In a new paper a team of researchers from Insilico Medicine present a new model called GENTRL (https://github.com/insilicomedicine/GENTRL) for molecule generation. This algorithm, given a protein target, has generated 6 viable compounds in 21 days, and after 25 more days of synthesis and testing, 4 passed the preliminary tests; the most potent one was tested on live mice, and its predicted biological and chemical properties were confirmed.

The model itself is quite interesting: they use a VAE with tensor trains as a latent space; this was pre-trained via regular ELBO maximization, and after that was optimised for novelty via REINFORCE. The code is freely available and is written in PyTorch.