ML is useful to design the reaction pathways to make these chemicals. If we find safer/better/cheaper ways to make medicines like Remdesivir with deep learning, and publish them, then chemists may have an easier time of it.<p>If Remdesivir data looks good this month, there will be a rush to produce it, and if there’s only one published way to do that, then the ingredients for that one approach will potentially be hard to find. Thus we can benefit from different approaches which start from different raw materials.<p>Lots of cool Arxiv papers on this and Graph Neural Nets, Soft actor-critic, or Transformers can be interesting approaches. The transport theory seems like a good way to make a value function. How much time and money does it take to produce a given chemical by a given set of reactions? That’s a gajillion dollar question.<p>I spent way too much time last year looking at permutation-invariant distance metrics similar to Fused Gromov Wasserstein to invent an Atom Mover Distance, please let me know if you figure that out! DeepChem library is a solid framework, as are Tensorflow and Pytorch...<p>If anyone’s looking for a way to contribute to the COVID-19 response, open source data/algorithms to design synthesis pathways can be a strong approach. Everyone loves to use Deep Learning to design drugs, but it is valuable to design ways to make drugs, too!