Search
Presentation Details
| Poster #108 Predicting Olfactory Receptor Activity Using Nanomechanical Sensors and Machine Learning |
| Noriaki Ota1, Yusuke Ihara1, Kosuke Minami2, Ryo Tamura3, Genki Yoshikawa2,4, Chiori Ijichi1 1Institute of Food Sciences and Technologies, Food Products Division, Ajinomoto Co., Inc., Kawasaki, --, Japan 2Research Center for Macromolecules and Biomaterials, National Institute for Materials Science (NIMS), Tsukuba, --, Japan 3Center for Basic Research on Materials, National Institute for Materials Science (NIMS), Tsukuba, --, Japan 4Graduate School of Pure and Applied Science, University of Tsukuba, Tsukuba, --, Japan |
Humans perceive billions of odors through ~400 olfactory receptors (ORs). Since odor information is thought to be condensed into the activity of 400 ORs, obtaining the ORs activity profile for each scent is important for understanding, digitizing, and utilizing the sense of smell. We have obtained the OR activity profiles by a cell-based assay for ~3000 odorants and foods. However, there are several drawbacks to collecting data using cellular assays: measurements may not be possible due to cytotoxicity, on-site measurements are difficult to perform, and the process is time-consuming. Therefore, we worked on developing an algorithm to predict ORs activity without cell-based assays by combining nanomechanical sensors and machine learning methods. For the nanomechanical sensor, we used “Membrane-type Surface stress Sensor (MSS)", which is characterized by high sensitivity, high selectivity, and fast response. In this study, we used an MSS array with 12 different receptor layers to acquire 12 different signals in one measurement. We acquired signal data from about 300 odorants, for which ORs activity profiles had already been obtained. We then extracted features from the acquired signal data, created regression models to predict activity values for 10 representative ORs, and compared the prediction accuracy for each. The results showed that prediction accuracy could be improved by optimizing the selection of receptor layer signals data used as features. The model with the highest prediction accuracy was developed using only three receptor layer signals data and predicted the activity of OR5K1 with a correlation coefficient (r) of about 0.733 in cross-validation. In this poster, we will report these results and discuss the relationship between receptor layers and OR activity prediction. |