy): The decoupling via orthogonality is presented as preserving unique modality information while enabling alignment, yet no analysis or ablation demonstrates that the constraints do not collapse meaningful cross-modal signal or that the similarity threshold in weak positives avoids erroneous repulsion. This directly affects the soundness of the tri-modal objective.
[Experiments] Experiments section: The claim that SOTA results confirm 'strong alignment ... with human olfactory intuition' lacks any direct perceptual validation (e.g., correlation with human odor similarity ratings or identification accuracy). Technical superiority on molecular/receptor/language tasks does not by itself establish human-like geometry in the embedding space.

minor comments (2)

[Abstract] The abstract would be strengthened by including one or two key quantitative results (e.g., accuracy or retrieval metrics) to allow readers to immediately gauge the claimed SOTA gains.
[Method] Notation for the three modality encoders and the orthogonality loss weight should be introduced consistently in the method section to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback, which highlights important areas for strengthening the presentation and validation of our work. We address each major comment point by point below, indicating planned revisions where appropriate.

read point-by-point responses

Referee: [Abstract] Abstract: The central claims of SOTA performance and excellent zero-shot generalization are asserted without any quantitative metrics, baselines, ablation results, or dataset details. This gap is load-bearing because the significance and the inference to human olfactory intuition rest entirely on these unshown experiments.

Authors: We agree that the abstract would benefit from concrete quantitative support to make the claims immediately verifiable. In the revised manuscript, we will expand the abstract to include specific performance metrics (such as accuracy or similarity scores on key benchmarks), explicit comparisons to baselines, references to ablation results, and brief dataset details. This revision will directly address the load-bearing nature of the claims while preserving the abstract's conciseness. revision: yes
Referee: [Method] Method section (orthogonal constraints and weak positive strategy): The decoupling via orthogonality is presented as preserving unique modality information while enabling alignment, yet no analysis or ablation demonstrates that the constraints do not collapse meaningful cross-modal signal or that the similarity threshold in weak positives avoids erroneous repulsion. This directly affects the soundness of the tri-modal objective.

Authors: We acknowledge the value of empirical validation for these design choices. The method section provides the theoretical rationale for orthogonality (to retain modality-specific information) and the weak-positive threshold (to mitigate data sparsity without introducing false negatives). To strengthen this, we will add dedicated ablation studies in the revised version, including performance comparisons with and without the orthogonal constraints, as well as sensitivity analysis over the similarity threshold, to confirm that cross-modal alignment is preserved and erroneous repulsion is avoided. revision: yes
Referee: [Experiments] Experiments section: The claim that SOTA results confirm 'strong alignment ... with human olfactory intuition' lacks any direct perceptual validation (e.g., correlation with human odor similarity ratings or identification accuracy). Technical superiority on molecular/receptor/language tasks does not by itself establish human-like geometry in the embedding space.

Authors: The zero-shot generalization results are evaluated on olfactory tasks whose ground truth derives from human perceptual annotations (e.g., odor description and similarity judgments). This provides indirect but task-relevant evidence that the learned geometry aligns with human intuition. We agree that an explicit correlation analysis with human similarity ratings would offer stronger direct validation. In the revision, we will moderate the language around 'confirming' alignment to 'supporting' or 'suggesting' alignment, add a limitations discussion clarifying the indirect nature of the evidence, and include any feasible correlation analysis using publicly available human rating datasets. revision: partial

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper defines NOSE via standard tri-modal contrastive learning augmented by orthogonal constraints and a weak-positive sampling heuristic. Claims of SOTA performance and zero-shot generalization are supported by experimental results on held-out molecular, receptor, and language benchmarks rather than by any self-referential reduction. No equations or sections reduce a claimed prediction to a fitted parameter or to a self-citation chain; the alignment inference with human intuition is an interpretive step resting on external benchmark outcomes, not a definitional tautology. The architecture and loss are constructed from first principles of contrastive objectives and modality decoupling, with no load-bearing self-citations or imported uniqueness theorems.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard contrastive learning assumptions plus two paper-specific choices: the orthogonality constraint strength and the definition of weak positives. No new physical entities are postulated.

free parameters (2)

Orthogonality constraint weight
Hyperparameter controlling how strictly the three modality embeddings are forced to be orthogonal; value chosen to balance decoupling and alignment.
Weak positive similarity threshold
Threshold or weighting used to decide which odor pairs count as weak positives; fitted or tuned on olfactory data.

axioms (2)

domain assumption The three modalities (molecular graph, receptor sequence, language) contain complementary but non-redundant information about olfaction that can be aligned in a shared space.
Invoked in the introduction and method sections to justify tri-modal fusion.
standard math Standard contrastive loss with orthogonal regularization preserves unique modality information while enabling cross-modal retrieval.
Background assumption from prior multimodal contrastive work.

pith-pipeline@v0.9.0 · 5493 in / 1503 out tokens · 49299 ms · 2026-05-10T16:05:16.207787+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

74 extracted references · 11 canonical work pages · 3 internal anchors

[1]

Michael H Abraham, Ricardo S \'a nchez-Moreno, J Enrique Cometto-Mu \ n iz, and William S Cain. 2012. https://doi.org/10.1093/chemse/bjr094 An algorithm for 353 odor detection thresholds in humans . Chemical senses, 37(3):207--218

work page doi:10.1093/chemse/bjr094 2012
[2]

T Acree. 2004. http://www.flavornet.org/flavornet.html Flavornet and human odor space . http://www.flavornet.org/flavornet.html

2004
[3]

Lucky Ahmed, Yuetian Zhang, Eric Block, Michael Buehl, Michael J Corr, Rodrigo A Cormanich, Sivaji Gundala, Hiroaki Matsunami, David O’Hagan, Mehmet Ozbil, and 1 others. 2018. https://doi.org/10.1073/pnas.1713026115 Molecular mechanism of activation of human musk receptors or5an1 and or1a1 by (r)-muscone and diverse other musk-smelling compounds . Proceed...

work page doi:10.1073/pnas.1713026115 2018
[4]

Steffen Arctander. 2017. Perfume and flavor materials of natural origin. Lulu.com

2017
[5]

Christian B Billesb lle, Claire A de March, Wijnand JC van der Velden, Ning Ma, Jeevan Tewari, Claudia Llinas Del Torrent, Linus Li, Bryan Faust, Nagarajan Vaidehi, Hiroaki Matsunami, and 1 others. 2023. https://doi.org/10.1038/s41586-023-05798-y Structural basis of odorant recognition by a human odorant receptor . Nature, 615(7953):742--749

work page doi:10.1038/s41586-023-05798-y 2023
[6]

Linda Buck and Richard Axel. 1991. A novel multigene family may encode odorant receptors: a molecular basis for odor recognition. Cell, 65(1):175--187

1991
[7]

Linda B Buck. 2004. Olfactory receptors and odor coding in mammals. Nutrition reviews, 62(suppl\_3):S184--S188

2004
[8]

Caroline Bushdid, Marcelo O Magnasco, Leslie B Vosshall, and Andreas Keller. 2014. https://doi.org/10.1126/science.1249168 Humans can discriminate more than 1 trillion olfactory stimuli . Science, 343(6177):1370--1372

work page doi:10.1126/science.1249168 2014
[9]

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PmLR

2020
[10]

Seyone Chithrananda, Judith Amores, and Kevin K Yang. 2024. Mapping the combinatorial coding between olfactory receptors and perception with deep learning. bioRxiv, pages 2024--09

2024
[11]

Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. 2020. Chemberta: large-scale self-supervised pretraining for molecular property prediction. arXiv preprint arXiv:2010.09885

work page arXiv 2020
[12]

Chulwon Choi, Jungnam Bae, Seonghan Kim, Seho Lee, Hyunook Kang, Jinuk Kim, Injin Bang, Kiheon Kim, Won-Ki Huh, Chaok Seok, and 1 others. 2023. Understanding the molecular mechanisms of odorant binding and activation of the human or52 family. Nature Communications, 14(1):8105

2023
[13]

Neelansh Garg, Apuroop Sethupathy, Rudraksh Tuwani, Rakhi Nk, Shubham Dokania, Arvind Iyer, Ayushi Gupta, Shubhra Agrawal, Navjot Singh, Shubham Shukla, and 1 others. 2018. Flavordb: a database of flavor molecules. Nucleic acids research, 46(D1):D1210--D1216

2018
[14]

Christiane Geithe, Gaby Andersen, Agne Malki, and Dietmar Krautwurst. 2015. A butter aroma recombinate activates human class-i odorant receptors. Journal of agricultural and food chemistry, 63(43):9410--9420

2015
[15]

Rosa S Gisladottir, Erna V Ivarsdottir, Agnar Helgason, Lina Jonsson, Nanna K Hannesdottir, Gudrun Rutsdottir, Gudny A Arnadottir, Astros Skuladottir, Benedikt A Jonsson, Gudmundur L Norddahl, and 1 others. 2020. Sequence variants in taar5 and other loci affect human odor perception and naming. Current Biology, 30(23):4643--4653

2020
[16]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, and 1 others. 2025. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633--638

2025
[17]

Ria Gupta, Aayushi Mittal, Vishesh Agrawal, Sushant Gupta, Krishan Gupta, Rishi Raj Jain, Prakriti Garg, Sanjay Kumar Mohanty, Riya Sogani, Harshit Singh Chhabra, and 1 others. 2021. Odorify: a conglomerate of artificial intelligence--driven prediction engines for olfactory decoding. Journal of Biological Chemistry, 297(2)

2021
[18]

Franziska Haag, Antonella Di Pizio, and Dietmar Krautwurst. 2022. The key food odorant receptive range of broadly tuned receptor or2w1. Food Chemistry, 375:131680

2022
[19]

Elizabeth A Hamel, Jason B Castro, Travis J Gould, Robert Pellegrino, Zhiwei Liang, Liyah A Coleman, Famesh Patel, Derek S Wallace, Tanushri Bhatnagar, Joel D Mainland, and 1 others. 2024. Pyrfume: A window to the world’s olfactory data. Scientific Data, 11(1):1220

2024
[20]

Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. 2020. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM international conference on multimedia, pages 1122--1131

2020
[21]

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

2022
[22]

International Fragrance Association (IFRA) . 2025. IFRA Fragrance Ingredient Glossary . https://ifrafragrance.org/publications/guidance-reference-document/ifra-fragrance-ingredient-glossary. Accessed: December 10, 2025

2025
[23]

Yongquan Jiang, Xin Xie, Yan Yang, Yuerui Liu, Kuanping Gong, and Tianrui Li. 2025. Dual-branch graph neural network for predicting molecular odors and discovering the relationship between functional groups and odors. Journal of Computational Chemistry, 46(6):e70069

2025
[24]

Andreas Keller, Richard C Gerkin, Yuanfang Guan, Amit Dhurandhar, Gabor Turu, Bence Szalai, Joel D Mainland, Yusuke Ihara, Chung Wen Yu, Russ Wolfinger, and 1 others. 2017. Predicting human olfactory perception from chemical features of odor molecules. Science, 355(6327):820--826

2017
[25]

Andreas Keller and Leslie B Vosshall. 2016. Olfactory perception of chemically diverse molecules. BMC neuroscience, 17(1):55

2016
[26]

TN Kipf. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

work page internal anchor Pith review Pith/arXiv arXiv 2016
[27]

Yogesh Kumar, Om Prakash, Himanshu Tripathi, Sudeep Tandon, Madan M Gupta, Laiq-Ur Rahman, Raj K Lal, Manoj Semwal, Mahendra Pandurang Darokar, and Feroz Khan. 2018. Aromadb: a database of medicinal and aromatic plant’s aroma molecules with phytochemistry and therapeutic potentials. Frontiers in plant science, 9:1081

2018
[28]

Murathan Kurfal , Pawel Herman, Stephen Pierzchajlo, Jonas Olofsson, and Thomas H \"o rberg. 2025. Representations of smells: The next frontier for language models? Cognition, 264:106243

2025
[29]

Maxence Lalis, Matej Hladi s , Samar Abi Khalil, Lo \" c Briand, S \'e bastien Fiorucci, and J \'e r \'e mie Topin. 2024. M2or: a database of olfactory receptor--odorant pairs for understanding the molecular mechanisms of olfaction. Nucleic Acids Research, 52(D1):D1370--D1379

2024
[30]

Hadas Lapid, Sagit Shushan, Anton Plotkin, Hillary Voet, Yehudah Roth, Thomas Hummel, Elad Schneidman, and Noam Sobel. 2011. Neural activity at the human olfactory epithelium reflects olfactory perception. Nature neuroscience, 14(11):1455--1461

2011
[31]

Brian K Lee, Emily J Mayhew, Benjamin Sanchez-Lengeling, Jennifer N Wei, Wesley W Qian, Kelsie A Little, Matthew Andres, Britney B Nguyen, Theresa Moloy, Jacob Yasonik, and 1 others. 2023. https://doi.org/10.1126/science.ade4401 A principal odor map unifies diverse tasks in olfactory perception . Science, 381(6661):999--1006

work page doi:10.1126/science.ade4401 2023
[32]

Paul Pu Liang, Zihao Deng, Martin Q Ma, James Y Zou, Louis-Philippe Morency, and Ruslan Salakhutdinov. 2023. Factorized contrastive learning: Going beyond multi-view redundancy. Advances in Neural Information Processing Systems, 36:32971--32998

2023
[33]

Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, and 1 others. 2023. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123--1130

2023
[34]

Shengzhong Liu, Tomoyoshi Kimura, Dongxin Liu, Ruijie Wang, Jinyang Li, Suhas Diggavi, Mani Srivastava, and Tarek Abdelzaher. 2023. Focal: Contrastive learning for multimodal time-series sensing signals in factorized orthogonal latent space. Advances in Neural Information Processing Systems, 36:47309--47338

2023
[35]

Yue Ma, Ke Tang, Yan Xu, and Thierry Thomas-Danguin. 2021. A dataset on odor intensity and odor pleasantness of 222 binary mixtures of 72 key food odorants rated by a sensory panel of 30 trained assessors. Data in brief, 36:107143

2021
[36]

Joel D Mainland, Yun R Li, Ting Zhou, Wen Ling L Liu, and Hiroaki Matsunami. 2015. Human olfactory receptor responses to odorants. Scientific data, 2(1):1--9

2015
[37]

Bettina Malnic, Junzo Hirono, Takaaki Sato, and Linda B Buck. 1999. Combinatorial receptor codes for odors. Cell, 96(5):713--723

1999
[38]

Emily J Mayhew, Charles J Arayata, Richard C Gerkin, Brian K Lee, Jonathan M Magill, Lindsey L Snyder, Kelsie A Little, Chung Wen Yu, and Joel D Mainland. 2022. Transport features predict if a molecule is odorous. Proceedings of the National Academy of Sciences, 119(15):e2116576119

2022
[39]

Grant D McConachie, Emily Duniec, Florence Guerina, Meg A Younger, and Brian DePasquale. 2025. Low rank adaptation of chemical foundation models generates effective odorant representations. bioRxiv, pages 2025--11

2025
[40]

Guillaume Ollitrault, Rayane Achebouche, Antoine Dreux, Samuel Murail, Karine Audouze, Anne Tromelin, and Olivier Taboureau. 2024. Pred-o3, a web server to predict molecules, olfactory receptors and odor relationships. Nucleic Acids Research, 52(W1):W507--W512

2024
[41]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, and 1 others. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748--8763. PmLR

2021
[42]

Aharon Ravia, Kobi Snitz, Danielle Honigstein, Maya Finkel, Rotem Zirler, Ofer Perl, Lavi Secundo, Christophe Laudamiel, David Harel, and Noam Sobel. 2020. A measure of smell enables the creation of olfactory metamers. Nature, 588(7836):118--123

2020
[43]

David Rogers and Mathew Hahn. 2010. Extended-connectivity fingerprints. Journal of chemical information and modeling, 50(5):742--754

2010
[44]

Vivek Sagar, Laura K Shanahan, Christina M Zelano, Jay A Gottfried, and Thorsten Kahnt. 2023. High-precision mapping reveals the structure of odor coding in the human brain. Nature neuroscience, 26(9):1595--1602

2023
[45]

Benjamin Sanchez-Lengeling, Jennifer N Wei, Brian K Lee, Richard C Gerkin, Al \'a n Aspuru-Guzik, and Alexander B Wiltschko. 2019. Machine learning for scent: Learning generalizable perceptual representations of small molecules. arXiv preprint arXiv:1910.10685

work page arXiv 2019
[46]

Charles S Sell. 2006. On the unpredictability of odor. Angewandte Chemie International Edition, 45(38):6254--6261

2006
[47]

Anju Sharma, Rajnish Kumar, Shabnam Ranjta, and Pritish Kumar Varadwaj. 2021. Smiles to smell: decoding the structure--odor relationship of chemical compounds using the deep neural network approach. Journal of Chemical Information and Modeling, 61(2):676--688

2021
[48]

Anju Sharma, Bishal Kumar Saha, Rajnish Kumar, and Pritish Kumar Varadwaj. 2022. Olfactionbase: a repository to explore odors, odorants, olfactory receptors and odorant--receptor interactions. Nucleic Acids Research, 50(D1):D678--D686

2022
[49]

Mrityunjay Sharma, Sarabeshwar Balaji, Pinaki Saha, and Ritesh Kumar. 2025. Navigating the fragrance space using graph generative models and predicting odors. Journal of Chemical Information and Modeling, 65(10):4818--4832

2025
[50]

Daniel Shin, Gao Pei, Priyadarshini Kumari, and Tarek R Besold. 2023. Optimizing learning across multimodal transfer features for modeling olfactory perception. In 29th ACM SIGKDD Cof. on Knowledge Discovery and Data Mining

2023
[51]

Sigma-Aldrich . 2025. Sigma-Aldrich Website . https://www.sigmaaldrich.com/SG/en. Accessed: December 10, 2025

2025
[52]

Kobi Snitz, Adi Yablonka, Tali Weiss, Idan Frumin, Rehan M Khan, and Noam Sobel. 2013. Predicting odor perceptual similarity from odor structure. PLoS computational biology, 9(9):e1003184

2013
[53]

Noam Sobel, Vivek Prabhakaran, John E Desmond, Gary H Glover, RL Goode, Edith V Sullivan, and John DE Gabrieli. 1998. Sniffing and smelling: separate subsystems in the human olfactory cortex. Nature, 392(6673):282--286

1998
[54]

Chih-Ying Su, Karen Menuz, and John R Carlson. 2009. Olfactory perception: receptors, cells, and circuits. Cell, 139(1):45--59

2009
[55]

Farzaneh Taleb, Miguel Vasco, Antonio Ribeiro, M rten Bj \"o rkman, and Danica Kragic. 2024. Can transformers smell like humans? Advances in Neural Information Processing Systems, 37:72032--72060

2024
[56]

The Good Scents Company . 2025. The Good Scents Company Information System . https://www.thegoodscentscompany.com/. Accessed: December 10, 2025

2025
[57]

Gary Tom, Cher Tian Ser, Ella M Rajaonson, Stanley Lo, Hyun Suk Park, Brian K Lee, and Benjamin Sanchez-Lengeling. 2025. Does this smell the same? learning representations of olfactory mixtures using inductive biases. Machine Learning: Science and Technology, 6(3):035063

2025
[58]

Hugo Touvron, Piotr Bojanowski, Mathilde Caron, Matthieu Cord, Alaaeldin El-Nouby, Edouard Grave, Gautier Izacard, Armand Joulin, Gabriel Synnaeve, Jakob Verbeek, and 1 others. 2022. Resmlp: Feedforward networks for image classification with data-efficient training. IEEE transactions on pattern analysis and machine intelligence, 45(4):5314--5321

2022
[59]

Ngoc Tran, Daniel Kepple, Sergey Shuvaev, and Alexei Koulakov. 2019. Deepnose: Using artificial neural networks to represent the space of odorants. In International Conference on Machine Learning, pages 6305--6314. PMLR

2019
[60]

Yuta Wakutsu and Hiromasa Kaneko. 2025. Molecular odor prediction using olfactory receptor information. Molecular Informatics, 44(3):e202400274

2025
[61]

Ivonne Wallrabenstein, Jonas Kuklan, Lea Weber, Sandra Zborala, Markus Werner, Janine Altm \"u ller, Christian Becker, Anna Schmidt, Hanns Hatt, Thomas Hummel, and 1 others. 2013. Human trace amine-associated receptor taar5 can be activated by trimethylamine. PloS one, 8(2):e54950

2013
[62]

David Weininger. 1988. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of chemical information and computer sciences, 28(1):31--36

1988
[63]

Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. 2020. On layer normalization in the transformer architecture. In International conference on machine learning, pages 10524--10533. PMLR

2020
[64]

Zhaoping Xiong, Dingyan Wang, Xiaohong Liu, Feisheng Zhong, Xiaozhe Wan, Xutong Li, Zhaojun Li, Xiaomin Luo, Kaixian Chen, Hualiang Jiang, and 1 others. 2019. Pushing the boundaries of molecular representation for drug discovery with the graph attention mechanism. Journal of medicinal chemistry, 63(16):8749--8760

2019
[65]

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2018. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826

work page internal anchor Pith review arXiv 2018
[66]

Keiichi Yoshikawa, Jun Deguchi, Jieying Hu, Hsiu-Yi Lu, and Hiroaki Matsunami. 2022. An odorant receptor that senses four classes of musk compounds. Current Biology, 32(23):5172--5179

2022
[67]

Xin Yuan, Zhe Lin, Jason Kuen, Jianming Zhang, Yilin Wang, Michael Maire, Ajinkya Kale, and Baldo Faieta. 2021. Multimodal contrastive training for visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6995--7004

2021
[68]

Shitong Zeng, Lili Zhang, Peng Li, Dandan Pu, Yingjie Fu, Ruiyi Zheng, Hui Xi, Kaina Qiao, Dingzhong Wang, Baoguo Sun, and 1 others. 2023. Molecular mechanisms of caramel-like odorant-olfactory receptor interactions based on a computational chemistry approach. Food Research International, 171:113063

2023
[69]

Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, and 1 others. 2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176

work page internal anchor Pith review Pith/arXiv arXiv 2025
[70]

Shu Zhong, Zetao Zhou, Christopher Dawes, Giada Brianz, and Marianna Obrist. 2024. Sniff ai: Is my'spicy'your'spicy'? exploring llm's perceptual alignment with human smell experiences. arXiv preprint arXiv:2411.06950

work page arXiv 2024
[71]

Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. 2023. Uni-mol: A universal 3d molecular representation learning framework. In The eleventh international conference on learning representations

2023
[72]

Pan Zhou, Caiming Xiong, Xiaotong Yuan, and Steven Chu Hong Hoi. 2021. A theory-driven self-labeling refinement method for contrastive representation learning. Advances in Neural Information Processing Systems, 34:6183--6197

2021
[73]

online" 'onlinestring :=

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...
[74]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...