Teaching AI to smell: A new model mimics the brain's olfactory system

Summary

Predicting how a certain chemical will smell is a notoriously tricky puzzle. Two molecules can look almost identical in their chemical structure but trigger completely different reactions in our noses and brains. Scientists have tried using massive artificial intelligence systems, known as foundation models, to predict these odor matches. However, out of the box, these general AI tools do not perform much better than older, hand-calculated methods. Both approaches end up looking at the same basic physical and chemical traits, leaving a lot of room for improvement.

To bridge this gap, researchers developed a new system called LORAX. Instead of relying on a general understanding of chemistry, LORAX fine-tunes these existing AI models to focus specifically on the biology of smell. Using specialized training techniques, the model adjusts its focus to better understand how molecules actually bind to scent receptors in the body.

This tailored approach worked. The researchers found that LORAX significantly outperforms previous models. It succeeds by organizing chemical information in a way that closely mirrors how our actual olfactory neurons, or smell-sensing brain cells, process odors. However, it is important to note that this research is currently based on early computer simulations of receptor binding, and the specific accuracy numbers or sample sizes are not yet detailed.

The story in numbers.

Novel fine-tuned model introduced (LORAX)

1

Quantitative performance metrics (not provided in abstract)

N/A