Automatic creation of molecular substructures for accurate estimation of pure component properties using connectivity matrices

Estimation of pure component physiochemical properties has received much attention in the last decades as they serve as the basis for design of chemical products and processes. In this work, we propose a connectivity matrix-based framework coupled with machine learning for automatic creation of molecular features used to accurately estimate pure component properties. The concept of connectivity matrix is employed to represent a molecule structure. An extraction strategy is proposed to extract a plethora of submatrices (or molecular structural fragments) with each representing the environment of an atom/bond automatically and systematically from this connectivity matrix. This extraction does not cause any loss of molecular information. The submatrices are then transferred into molecular features based on matrix eigenvalues. Frequency and Pearson correlation analysis are used to extract key features, which are further reduced using principal component analysis. Machine-learning methods such as the artificial neural network (ANN) and Gaussian process regression (GPR) are used to develop prediction models, respectively. The capability and advantages of the proposed framework in comparison to existing methods are illustrated through estimation of normal boiling point of pure compounds.