A Gaussian process model for chemoinformatics with application to the hazard classification of organic solvents

With the proliferation of screening tools for chemical testing, it is now possible to create vast databases of chemicals easily. However, rigorous statistical methodologies employed to analyse these databases are in their infancy, and further development to facilitate chemical discovery is imperative. In this paper, we address the challenge of predicting an organic solvent's water pollution class based on its chemical structure. We hypothesise that solvents with similar chemical structure are more likely to belong in the same hazard class. This chemical similarity is measured by the Tanimoto distance, a non-Euclidean metric on the chemical space. To incorporate the similarity between chemical structures in the model, we propose a Gaussian process model on the chemical space, with the kernel being a function of the Tanimoto distance. A novel feature of the proposed model is the inclusion of a scaling parameter in the kernel, which controls the strength of the correlation between compounds and offers additional flexibility. We find that accounting for correlation between chemical compounds substantially improves predictive performance over the uncorrelated model, where compound similarity is unaccounted for. Our model also compares favourably against other established models in the literature. Furthermore, we present a genetic algorithm designed to identify important features to a compound's efficacy, with the goal of facilitating chemical discovery. The algorithm operates based on two criteria derived from the proposed model. Simulation studies are conducted to demonstrate the suitability of the proposed methods.