When we look at a color, our brains process a clear physical metric: the wavelength of light. This wavelength maps predictably to a low-dimensional sensory space and projects onto an organized spatial map in the visual cortex. Hearing operates similarly, mapping sound frequencies to a topographic layout of the auditory cortex.
Olfaction enjoys no such luxury. As Venkatesh N. Murthy notes, the olfactory system lacks a systematic topographic map of odors. There is no single physical metric—like wavelength or frequency—that dictates how a molecule will smell, as highlighted by E. Darío Gutiérrez et al. and Ngoc B. Tran et al.. Instead, the input space consists of millions of potential airborne molecules, each defined by thousands of chemical properties, interacting with hundreds of different receptor types (Alexei A. Koulakov et al.).
To solve this olfactory mapping challenge, a powerful research direction has emerged: using machine learning to discover a continuous, low-dimensional latent manifold that acts as the missing coordinate system for smell.
The Problem of Non-Linearity and Activity Cliffs
Attempting to map chemical structures directly to human odor perception using simple, deterministic rules quickly breaks down. As Guillaume Bosc et al. point out, the structure-odor relationship (SOR) is highly non-linear and complex.
The primary obstacle to a predictable model is the phenomenon of activity cliffs. As detailed by Wesley Wei Qian et al. and Priyadarshini Kumari et al., an activity cliff occurs when an incredibly minor change in a molecule’s chemical structure causes a massive, disproportionate shift in its perceived scent. Conversely, two molecules with completely different chemical backbones can smell nearly identical. Because the raw physical features of a molecule do not scale smoothly with human perception, traditional hand-engineered chemical descriptors struggle to generalize.
Smoothing the Space via Latent Manifolds
To bypass these activity cliffs, researchers have turned to probabilistic and generative models to uncover the hidden, smooth mathematical space underlying neural and behavioral responses.
Instead of mapping a molecule's raw features directly to a verbal label, these models project molecules into a continuous latent space where proximity correlates with perceptual similarity. For instance, Anqi Wu et al. applied a Gaussian Process Latent Variable Model (GPLVM) to neural recordings from the mouse piriform cortex. They demonstrated that even though the input chemical space is fractured, the resulting neural responses sit on a low-dimensional latent manifold that can accurately predict how the brain will respond to entirely unseen odors.
On the structural side, Ngoc B. Tran et al. developed DeepNose, a framework that treats molecules as 3D spatial objects and uses a convolutional autoencoder to learn a low-dimensional representation. By forcing the network to compress and reconstruct 3D molecular structures, the model implicitly creates an in silico ensemble of receptors that matches or outperforms traditional chemical descriptors at predicting human smell.
Why Chemical Foundation Models Can "Smell"
One of the most striking insights from this domain is that deep learning models do not even need to be trained on sensory data to understand smell.
In recent work, Farzaneh Taleb et al. and Farzaneh Taleb et al. demonstrated that transformer foundation models (such as MoLFormer) pre-trained purely on massive databases of chemical structures naturally develop internal representations that align with human olfactory perception.
Why does this happen? The underlying "logic" of smell is deeply embedded within the physical and chemical constraints of molecular structures. Because human olfactory receptors evolved to interact with these exact physical structures, a self-supervised model that masters the grammar of chemistry inadvertently masters the grammar of olfaction. The model's internal latent space naturally mirrors the biological alignment of our own sensory apparatus.
Grounding the Map in Human Feedback
To make these latent spaces useful, they must be anchored to human perception. However, human language is notoriously poor at describing scents objectively, a limitation that complicates the evaluation of artificial olfactory systems (David Harel).
To circumvent the ambiguity of language, researchers use two main strategies:
- Semantic Embedding Alignment: E. Darío Gutiérrez et al. leveraged natural language processing (NLP) to show that the semantic distances between words in everyday language can predict how a molecule will smell, effectively mapping chemical structures to a linguistic "odorwheel."
- Relative Similarity Feedback: Rather than asking subjects to name a smell, Priyadarshini Kumari et al. used machine learning to ingest simple "triplet" feedback from non-experts (e.g., "Does Odor A smell more like B or C?"). This relative feedback successfully warps the chemical space into a continuous perceptual map without requiring expert nomenclature.
Ultimately, these approaches converge toward universal representations like the Principal Odor Map (POM) developed by Wesley Wei Qian et al.. By training deep networks on human psychophysical data, the authors structured a map that generalizes across species and scales. Crucially, they discovered that the map is organized by metabolic activity and the natural co-occurrence of molecules in the environment. The brain, it turns out, does not map chemistry for the sake of chemistry; it maps the functional, metabolic meaning of the world.
Key Foundational Work
Olfactory Maps in the Brain (2011) — contributed 3 concept elements to this direction.
Niépce–Bell or Turing: how to test odour reproduction (2016) — contributed 2 concept elements to this direction.
h(odor): Interactive Discovery of Hypotheses on the Structure-Odor Relationship in Neuroscience (2016) — contributed 1 concept elements to this direction.