Retrosynthetic Planning with Dual Value Networks
Summary
Apply reinforcement learning for retrosynthesis.
My contribution
Not much: I had some early conversations with the main authors and contributed a bit of code.
Thoughts
(as of 2023-11-24)
Let me preface by saying these are my thoughts and do not represent the thoughts of the other co-authors.
This algorithm does what it claims to do (produce possible synthesis routes), but I wouldn’t actually deploy it in practice. Since single-step reaction models are imperfect I think this technique (and other similar techniques) will just learn to exploit the model. My vision for retrosynthesis is to accept the presence of imperfect models and try to model their imperfections (e.g. my “retro-fallback” paper).
Also, as far as I can tell there is no publicly available code for this paper, which I think is shameful…