Goal‑directed molecular design has progressed rapidly, yet many generated molecules remain hard to synthesize, limiting practical impact. Existing synthesizability‑aware approaches either project molecules back to synthesizable analogs, deviating from the intended target, or optimize directly in discrete synthesis spaces that lack a continuous landscape for efficient search. We argue that the bottleneck lies in the search space rather than the optimizer.
To address this, we introduce RouteFlow, a framework that reformulates synthesizable molecular design as a search over a continuous route latent space. Each latent vector in this space maps to a complete synthesis route, inherently preserving synthesizability.
We navigate this space using reward‑guided flow matching as an efficient sampler, steering the search toward regions with high desired properties.
Because reward optimization can push latents off the manifold of real synthesis routes, making decoding unreliable, we add a cycle‑consistency mechanism that stabilizes fine‑tuning by keeping latents on‑manifold.
We evaluate RouteFlow on 16 optimization tasks from the Therapeutic Data Commons. The method achieves the best sample efficiency, synthetic accessibility, and retrosynthesis success rate among synthesizability‑aware baselines.
Our results also confirm that cycle‑consistency reliably maintains on‑manifold optimization while improving target properties, supporting effective synthesizable molecular discovery.
Review: RouteFlow’s continuous route latent space, combined with reward‑guided flow matching and cycle‑consistency, successfully reconciles synthesizability with efficient property optimization, offering a more practical solution for drug discovery pipelines.