Generative molecular AI

This project was started between my PhD and postdoc while I was exploring data-driven methods in chemistry. It is unlikely to be continued as I pursue other projects I deem more interesting (not to mention those from my employment).

I started working on a generative model for molecule design.

Admittedly, I underestimated the complexity of this endeavour. My eventual goal is to train an equivariant message-passing graph variational autoencoder with variable input and output, written in PyTorch, and use a basic reinforcement learning algorithm to search for a molecule similar to those in QM9 with the highest HOMO-LUMO gap (as a surrogate to more interesting and complex properties). Additional molecules generated would have their targets calculated with PySCF.

I have some functions done for the final model (e.g. equivariant layers, PySCF helpers) but at the moment the model only does basic predictions, and takes padded fixed-size inputs. I hope to revisit this project more earnestly once I have released PyTCHInt.