RxnFlow

A generative model can propose a molecule that docks beautifully into a target pocket and that no chemist can actually make. Most de novo designers assemble atoms or fragments into a molecular graph and treat synthesizability as a score bolted on afterwards, so a promising design dies at the moment someone has to order reagents. RxnFlow removes the question by changing what the model emits: not a graph, but a synthetic pathway — a purchasable starting block followed by up to three reactions drawn from a fixed template set. The route is the output, so synthesizability is a property of the generative process rather than a filter applied to it.

Building molecules this way is easy in principle and expensive in practice. The action space is the product of a building-block catalog and a reaction template library — roughly 10 million pairs once 1.2 million Enamine blocks meet 71 templates. Earlier reaction-based generators shrank the catalog to a few thousand blocks to stay tractable, which narrows the reachable chemistry and pushes the model toward longer syntheses. RxnFlow instead trains its generative flow network on a random subsample of the allowable actions at each step and corrects the state-flow estimate by importance sampling, so memory and compute track the subsample rather than the catalog.

The work is by Seonghwan Seo and colleagues at KAIST, with collaborators at Simon Fraser University, POSTECH, OMELET and HITS, and appeared at ICLR 2025. Two checkpoints ship under the one name: an unconditional prior trained on drug-likeness, used as a warm start for per-target optimization runs, and a pocket-conditional model that generates against a binding site it has never seen with no training iterations.

#Key Features

• Synthesis as the generative action: Each trajectory picks a starting block, then uni- or bi-molecular reaction templates, so the sampled object is a route and the molecule is what the route produces.
• Action space subsampling: The forward policy is estimated from an importance-weighted subset of allowable actions, cutting the cost of the block-template product from the full catalog to a controllable sample size.
• Non-hierarchical action selection: Template and reactant block are chosen jointly in one continuous action embedding rather than template-then-block, keeping flow estimates consistent when the block library changes.
• Zero-shot pocket conditioning: A pocket-conditional checkpoint samples binders for an arbitrary pocket without a per-target optimization loop, using a pharmacophore-based proxy for the docking score as its reward.
• Objectives and libraries added without retraining: Because actions carry chemical context, a trained policy can be pointed at newly available blocks or constrained toward a new property, such as low topological polar surface area, at sampling time.

#Technical Details

A graph transformer embeds the partially built molecule; building blocks are embedded from fingerprints and physicochemical properties into a shared continuous action space, and MLP heads score the stop and reaction actions. Training uses the trajectory balance objective with a 1% subsampling ratio. The template set holds 13 uni- and 58 bi-molecular reactions, and trajectories run to at most three steps. On the CrossDocked2020 pocket-conditional benchmark — 100 test pockets, 100 molecules each — RxnFlow reaches an average Vina score of −8.85 kcal/mol (median −9.03), QED 0.67, 34.8% synthesizability under AiZynthFinder retrosynthesis, and 0.81 diversity, in about 4 seconds per pocket. MolCRAFT in its larger configuration scores better on docking alone (−9.25) at 3.9% synthesizability, and the diffusion baseline TargetDiff reaches 0.9%. Under pocket-specific optimization across 15 LIT-PCBA targets the paper reports RxnFlow ahead of every reaction- and fragment-based baseline on hit ratio; on the five targets tabulated it also needs the fewest synthesis steps, 1.95–2.42 on average against 2.45–3.83 for those baselines.

#Applications

The practical use is hit generation against a named binding site where the output has to be orderable: the pocket-conditional checkpoint proposes candidates for a new target in seconds, and the per-target optimization scripts fine-tune against GPU-accelerated UniDock docking when a campaign justifies the compute. Because the action space is a catalog, teams can point the same policy at their own building-block inventory or a make-on-demand library, and add a mid-campaign objective such as solubility without starting over. The authors have since built a commercial screening product on the approach with eMolecules and HITS.

#Impact

The contribution that carries beyond this model is the demonstration that the action-space size which made reaction-based generation expensive can be handled by subsampling rather than by pruning the catalog, and that a synthesis-oriented generator can be run pocket-conditionally at all. The limits are worth stating plainly. Every reward in the paper is computational — docking scores and a retrosynthesis solver's verdict — with no experimental validation, and the biological grounding is the protein pocket used for conditioning rather than biological data the model was trained on. The reachable chemical space is bounded by the template set and the block library, and the Enamine catalog used in the paper is available only on request. The maintained branch no longer reproduces the published numbers exactly; a tagged archive holds the paper version.