Chemometric modelling of anticancer drugs using CatBoost regression and graphical derivatives

Abstract

In this work, a chemometric methodology based on graph topology descriptors and CatBoost regression is proposed for predicting the physicochemical properties of anticancer drugs. Molecular structures were encoded as graphs, and degree-based topological descriptors were derived to capture their complexity. These descriptors were used in the construction of regression models predicting boiling point, molar refractivity, and polarizability. The first statistical analysis with linear and cubic regression demonstrated that models of order higher than unity were able to take into account the non-linear dependence of descriptors vs. molecular properties. CatBoost regression model was used for improved predictability and better interpretability. This model exhibits a coefficient of determination $\left(R^2\right)$ of 0.997 for the prediction of boiling point and superior performance across all the other two properties, with average absolute errors lower than 2%. Of importance, we identified several graph descriptors as important predictors, which confirmed their chemometric relevance. The method may contribute with useful information as a complementary method to current machine learning-based models used for prediction of drug properties in chemoinformatics or pharmaceutical drug development, it integrates chemical graph theory with intelligent reasoning and modeling for a more fault tolerant and generalized 1 solution to drug property prediction.