Genetic optimization of training sets for improved machine learning models of molecular properties

TL;DR

This paper demonstrates that genetic algorithms can optimize training set composition for molecular property prediction models, significantly improving their accuracy over randomly selected training sets.

Contribution

It introduces a genetic algorithm-based method for selecting training molecules, leading to more accurate machine learning models of molecular properties.

Findings

Reduced mean absolute errors by ~25% for some properties

Optimized training sets outperform random sampling

Class-based training sets enable transferability to similar molecules

Abstract

The training of molecular models of quantum mechanical properties based on statistical machine learning requires large datasets which exemplify the map from chemical structure to molecular property. Intelligent a priori selection of training examples is often difficult or impossible to achieve as prior knowledge may be sparse or unavailable. Ordinarily representative selection of training molecules from such datasets is achieved through random sampling. We use genetic algorithms for the optimization of training set composition consisting of tens of thousands of small organic molecules. The resulting machine learning models are considerably more accurate with respect to small randomly selected training sets: mean absolute errors for out-of-sample predictions are reduced to ~25% for enthalpies, free energies, and zero-point vibrational energy, to ~50% for heat-capacity, electron-spread, and polarizability, and by more than ~20% for electronic properties such as frontier orbital eigenvalues or dipole-moments. We discuss and present optimized training sets consisting of 10 molecular classes for all molecular properties studied. We show that these classes can be used to design improved training sets for the generation of machine learning models of the same properties in similar but unrelated molecular sets.