10:00 - 12:00
Sat-S11
Goethe Hall
Chair/s:
Antonella Di Pizio, Sébastien Fiorucci
Computational approaches are widely used to get insights into the chemistry and biology of chemosensation. The ECRO Special Interest Group Computational Chemosensation aims to gather researchers working in computational chemosensation, to facilitate their interaction and advance computational techniques for chemical senses, but also promote the potential of computational works to promote collaborations with experimentalists. The proposed symposium is the first initiative of the group and aims to highlight computationally guided advancements in chemosensation, ranging from machine learning based predictors, to the use of computer-aided drug design tools for ligand discovery, to the multiscale simulations of chemosensory receptors, to network analyses of proteins and signaling events. Works on both taste and smell will be presented in the symposium. We expect that bringing together computational researchers from different fields will provide stimulating and fruitful discussions about future perspectives. Moreover, during the symposium, the ECRO special interest group will be introduced to the audience.
Back
Neural networks learn a map of odor that achieves human-level performance at describing odor character
Sat-S11-001
Presented by: Joel Mainland
Emily Mayhew 1, 2, Kelsie Little 1, Matthew Andres 1, Britney Nguyen 1, Jane Parker 3, Richard Gerkin 4, Joel Mainland 1, 5
1 Monell Chemical Senses Center, 2 Michigan State University, 3 University of Reading, 4 Arizona State University, 5 University of Pennsylvania
Predicting olfactory perception from molecular structure is an enduring challenge in olfaction, largely because similarity in chemical structure is an unreliable predictor of perceptual similarity. Enantiomers are structurally similar, but can evoke distinct odors, while musks from very different structural classes can be perceptually similar. Neural networks show promise for resolving this issue, as they can remap traditional structural metrics into an odor space that better predicts perceptual differences. Here, we used a neural network (NN) to generate a novel map from molecular structure that preserves perceptual relationships and enables odor quality prediction of novel odorants. We found that the NN is as reliable as a human in describing odor quality: on a prospective validation set of 400 novel odorants, the NN-generated odor profile more closely matched the trained panel mean (n=15) than the median panelist. The NN embeddings outperformed both physicochemical descriptors and chemical fingerprints on other odor prediction tasks, suggesting the NN learned a generalized representation of the structure-odor relationship. This model allows us to predict the odor of any molecule and paves the way toward digitizing odors.
Back