Scientists train AI to predict smells based on chemicals

Scientists have made significant progress in understanding the relationship between chemical structure and smell perception by training an AI neural network to predict smells based on chemical composition. The research, published in the journal Science, aimed to shed light on the mysterious field of smell perception, which remains less understood compared to other senses like sight and hearing.

While humans can distinguish over a trillion odours, the challenge lies in deciphering the patterns and logic that govern how chemical structure determines smell. To address this, the research team trained the neural network using 5,000 compounds from two perfumery databases, along with corresponding smell labels such as ‘fruity’ or ‘cheesy’. The AI was then able to create a ‘principal odour map’ that visually represented the relationships between different smells. Impressively, when introduced to a new molecule, the neural network accurately predicted its smell through descriptive analysis.

To validate the AI’s predictions, a panel of 15 adults from diverse backgrounds in Philadelphia was asked to smell and describe the same odour. Surprisingly, the neural network’s descriptions were found to be better than the average panellist’s descriptions most of the time. This demonstrates the AI’s capability to capture certain aspects of smell perception related to chemical structure.

However, it is important to note that smell perception is deeply personal and influenced by memories, culture, and individual experiences. This complexity makes it challenging to determine a universally ‘best’ description of a smell. Nonetheless, the neural network’s odour map successfully captures the common aspects of smell perception driven by chemistry.

Although computer models have been used before to explore the relationship between chemical structure and smell perception, this study is significant because the AI neural network created an odour map and accurately predicted the smells of new molecules. However, there are limitations to this model. It solely considers chemical structure and smell, overlooking other factors such as interactions between chemicals and olfactory receptors, as well as the impact of varying concentrations of odorants. Moving forward, models need to incorporate biology and changing odour perception.

The researchers envision a future where smell can be fully digitised, similar to sound and vision. Machines equipped with the ability to detect and describe smells could revolutionise areas such as diagnostics, food quality assessment, and entertainment. However, the current model predicts smells on a molecule-by-molecule basis, whereas in reality, most smells are a combination of various odorants. Therefore, the researchers’ next step is to explore how neural networks can predict the smells of blends of chemicals.

While this study represents a significant advancement in understanding smell perception, further research is necessary to unravel the complexities of this sense and how the brain interprets smell. The ultimate goal is to digitise smell, enabling machines to detect and describe smells accurately. However, the researchers acknowledge that this study is just the first step towards achieving this ambitious objective.