Overview / Usage

Predicting odor perception from odor structure is a major goal in olfaction research. Recent developments of software packages allowing the calculation of large number of structural properties and the availability of multi-variety analysis and machine learning tools improved these predictions. In this work, we used a large dataset of odor quality ratings published for an open public challenge recently, in order to establish a deeper link between molecules structure and their perception. We suggest newly develop algorithms to achieve this goal.

Methodology / Approach

We used an adaptation of dimensionality reduction techniques and adaptation of the RotationForest method (Rudriguez et al. 2006). In short, we randomly chose subspaces of athe original space and performed our predictions by k-nn regression on this random space. We repeated this process many times. for different subsets. Instead of selecting the descriptors that contribute to this prediction the most, we averaged all the predictions made on different spaces. There was no learning phase for this algorithm, making it potentially less than optimal solution, however as it makes less assumptions (e.g. that the data distribution in the training set reflects the testing set), it is more robust and generalizable. Our results using this algorithm suggest that molecular structure can predict perceptual qualities.

Technologies Used

Code was written in Python and Matlab. The algorithms used the machine learning and statistics toolboxes, and newly developed algorithms.