Combining variational autoencoder representations with structural descriptors improves prediction of docking scores

Molecular hypergraph grammar variational autoencoders are a generative model with great potential for de novo design in drug discovery. A numerical experiment with the aim of predicting docking scores for the D2 dopamine receptor shows that combining representations from this model with structural descriptors of the docking pose attains state-of-the-art performance. The results suggest that incorporating structural information in the training of variational autoencoders could lead to better representations and accelerate guided virtual screening.