Responding to his issues with Vijay Pande's work (I'm not affiliated), graph convolutions really are better than char-RNN for this. There's a good theoretical motivation for why, and people have spent a lot of time trying to find better embeddings of chemical space. (Admittedly, still not great - I wouldn't dispute the overall thesis that AI in drug discovery is still very early.)

I did a project a year or so ago to reimplement one of Aspuru-Guzik's papers, a variational autoencoder, (https://github.com/maxhodak/keras-molecules) and when I did that I compared to char-RNN and the VAE did get much more interesting results. I also saw results from other people around that time showing that using graph convolutions on the front end instead of one-hot encoded SMILES strings was even better.

Also, as for OP's objection about GANs not working because it's a "perfect discriminator," this is an obvious result that becomes apparent after spending 20 seconds with the problem. (Eg, this thread here: https://github.com/maxhodak/keras-molecules/issues/55) I haven't read the Harvard paper referenced but I'd be absolutely shocked if this was lost on them. There are definitely ways to work through it.