P025 The opportunities and limitations of AI as a tool in neuroscience: how does the nose know what it knows?
James M. Bower Biome Farms, Veneta Oregon Introduction
There has been a dramatic increase in the use of AI in the analysis of neurobiological data.For example, recently a graphical neural network trained to predict odor percepts from molecular structures has suggested that olfactory discrimination may be based on the metabolic relationships between molecules rather than their physio-chemical structures (Qian et al., 2023).Although these authors were unaware, Chris Chee in my laboratory had discovered the same result 25 years earlier using a different kind of analysis of olfactory perception (Ruiter-Chee, 2000).This talk will describe each approach, the neurobiological significance of the results, then considering the value and limitations of AI as an abstract data analysis tool.
Methods In the first study, a graphical neural network constructed a map from molecular structure to odor descriptors, the results tested comparing model results to those of human experts (Lee et al., 2023). Chemical relationships within the AI produced odor map were then examined (Qian et al., 2023).The second approach used a cross-entropy analysis of the co-occurrence of individual descriptors in the human identified profiles of 822 molecules.The resulting directed graph was then analyzed for the locations of odorants containing nitrogen and sulfur (Chee-Ruiter, 2000).
Results Both studies suggest that human olfactory discrimination reflects the metabolic relationships between molecules rather than their strict physio-chemical properties.Metabolically related but structurally dissimilar molecules were grouped together in the AI generated map, while molecules containing sulfur and nitrogen co-localized in the directed graph.While both studies reached similar conclusions, the cross-entropy analysis lead directly to further studies of the binding properties of olfactory receptors as well as realistic modeling studies of the organization of efferent and intrinsic pathways within olfactory cortex, both suggesting that the olfactory system intrinsically “knows” about the metabolic structure of the world.
Discussion The results of both studies suggest that the assumption, first proposed by the Roman poet and philosopher Lucretius who in 50 B.C.E, that the olfactory system recognizes and categorizes odorant molecules based on their general physio-chemical properties is fundamentally wrong. Accordingly at a minimum the physio-chemically organized (i.e. carbon length chain) panels of odor stimuli traditionally used in olfactory experiments are unlikely to reveal how the olfactory system works. Beyond that, however, the additional studies conducted in our laboratory call into question whether the neurobiological basis for olfactory discrimination, for example, is learned or intrinsic, a question that cannot be addressed by the AI model.
Acknowledgements d I acknowledge the alpacas, emus, and horses that watch me very day as I work on my books and papers. Otherwise, I am completely self funded, as I am simulating an 18th century landed gentry scientist. References Bailey, C., (1959)Lucreti De Rerum Natura Libri Sex,2nd edition (Oxford Press)
Chee- Ruiter, CWJ. 2000. The biological sense of smell: olfactory search behavior and a metabolic view for olfactory perception. Dissertation (Ph.D.), California Institute of Technology
Lee, B.K. et al. (2023) A principal odor map unifies diverse tasks in olfactory perception. Science 381: 999.
