@rasyn_lab: Today, we are launching Synthon 350M, our single-step retrosynthesis model. Synthon is the strongest publicly reported …

Today, we are launching Synthon 350M, our single-step retrosynthesis model.

Synthon is the strongest publicly reported model for this task when every model gets a single shot at the answer. Chemists make molecules by combining simpler ones. Retrosynthesis is that in reverse: start from the molecule you want, and work out what to combine to get it. A synthesis plan is that question asked over and over, one step at a time, until everything left is something you can order, and getting one step wrong wastes the rest of the plan. Synthon answers it with a ranked list, and the right answer is its first suggestion two thirds of the time and inside its top five nine times out of ten.

It works by reading the product as a root-aligned SMILES string, which lets it attack the same molecule from twenty different starting atoms and produce twenty independent sets of candidate reactants. Those sets are pooled with log-sum-exp across roots, so a reactant set that recurs across many views outranks one that a single view was confident about.

We benchmarked Synthon on USPTO-50K. It is fifty thousand real reactions curated from United States patent filings, with 5,007 of them held back for testing, and it is the standard benchmark every retrosynthesis model reports against, so the numbers below are directly comparable to published work.

With the reaction class withheld, Synthon reaches 66.23 percent top-1 and 91.41 percent top-5. Given the class, top-1 rises to 74.12 percent. In the single-shot setting, with no test-time augmentation for anyone, it reaches 62.67 percent.

Synthon is available on Marigold now for you to try. The full comparison against every published model on this benchmark is in the thread.

Our goal is to automate chemistry, the field sitting underneath medicine, agriculture, batteries, materials and most of manufacturing. All of it still moves at the speed of how fast people can plan, run and interpret experiments, and planning comes first.

Planning, for a chemist, means working out how to build the molecule. Nothing downstream starts until that is solved. You can predict that a compound binds its target, dissolves properly and is not toxic, and none of it counts until someone makes it in a lab. Every drug, every pesticide, every polymer, every battery electrolyte needed someone to find a route to it first, and for a hard molecule that search takes weeks and often ends in a dead end.

So that is the step we went after first. Synthon is the first model we have shipped toward that goal, and more are coming behind it.

You send a molecule and get back a ranked list of reactant sets that would make it.

There are two modes. One pass over the molecule takes about 0.4 seconds and gets top-1 right 62.67 percent of the time. Twenty passes pooled together takes four to six seconds and gets 66.23. That is 3.5 points for ten times the compute, which is worth it when you are checking one molecule and probably not when you are screening a library.

If you already know what kind of reaction you want, pass the class in with the molecule. Top-1 goes to 74.12 percent.

Synthon is on Marigold now.

http://app.rasyn.ai