Retrosynthesis works backwards from a target molecule: at each step a chemist (or, here, a model) picks a bond to disconnect, splitting the target into two simpler synthons that could react forward to remake it. Repeating this down to cheap, purchasable starting materials gives a synthesis plan. Because several disconnections are usually chemically valid at once, planning software must search a branching tree of possibilities rather than follow one fixed path.
Real AI retrosynthesis tools (e.g. AiZynthFinder, Chematica/Synthia) pair exactly this kind of tree search with a neural network trained on millions of literature reactions to score each disconnection — the tree search decides where to look, and the learned model decides what looks promising, exactly the two-part design MCTS + neural policy that made AlphaGo work for the game of Go.
Watch an AI retrosynthesis planner unravel a target molecule into simpler precursors. At every node the model proposes 2-3 candidate disconnections with a predicted success probability, and a Monte Carlo tree search — the same idea AlphaGo used — walks the tree, backs up results, and locks onto the cheapest, highest-confidence route down to purchasable building blocks.
Real retrosynthetic disconnections — ester → acid + alcohol, Grignard product → alkyl/aryl halide + carbonyl, amide → acid + amine, Suzuki biaryl → boronic acid + aryl halide, and more — branch a target molecule into a genuine search tree. Monte Carlo tree search explores that tree the way AlphaGo explored Go moves: select the most promising under-explored branch, expand it, estimate its value, and back the result up through every branch on the path, so visit counts gradually converge on the strongest route.
Set Complexity (2-5 steps) to size the target's retrosynthetic tree, then press Run retrosynthesis search to animate the Monte Carlo search live — watch disconnection candidates light up as they're visited and the tree grow downward toward purchasable reagents. Toggle Show alternative paths to keep rejected branches dimly visible instead of hiding them once the best route locks in, and use Reset to draw a brand-new random target at the current complexity.
Production AI retrosynthesis tools such as AiZynthFinder pair this exact tree-search skeleton with a neural network trained on millions of literature reactions to score each disconnection — the search decides where to look, the learned model decides what looks promising, the same two-part recipe that made AlphaGo superhuman at Go.