Lab Newsletter — September 17, 2026: How to Make a Molecule

Thu, 17 Sep 2026 03:00:12 +0000

This month we’ve watched AI design a great deal — small molecules, proteins, mRNA, antibodies. But design is only half of discovery. A molecule a model dreams up is worthless until someone can actually make it — and figuring out how to make a complex organic compound is itself a hard, creative search. Chemists call it retrosynthesis: recursively breaking a target molecule down into simpler, purchasable precursors. Today’s digest is about the deep- learning systems that learned to do that planning — and then to predict what a reaction will produce, and finally to hand the whole route to a robot.

🧠 The founding move: learn the rules of reactivity

Rule-based synthesis planners had been around for decades, but they were brittle. Segler & Waller (Chemistry – A European Journal, 2017) diagnosed why: “reaction rules often fail because they ignore the molecular context, which leads to reactivity conflicts.” Their fix was to learn from the whole literature — “deep neural networks can learn to resolve reactivity conflicts and to prioritize the most suitable transformation rules.” Trained “on 3.5 million reactions taken from the collective published knowledge of the entire discipline of chemistry,” the model reached “a top10-accuracy of 95% in retrosynthesis and 97% for reaction prediction.” The move that would define the field: don’t hand-code chemistry’s rules — learn them from every reaction ever published.

🔗 Knowledge-free: rank disconnections by analogy

Do you even need explicit rules? Coley, Rogers, Green & Jensen (ACS Central Science, 2017) showed you can lean on precedent instead. They “demonstrate molecular similarity to be a surprisingly effective metric for proposing and ranking one-step retrosynthetic disconnections based on analogy to precedent reactions.” The approach “mimics the retrosynthetic strategy defined implicitly by a corpus of known reactions without the need to encode any chemical knowledge.” Tested against reality — “using 40,000 reactions from the patent literature … the recorded reactants are among the top 10 proposed precursors in 74.1% of 5000 test reactions.” A planner that works by remembering what chemists have already done, and reasoning by analogy.

🌳 The landmark: search the tree of syntheses

A single disconnection isn’t a synthesis — you need to chain them into a full route, which is a vast search. Segler, Preuss & Waller (Nature, 2018) brought the game-playing toolkit to chemistry. Noting that “computer-aided retrosynthesis would be a valuable tool but at present it is slow and provides results of unsatisfactory quality,” they combined “Monte Carlo tree search and symbolic artificial intelligence” with “an expansion policy network that guides the search, and a filter network to pre-select the most promising retrosynthetic steps.” Trained “on essentially all reactions ever published in organic chemistry,” the system “solves for almost twice as many molecules, thirty times faster than the traditional computer-aided search method.” And it passed the ultimate test — in “a double-blind AB test,” chemists rated its routes on par with literature syntheses. AlphaGo’s idea, aimed at the synthesis tree.

🔮 The forward problem: predict what a reaction makes

Planning backward is only trustworthy if you can also predict forward — given reactants, what actually forms? Schwaller et al. (ACS Central Science, 2019) framed it as language. Treating “reaction prediction as a machine translation problem between … (SMILES) strings … of reactants, reagents, and the products,” their Molecular Transformer “outperforms all algorithms in the literature, achieving a top-1 accuracy above 90% on a common benchmark data set,” “requires no handcrafted rules,” and — crucially for autonomy — “can accurately estimate its own” uncertainty. A model that knows when it doesn’t know is exactly what you want before you commit reagents.

🧰 Make it open, make it fast: a tool anyone can run

Landmark results only change practice when they ship as usable software. Genheden et al. (Journal of Cheminformatics, 2020) delivered that with AiZynthFinder: “open-source software that can be readily used in retrosynthetic planning,” whose “Monte Carlo tree search … recursively breaks down a molecule to purchasable precursors,” guided by “an artificial neural network policy.” It’s built for real use — “fast and can typically find a solution in less than 10 s” — and, notably, the authors foregrounded software engineering: “automatic testing, system design and continuous integration leading to robust software with high maintainability.” It’s the same publish-the-tool ethos behind the lab’s BioImage Model Zoo and BioEngine: a method becomes infrastructure only when others can run it.

🤖 Closing the loop: from plan to robot to molecule

The endpoint of all this is not a route on a screen but a compound in a vial. Coley et al. (Science, 2019) built exactly that bridge, “combining artificial intelligence-driven synthesis planning and a robotically controlled experimental platform.” Routes are “proposed through generalization of millions of published chemical reactions and validated in silico to maximize their likelihood of success,” then “executed by a modular continuous-flow platform that is automatically reconfigured by a robotic arm to set up the required unit operations.” They demonstrated it “for 15 drug or drug-like substances.” AI plans, a robot builds — a chemistry analogue of the self-driving lab.

🧫 Why it’s our kind of problem

Read across the six and it’s the missing half of everything the lab has been digesting this month. Generative models can now propose molecules, proteins, and antibodies at will — but a design you can’t synthesize can’t be tested, and a hypothesis you can’t test can’t teach the model anything. Retrosynthesis plus robotic execution closes the design → build → test → learn loop that underpins the lab’s self-driving-lab and autonomous-discovery ambitions — the same loop Agent-Lens and the REEF imaging farm run on the imaging side, here run on the making side. The recurring pattern is one the lab keeps betting on: learn from the entire corpus of published results, search intelligently, quantify your uncertainty, and ship it as open, automatable infrastructure (Molecular Transformer’s calibration and AiZynthFinder’s open code are both squarely that). And it points at the bigger prize — a closed loop where an AI agent designs a candidate, plans its synthesis, runs it on a robot, measures the result, and updates itself. Discovery isn’t just imagining the molecule; it’s knowing how to make it.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits; a hypha-search surrogate sweep surfaced only virtual-cell horizon items, nothing breaking. Anchors were verified via NCBI E-utilities.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.

autonomous-labs | AICell Lab