In the search for new medicines, the bottleneck is rarely imagining a molecule—it is figuring out how to build it. The chemical space of drug-like molecules is estimated at $10^{60}$ candidates, an astronomical figure that dwarfs the number of stars in the observable universe. Navigating this space requires retrosynthesis: the process of working backward from a target molecule to identify a sequence of chemical reactions that lead to simple, commercially available building blocks.
Because each step in a synthesis can involve dozens of potential precursors, and each precursor requires its own synthesis, the search space suffers from a "combinatorial explosion." Researchers are increasingly treating this not just as a chemistry problem, but as a discrete search and optimization challenge, applying the same logic used to master games like Chess and Go.
The Asymmetry of Chemical Search
Retrosynthesis is fundamentally harder than forward reaction prediction. As Peilin Zhao and colleagues explain in A Unified View of Deep Learning for Reaction and Retrosynthesis Prediction, forward chemistry is a many-to-one mapping: given specific reactants, the outcome is usually deterministic. Retrosynthesis is a one-to-many mapping. A single target molecule might be produced by thousands of different reactant combinations, many of which involve "leaving groups"—atoms that disappear during the reaction and are thus absent from the final product.
This asymmetry creates a massive branching factor. Early pioneers like N. S. Sridharan recognized this in Search Strategies for the Task of Organic Chemical Synthesis, framing organic chemistry as a non-linear search through millions of potential subgoals. If a search algorithm explores the wrong branch, it quickly becomes lost in a "cloud" of unrealistic chemical intermediates.
Modeling Synthesis as a Game
To tame this complexity, researchers have modeled chemical synthesis as a formal game. Abraham Heifets and Igor Jurisica demonstrated in Construction of New Medicines via Game Proof Search that drug synthesis can be treated as a two-player zero-sum game. By applying Proof-Number Search (PNS)—an algorithm designed to find the most efficient path to a "proven" state—they could navigate the discrete search space of organic chemistry to recreate the synthetic route for blockbuster drugs like Lipitor.
The field saw a major shift with the arrival of Towards "AlphaChem". In this work, Marwin H. S. Segler and his team combined Monte Carlo Tree Search (MCTS) with deep neural networks. Much like AlphaGo evaluates board positions, their system uses neural "policy" networks to suggest the most promising reactions and "value" networks to estimate the likelihood that a specific precursor can eventually be synthesized from available materials.
The AND-OR Tree Structure
The most effective way to visualize this search is through an AND-OR tree.
- An OR node represents a choice: "To make Molecule A, I can use Route 1 or Route 2."
- An AND node represents a requirement: "To use Route 1, I need Molecule B and Molecule C."
The challenge is that the tree grows exponentially with every step. Recent work on AOT* improves this by using Large Language Models (LLMs) to "atomically map" entire synthetic pathways onto the search tree. This allows the system to skip the step-by-step grind of traditional search, requiring 3–5× fewer iterations to find a valid route.
Closing the Synthesizability Gap
A recurring problem in AI-driven drug discovery is the "synthesizability gap": models design "fantasy molecules" that look effective on paper but are impossible to build. Michael Sun and Connor W. Coley highlight this in Procedural Synthesis of Synthesizable Molecules, noting that the astronomical number of building blocks makes it difficult for models to stay within the bounds of known chemistry.
To solve this, new frameworks like MoleBridge use Discrete Markov Bridges to project theoretical designs into "synthesizable space." Instead of designing a molecule and then hoping it can be built, these models generate the "recipe" (the sequence of reactions) and the molecule simultaneously. Similarly, ReaSyn, developed by Seul Lee and colleagues, uses a bidirectional search—moving both top-down from the target and bottom-up from building blocks—to ensure the final design is "cookable" in a lab.
Accelerating the Search
As the field moves toward high-throughput screening, speed is becoming as important as accuracy. Traditional tree searches are slow because they must evaluate thousands of potential branches.
- Speculative Decoding: Mikhail Andronov and Jürgen Schmidhuber adapted "speculative beam search" from LLMs to chemistry in Fast and scalable retrosynthetic planning, allowing the model to solve up to 86% more molecules under strict time constraints.
- Evolutionary Algorithms: In Evolutionary Retrosynthetic Route Planning, Yan Zhang and team framed retrosynthesis as an optimization problem. By "evolving" a population of potential routes and using parallel processing, they reduced search times by 84% compared to MCTS.
By treating chemistry as a structured, searchable language, these researchers are transforming drug discovery from a process of trial-and-error into a precise computational discipline. The goal, as argued in AI Meets Chemistry, is to move beyond digital victories in games to real-world breakthroughs in medicine.