Molecular Representations in Deep-Learning Models for Chemical Property Prediction

A molecular property prediction model is dependent on the interplay between the quality of data, and expressive representation (or descriptor), and a suitable algorithm to relate the descriptors to the target property. In this work, a deep neural network (DNN) is used to regress two types of descriptors: fixed descriptors (Group fragments and Morgan fingerprints) and learned descriptors (from a Graph Neural Network, GNN). Bayesian optimization was used for hyperparameter tuning and a set of 5 models were benchmarked and used to predict the enthalpy of formation of organic compounds. GNN based models provided the best overall results compared to descriptor-based models which the attentive fingerprint model that combines RNN and graph attention mechanism (AFP) achieved the best results of 5.9 kJ/mol mean absolute deviation and a coefficient of determination of 0.99 in the training, validation, and test set. Despite not achieving chemical accuracy of 4 kJ/mol, the model has shown great promise in distinguishing between isomers and provides a baseline for future improvements to achieve chemical accuracy.