Why do animals as different as insects and mammals process smells in nearly the same way? A new theoretical study published in PNAS takes on this question by combining ideas from neuroscience, physics, and machine learning to explain why evolution has repeatedly converged on a shared design for olfaction.
The work, led by first author and Harvard biophysics graduate student Juan Carlos Fernández del Castillo, along with postdoctoral research associate Farhad Pashakhanloo of the MCB lab of Venkatesh Murthy and Jacob Zavatone-Veth, a Junior Fellow of the Harvard Society of Fellows, examines “canonical olfaction”—a set of organizing principles that define the earliest stages of smell across many species. In this widely observed architecture, receptors are broadly responsive to many odor molecules, each sensory neuron expresses only one receptor type, and neurons carrying the same receptor send their signals to shared structures in the brain known as glomeruli.
This three-part design recurs repeatedly across evolution, from flies to mice to humans. Its repeated emergence suggests that it is not arbitrary, but reflects a deeper principle about how biological systems process information.
“We thought this is interesting,” said Zavatone-Veth. “Can we model when you would expect to see canonical olfaction and when you would expect to see non-canonical olfaction?”
That initial question was motivated by recent experimental discoveries of exceptions—cases where neurons appear to express more than one receptor type. But as the project evolved, the team found themselves returning to a more fundamental problem.
Why is canonical olfaction so widespread?
A system shaped by information
To address this, the researchers turned to efficient coding, a long-standing idea in theoretical neuroscience that proposes sensory systems are organized to maximize the information they capture about the environment while operating under biological constraints. The olfactory system faces a particularly challenging version of this problem: it must represent a vast, high-dimensional space of possible odor molecules using a limited number of receptors and neurons.
Rather than proposing a new theory, the team extended this framework using modern computational methods, allowing them to model multiple stages of olfactory processing in a more realistic way.
“We used ideas from efficient coding… but took advantage of some recent advances in machine learning to make things much more realistic and scale them up,” said Fernández del Castillo.
Their approach breaks the problem into sequential steps, optimizing how information is transmitted at each stage—from receptors to neurons to downstream circuits.
Using mutual information, a statistical measure of the co-dependence between two variables, the authors designed a model to maximize the mutual information between the odor input and the neural response. Using recent breakthroughs in machine learning, the model performs this procedure across multiple layers of the circuit under biologically plausible constraints, and their optimized model naturally recovered the defining features of canonical olfaction.
By maximizing mutual information between the odor input and the neural response, the model naturally recovered the defining features of canonical olfaction. Mutual information is a statistical measure of the dependency between two variables, essentially capturing how much of the input signal is contained in the response. By using recent breakthroughs in machine learning, the authors were able to perform this procedure across multiple layers of the circuit, under biologically plausible constraints. Across a wide range of simulated environments, these same structural motifs consistently emerged as highly effective solutions.
“Our work suggests that the canonical story is basically optimal, or close to optimal, across many different environments,” said Fernández del Castillo.
A unifying perspective
The findings help explain why similar olfactory architectures have evolved independently across distant species; in short, this organization may arise because it is a robust solution to a shared computational problem.
“The fact that it looks so similar in different animals… is very suggestive that that pattern might be advantageous,” Zavatone-Veth said.
The study also synthesizes decades of work in sensory neuroscience, bringing together experimental observations and theoretical models within a single framework.
“I think the paper is more a kind of unifying synthesis,” Zavatone-Veth said. “The framework is not really new… but we can wrap a bunch of different strands of work into this one framework.”
When do exceptions matter?
The work also addresses the growing interest in non-canonical olfaction. In certain species, including some mosquitoes, neurons can express multiple receptor types—a departure from the standard one-receptor-per-neuron rule. These findings have raised the possibility that alternative circuit designs may be beneficial in specific ecological contexts.
“Perhaps when important chemicals an animal needs to detect are always present together in the environment, combining the signals early on may be fine, and even advantageous”, says Murthy.
The new analysis suggests a more nuanced interpretation. While such deviations can occur, they do not necessarily improve the system’s ability to represent odor information.
“This type of minor co-expression is tolerated from an information processing perspective,” Zavatone-Veth said, “but it doesn’t produce a huge benefit.”
Rather than identifying strict conditions under which non-canonical designs should arise, the study shows that the canonical architecture performs well across a broad range of environments. This makes it difficult to predict when exceptions will occur—but also highlights their potential significance.
“If you found some really clear, crisp example of non-canonical olfaction, that would be very suggestive that something interesting is going on,” Fernández del Castillo said.
By framing canonical olfaction as a near-optimal baseline, the study provides a powerful tool for interpreting new discoveries. It also generates predictions about how olfactory circuits may be organized in species that have not yet been fully characterized, linking circuit structure to properties of receptor families and the statistical structure of odor environments.
More broadly, the work highlights how combining theoretical neuroscience with advances in machine learning can shed new light on long-standing biological questions. In doing so, it suggests that evolution has repeatedly arrived at the same solution for smell not by chance, but because it is one of the most effective ways to encode information about the chemical world. Murthy notes, “stepping back even further, it is interesting to consider whether similar principles apply to other chemical sensing systems, including G-protein coupled receptor signaling and the immune receptor pathways.”
(PDF)
