Mapping Molecular Structure to Human Olfactory Perception

16 concepts13 papers2 proposals

In vision and audition, the mapping between physical stimuli and human perception is well-behaved. Light wavelengths correspond predictably to color, and sound frequencies map directly to pitch. In olfaction, however, this mapping breaks down. The physical stimulus is a molecule, and the relationship between a molecule's chemical structure and its perceived smell—known as the Structure-Odor Relationship (SOR)—is highly non-linear, complex, and poorly understood (Licon et al., 2019; Zhang et al., 2020). A tiny modification to a molecular graph can transform a pleasant floral scent into a putrid one, while two entirely different chemical structures might smell identical.

To solve this stimulus-percept problem, a growing body of research applies machine learning to decode how chemical features translate into subjective human experiences. The goal is to build computational systems capable of digitizing, predicting, and synthesizing scents (Carmel et al., 2003).

The Sensory Data Scarcity Bottleneck

The primary obstacle to training robust olfactory models is a severe data asymmetry. While modern deep learning thrives on massive datasets, high-quality human sensory data is exceptionally difficult, expensive, and time-consuming to collect at scale (Gutiérrez et al., 2018; Bender et al., 2023). We possess millions of known chemical structures, but only a tiny fraction have been annotated with human perceptual labels (Taleb et al., 2024).

Furthermore, human descriptions of smell are inherently subjective and noisy. To train models that truly ground themselves in human perception, researchers must find clever ways to bridge the gap between abundant, unstructured chemical data and scarce, structured sensory labels.

Engineering the Chemical Input Space

To predict how a molecule smells, a model must first "see" the molecule effectively. Early computational approaches relied on hand-crafted physicochemical descriptors (such as molecular weight, atom counts, or functional groups) paired with Multi-Layer Perceptrons (MLPs) to classify specific odor profiles, such as identifying "green" notes (Bo et al., 2022; Yang et al., 2024).

However, flat descriptors and 2D topological graphs fail to capture a critical biological reality: stereoisomers. Stereoisomers are molecules with the exact same atomic connectivity but different 3D spatial orientations, yet they can trigger entirely different olfactory receptors. To capture these geometric nuances, Tran et al. (2024) introduced DeepNose, a 3D equivariant convolutional neural network. By treating molecules as 3D spatial objects rather than flat graphs, the model inherently respects spatial rotations and reflections, allowing it to successfully distinguish the subtle perceptual differences between stereoisomers.

[2D Molecular Graph]  ──(Loses 3D geometry)──► Struggles with Stereoisomers
[3D Spatial Coordinates] ──(Equivariant CNN)──► Preserves Chirality & Perceptual Nuance

Aligning Chemistry with Human Language

How do we map a 3D chemical representation to a human descriptor like "smoky" or "fruity"? One powerful insight is to use natural language as a proxy for human perception.

Gutiérrez et al. (2018) demonstrated that the semantic relationships between words in everyday language can predict how a mono-molecular odorant will be described. Because the semantic distance between words (e.g., "orange" and "lemon") mirrors the perceptual distance between smells, natural language processing (NLP) can map chemical structures directly onto a continuous "odorwheel" of descriptors, bypassing the need for exhaustive sensory trials.

This alignment can also happen implicitly within large-scale models. Taleb et al. (2024) investigated chemical foundation models (like MoLFormer) that were trained purely on self-supervised molecular structure prediction without any sensory data. They discovered a striking emergent property: the internal representations learned by these foundation models naturally align with human olfactory perception. The model learns the underlying grammar of chemistry, which inherently dictates how those molecules interact with human biological receptors.

For complex, real-world sensory experiences like wine tasting, the data is even noisier, consisting of images, text reviews, and expert ratings. To unify these modalities, the FEAST algorithm developed by Bender et al. (2023) projects wine labels, review text, and sparse human sensory data into a shared, multimodal "flavor space." This allows the model to leverage abundant unstructured data (like online reviews) to ground and predict scarce human perceptual experiences.

From Single Molecules to Complex Mixtures and Intensity

While predicting the scent of a single molecule is a vital step, the real world is composed of complex chemical mixtures. To address this, Tom et al. (2025) developed POMMix. This architecture takes a pre-trained Principal Odor Map (POM) designed for single molecules and extends it to mixtures using graph neural networks and attention mechanisms. The attention layers learn how different molecules mask or enhance one another, allowing the model to accurately predict the perceptual similarity of complex mixtures even when training data is limited.

Beyond the quality of a scent, fragrance design requires predicting its strength. Fichtelmann (2026) tackled this by training machine learning models on ordinal odor strength data, enabling the systematic, computational design of balanced fragrances where components do not overpower one another.

Finally, researchers have realized that a single, global mathematical rule cannot capture the entire structure-odor map. Instead, mining "local exceptional models" (Bosc et al., 2015) or discovering multiple distinct chemical rules for a single odor quality (Licon et al., 2019) allows algorithms to capture the highly contextual, non-linear nature of how our noses interpret chemistry.

The Frontier: True Reasoning vs. Lexical Shortcuts

As large language models (LLMs) are increasingly applied to scientific domains, evaluating their true understanding of olfaction has become critical. Makri et al. (2026) introduced the Olfactory Perception (OP) benchmark to test whether LLMs can reason about smell directly from molecular structures.

The benchmark revealed a key limitation: while LLMs achieve decent accuracy, they primarily rely on memorized lexical associations from compound names rather than true structural molecular reasoning. When given a molecule with a masked or unfamiliar name, their performance drops.

Moving the field forward requires moving past these text-based shortcuts. By tightly coupling 3D geometric chemistry with perceptually aligned embeddings, computational olfaction is steadily transforming the subjective human sense of smell into a predictable, digital modality.

Key Foundational Work

Vers la découverte de modèles exceptionnels locaux : des règles descriptives liant les molécules à leurs odeurs (2015) — contributed 1 concept elements to this direction.
Towards an odor communication system (2003) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sensory_labeling_bottleneck_bd34840e,
  title = {sensory labeling bottleneck},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sensory-labeling-bottleneck/rdr_bd34840e720d475abb9512a71087bb2a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_sensory_labeling_bottleneck_bd34840e_v1,
  title = {sensory labeling bottleneck},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sensory-labeling-bottleneck/rdr_bd34840e720d475abb9512a71087bb2a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake