A team of chemistry, life science, and AI researchers are using graph neural networks to identify molecules and predict smells. Models made by researchers outperform current state-of-the-art approaches and the top-performing model from the DREAM Olfaction Prediction Challenge, a competition for mapping the chemical properties of odors.
The work was created by researchers from Google, Canadian Institute for Advanced Research, Vector Institute for Artificial Intelligence, University of Toronto, and Arizona State University.
The researchers believe progress in machine learning application of molecule identification can help deliver machine intelligence that’s able to predict smell similar to the way AI that can imitate other senses like vision and hearing has advanced in recent years. With grasping challenges, researchers are also trying to help robotic hands tackle the human sense of touch as well.
“Progress in deep learning for olfaction would aid in the discovery of new synthetic odorants, thereby reducing the ecological impact of harvesting natural products. Additionally, new representations of molecules derived from a model trained on odor recognition tasks may contribute our understanding of sensory perception in the brain,” a paper on the topic reads.
IBM Research and fragrance company Symrise have also worked to create new smells with machine learning.
Researchers say graphs with neural networks is a fitting approach for what’s called quantitative structure-odor relationship (QSOR) modeling because it’s a good way to predict relationships between molecule properties such as smells and cluster similar molecules in a vector space. In this regard, smell can be treated like a multi-label classification problem and what researchers refer to as “odor embedding,” similar to how RGB is treated as an embedding for vision.
“By viewing atoms as nodes, and bonds as edges, we can interpret a molecule as a graph,” researchers detail in a paper titled “Machine Learning for Scent: Learning Generalizable Perceptual Representations of Small Molecules.”
“We propose the use of graph neural networks for QSOR, and show they significantly out-perform prior methods on a novel data set labeled by olfactory experts … analysis shows that the learn embeddings from graph neural networks capture a meaningful odor space representation of the underlying relationship between structure and odor, as demonstrate by strong performance on two challenging transfer learning tasks,” the paper states.
The researchers trained their model using a curated data set of 5,030 molecules from perfume materials databases. Each molecule is labeled with one or more descriptors like fruity or bready by olfactory experts, primarily practicing perfumers. Descriptors like fruity or bready were clustered together.
To support progress in the field of AI for smell prediction, Google plans to open-source more data sets related to odor prediction in the future. Future research in the space could work to digitize scent or augment the lives of people who cannot smell.