A preprint on my review article was just posted on chemRxiv!

This is my magnum opus and it's two things at once: first, a comprehensive review of the synthesis planning literature. We argue that the field has been latently undergoing a phase transition from era of navigability to an era of validity. We describe the practices of the former and propose a framework for evolution of the field. The field has been fortunate to have contributions from the wide ML community, but there's been a salient misconception that chemical templates, as is, guarantee chemical validity, which they do only to an extent.

We formalize that extent into a hierarchy of chemical validity: Syntactic Validity (Tier-0), Topological Validity (Tier-1, what templates guarantee), Selectivity (Tier-2) and Executability (Tier-3).

The second part of the review answers the question - but why should I care about synthesis planning? We argue it's an underexplored path to a true foundation model. We argue chemistry is in its pre-GPT era as the models lack emergence of zero-shot generalization. The key to knowing chemistry is understanding functional groups. We think that multistep synthesis planning provides a unique epistemic environment for a model to learn those FGs: a full plan contains not only information about which groups react and how, but also implicitly which groups can coexist, which are conflicting with each other (and so necessitate protective groups). And crucially, the same quantum mechanical reasons that determine the reactivity of functional groups determine their properties. So we hypothesize molecular property prediction might be an emergent characteristic of a scaled synthesis planner.

As such, our argument is a falsifiable scientific conjecture that turns Artificial Chemical Intelligence from a brand term into a concrete research roadmap.

Because this is quite a behemoth of a review, while I recommend reading it in full, we also provide an interactive portal: https://ischemist.com/syntax-of-matter

You can look at key ideas, play with the calculator explaining the difficulty of the problem, browse existing approaches, explore the validity framework.

You can also take a reading path for a general quick overview or one tailored to a chemist or an ML practitioner.

On the page with full text you'll also find NotebookLM generated deep dive podcasts on the topic. I've listened through all of them and they give a good overview of the work.

Synthesis Planning has been recently identified as one of the Grand Challenges in small molecules drug discovery, and we hope our review helps more people understand the current landscape and what problems need to be solved (we also have a dedicated open problems page on the interactive portal!)