Working backwards from a molecule.
What retrosynthesis is, why it's one of the hard problems in chemistry, what I'm building for Season 1, and an honest ledger of what has and hasn't been proven yet.
01 · THE PROBLEMMaking a molecule is a route-finding problem.
Every medicine, pesticide, dye and plastic starts as a target structure on paper. Getting from cheap, available chemicals to that exact target takes a sequence of reactions, sometimes three steps, sometimes thirty. Each step has to work under real conditions, give usable yield, and not wreck the parts of the molecule you already built.
Retrosynthesis is the method chemists use to find that sequence. You start at the target and work backwards: which bond could I have formed last? What would the molecule look like just before that? Repeat until you reach things you can buy.
Aspirin takes one step. A modern drug can have dozens of bonds that could be disconnected at each stage, and every choice branches. The number of possible routes grows explosively, and most of them fail in the lab for reasons that are hard to see on paper: the wrong side of a ring reacts, a protecting group falls off, a step that works in theory gives two percent yield.
02 · HOW IT'S DONE TODAYExpert intuition, and databases of the past.
For most of the field's history, retrosynthesis lived in the heads of experienced chemists. Over the last decade, software has joined in. The leading tools mostly learn from large databases of reactions that have already been published and reported to work.
That approach is powerful, and it has a ceiling. A tool that learns from the past is strongest where the past is dense. It gets weaker on genuinely new structures, on reaction types that are under-reported, and wherever the published record is biased toward what worked once rather than why.
The Season 1 bet: derive the route from the structure itself, instead of looking it up.
03 · WHAT I'M BUILDINGA compiler for atoms.
The engine takes a target molecule and works out a route from first principles: which bonds can be formed, under what conditions, in what order, and roughly how well each step should go. It uses no reaction database and no model trained on published reactions.
Think of a compiler. You hand it a description of what you want, and it works out the low-level steps that produce it. Here the description is a molecule, and the steps are reactions.
I'm not publishing how the engine works. What I will publish, as the season goes on, is what it gets right, what it gets wrong, and how it compares to the tools chemists already use.
04 · THE LEDGERWhat's been checked so far.
These are internal checks against known chemistry, run in May 2026. Each one asks the engine for an answer chemists already know and confirms it got there without being told.
Every check above is internal: known answers, chosen by me. The engine has not been tested on molecules it has never seen, scored by someone else, or compared head-to-head with established tools. Until it is, the honest status is unbenchmarked.
05 · WHAT SEASON 1 WILL SHOWThe first real test, on camera.
- An out-of-distribution benchmark: molecules held out by structure, so the test can't be passed by familiarity.
- A side-by-side with the tools chemists use today, on the same targets.
- The misses, published as clearly as the hits. That's the deal on this channel.
- If it holds up, a way for chemists to try it on their own targets.
If you're a chemist, a lab or a company that makes molecules and you'd like to be part of that test, I'd like to hear from you.