A group of experts in AI, life sciences, and chemistry are using graph neural networks to identify molecules and predict odors. The performance of the models built by these experts has surpassed all current methods. Especially, they are uncomparable with those born in the DREAM Olfactory Prediction Challenge.
These researchers are mainly from Google, the Canadian Institute of Advanced Studies, the Toronto Vector Artificial Intelligence Institute, the University of Toronto, and Arizona State University. They believe that with the increased level of application of machine learning in the field of molecular recognition, machine intelligence will be able to perform odor recognition. It is similar to the case when AI simulates other perceptions such as sight and hearing. In addition, researchers are still trying to get the robotic arm to feel tactile.
A related paper writes: ‘The advancement of deep learning in olfaction can help discover new chemical compounds, thereby reducing the demand for natural crops and reducing the impact on the ecological environment. The molecular structure is derived from the odor recognition model. It can help us understand how the brain’s olfactory perception works.’
How AI Will Have A Sense of Smell?
IBM Research and perfume company Symrise are also trying to design new flavors through machine learning. Researchers have shown that the graph neural network is well suited to the structure-odor quantitative relationship model (QSOR). It is capable of premolecular properties (such as odor) and the relationship of cluster-like molecules in vector space. From this point of view, odor recognition can be regarded as a multi-label classification problem, which researchers call ‘olfactory embedding’. It is similar to the computer decomposition of images into red, blue, and green.
One of the researchers explained: ‘By considering the atom as a node and the chemical bond as an edge, we can think of the molecule as an image. We propose to apply the graph neural network to the QSOR model and prove it with the database provided by the olfactory expert. Its performance far exceeds the existing methods. The analysis shows that the analytical embedding of graph neural networks can uncover the potential relationship between molecular structure and odor.’
The researchers used the molecular data of 5030 perfume materials in the database to train their models. Every molecular data puts the olfactory expert labeled, including fruity, toasted, and so on, and disrupts it.
In order to speed up the prediction of olfactory prediction AI, Google plans to disclose more relevant data sets in the future. Research in this area will be able to digitize the scent and help people find more smells that they can’t smell.