New AI-based Framework Could Explain Why Evolution Gave So Many Species the Same Smell Circuit
Researchers have faithfully recreated key features of the initial stages of olfaction using mathematical theory and computer models
Venkatesh N. Murthy (left), associate faculty member at the Kempner Institute and Raymond Leo Erikson Life Sciences Professor of Molecular and Cellular Biology, is one of the senior authors of the new study.
At a Glance
- Researchers used mathematical theory and AI tools to suggest an explanation for why many animal species organize the initial stages of their smell circuits in the same way.
- The study indicates that this shared circuit design might help brains capture as much information as possible about odors.
- The findings could help scientists predict the olfactory circuits of less well-studied animals — and why some of them might deviate from the common design.
Why do the smell circuits of flies, mice, and humans look so remarkably alike?
A new study from Harvard researchers offers a possible explanation: this shared design may be the mathematically optimal way to process information about the olfactory world. Published in the Proceedings of the National Academy of Sciences (PNAS), the study presents a new AI-based framework designed to optimize how much information a model circuit can extract from odor data. The researchers showed that when a model is trained using this framework, it recreates the same basic blueprint that characterizes the initial stages of olfactory circuits in the brains of real animals.
In evolutionary history, insects and mammals diverged hundreds of millions of years ago, yet they converged on a strikingly similar architecture for their “early” olfactory circuits — the initial processing stages of their larger olfactory systems. The new study demonstrates computationally that this convergence can arise from a drive toward efficiency — by maximizing the information that the early olfactory system can encode about smells.
“The early olfactory system is organized in a particular way in almost every animal we’ve looked at,” said Venkatesh N. Murthy, associate faculty member at the Kempner Institute for the Study of Natural and Artificial Intelligence, Raymond Leo Erikson Life Sciences Professor of Molecular and Cellular Biology, and a senior author of the paper. “Evolution found this particular organization to be a good solution.”
This solution, often called “canonical olfaction,” involves a set of distinctive features within the early olfactory circuit. The circuit itself processes smells in three stages: first using odor receptors, which are molecular detectors that respond to one or more types of odor molecules; second, using olfactory neurons, which are odor-sensitive neurons with embedded odor receptors; and third using glomeruli, which are hubs in the brain that receive signals from olfactory neurons
In species that display canonical olfaction, each of these processing stages shows a distinct feature. Specifically, each type of odor receptor responds to many different types of odor molecules, rather than being specialized for just one type; each olfactory neuron carries just one type of receptor, rather than carrying a mix of several types; and each glomerulus only receives inputs from olfactory neurons with the same receptor type, rather than receiving a mix of signals from neurons that express different receptor types.
This canonical arrangement is widespread in the natural world. But is it optimal? To answer this question, the researchers built artificial olfactory models that began with random wiring. They then trained each model to maximize the amount of information it could transmit about odors.
“If you start with a random architecture and try to maximize the amount of information that it conveys, you robustly converge on the way animals do it,” said Murthy.
Accurately processing odor information involves overcoming formidable challenges, and Murthy and his collaborators think maximizing information might be crucial to how the olfactory system does so. A natural environment can contain thousands of possible odor molecules, but at any moment only a few are present. The brain must identify these molecules — which may come from food sources, predators, or potential mates — even though odor signals are intermingled and the number of receptors is smaller than the number of possible odor molecules.
Murthy and his collaborators saw a potential approach to this challenge in a long-standing idea in neuroscience known as the “efficient coding” hypothesis, which proposes that sensory systems are shaped to make optimal use of limited resources, especially in initial stages of processing. While the hypothesis has helped explain aspects of vision and hearing, according to Murthy, “the [olfactory] system hasn’t really been formulated this way before.”
Computational complexity may be one reason the efficient coding hypothesis has rarely been used to predict details about the early olfactory system. The key quantity that must be optimized — “mutual information,” a statistical measure of how much information is captured by a system — is extremely hard to compute in complex systems like the early olfactory circuit.
To overcome that obstacle, the researchers created a new efficient coding framework that leveraged the power of modern AI and machine learning (ML) tools. The framework allowed the researchers to approximate mutual information without calculating it exactly.
“Previous studies used older computational strategies for approximating mutual information, but we drew on some relatively recent developments in ML,” said Jacob Zavatone-Veth, a Junior Fellow of the Harvard Society of Fellows and one of the study’s senior authors.
What’s more, the new framework also enabled the researchers to optimize olfactory models that were more biologically realistic than earlier models. “We were able to extend the existing [biologically-inspired] models for early olfactory processing, and then optimize them by drawing on this body of [recent ML research],” said Zavatone-Veth.
The software models developed for the study included three biologically-inspired software “layers” that are based on the first three stages of the olfactory system: receptors that detect odor chemicals, olfactory neurons that express those receptors, and glomeruli that gather signals from olfactory neurons. The researchers trained several different software models, each starting with random connections, so that they processed olfactory information in an optimal way.
Across a wide range of computer simulations that tested different environmental conditions, the results were consistent with the canonical structure seen in animals. In each case, the model’s receptors became sensitive to a broad range of molecules. Individual olfactory neurons ended up expressing just one receptor type. And olfactory neurons that expressed the same receptor type converged onto the same glomerulus.
To further test the accuracy of their framework, the team then compared one of their models to detailed measurements from fruit fly larvae, whose 21 types of odor receptors have been carefully mapped. The model reproduced key features of real receptor types, including their wide range of sensitivities.
Because the new efficient coding framework is based on general mathematical principles rather than species-specific assumptions, it can also generate predictions. For species whose olfactory systems have not yet been fully mapped, the framework predicts the same three canonical features seen in species that evolved in typical environments.
The study’s framework may also help explain exceptions to the canonical blueprint. In some species, including the Aedes aegypti mosquito, individual olfactory neurons express more than one receptor type, violating the second feature of the canonical circuit. The framework suggests this arrangement may be useful in environments where odor molecules frequently appear together, or when receptors have similar abilities to detect odor molecules. These overlaps create strong correlations among receptor encodings, changing what counts as efficient coding. This finding complements recent work by Boston University researchers Caitlin Lienkaemper, Meg A. Younger and Gabriel Koch Ocker, also recently published in PNAS.
While the study indicates that efficient coding could account for why canonical olfaction arises in multiple species, it does not suggest that evolution creates an explicit biological mechanism that optimizes mutual information. Rather, it proposes that over millions of years, natural selection has shaped sensory systems to behave as if they maximize mutual information.
In addition to deepening our understanding of the neurobiology of smell, the researchers’ efficient coding framework has the potential to guide engineering. For example, designers of “electronic noses” — devices used in applications from food safety to medical diagnostics — face the same problem as evolution: creating the ability to detect many possible chemicals with limited sensors. In the future, such designers might be able to use this understanding of efficient olfactory coding to create better real-world detectors of odor molecules and other airborne chemicals.
“If I want to design a general-purpose chemical sensor — not a specific carbon monoxide or smoke detector but something to detect any chemical — then this principle could help,” Murthy said.
*****
To learn more about this research, read the paper in PNAS.
