Odorant clustering based on molecular parameter-feature extraction and imaging analysis of olfactory bulb odor maps

Progress in the molecular biology of olfaction has revealed a close relationship between the structural features of odorants and the response patterns they elicit in the olfactory bulb. Molecular feature-related response patterns, termed odor maps (OMs), may represent information related to basic odor quality. Thus, studying the relationship between OMs and the molecular features of odorants is helpful for better understanding the relationships between odorant structure and odor. Here, we explored the correlation between OMs and the molecular parameters (MPs) of odorants by taking OMs from rat olfactory bulbs and extracting feature profiles of the corresponding odorant molecules. 178 images of glomerular activities in olfactory bulb that are corresponding to odorants were taken from the OdorMapDB, a publicly accessible database. The gray value of each pixel was extracted from the images (178 × 357 pixels) to fabricate an image matrix for each odorant. Forty-six molecular feature parameters were calculated using BioChem3D software, which was used to construct a second matrix for each odorant. Correlation analysis between the two matrixes was first carried out by establishing coefficient maps. Results from hierarchical clustering showed that all parameters could be segregated into seven clusters, and each cluster showed a relatively similar response pattern in the olfactory bulb. Using the information from the OMs and MPs, we mapped odorants in 2D space by incorporating dimension-reducing techniques based on principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE). Artificial neural network models based on the OM and MP feature values were proposed as a means to identify odorant functional groups. An OM-PCA-based model calibrated via extreme learning machine (ELM) was 94.81% and 93.02% accuracy for the calibration and validation sets, respectively. Similarly, an MP-t-SNE-based model calibrated by ELM was 86.67% and 93.35% accuracy for the calibration set and the validation set, respectively. Thus, this research supports a structure-odor relationship from a data-analysis perspective.