Predicting Smell Perception from Molecular Descriptors Using Machine Learning Approach

As compared to smell, other senses like touch, taste, vision, and hearing are highly predictable. Though, it is very difficult to predict the smell perception of a molecule from its molecular properties. It is reported that structurally very similar types of molecules can produce a different smell, and structurally very different types of molecules can produce the nearly same smell. The goal of the manuscript is to have a comparative performance measure between the existing machine learning algorithms like ANN, SVR, DTR, RFR, and CNN, while using the DREAM olfaction challenge dataset and a series of defined feature set. In this paper, we have used machine learning approach to predict the Molecular olfactory perception based on the molecular descriptors of structurally different molecules. 4884 molecular descriptors of 472 structurally different molecules were used in this study. In terms of machine learning, it's a multi-input and multi-output regression problem and every feature need to be combined to give output in 21 targets. In this study, Artificial Neural Network (ANN), Decision Tree Regressor (DTR), Support Vector Machine (SVR), Convolution Neural Network (CNN), and Random Forest Regressor (RFR) has been used to predict the olfactory perception of a molecule. For the validation of our model R-Squared method with residual plots has been used. Mean Squared Error (MSE) and Mean Absolute Error (MAE) has been used for Error calculation. The ANN regression model performed better than all the other models used in this study. For the ANN regression model, MSE and MAE were 44.56 and 4.19 respectively.