In this paper, we present SynTwins, a novel framework that bridges the gap between desirable properties of AI-generated molecules and actual synthesizability. SynTwins mimics the strategy of expert chemists by designing synthesizable molecular analogs through three steps: retrosynthesis, exploration of similar building blocks, and virtual synthesis. It outperforms existing machine learning models in generating synthesizable analogs and, when integrated with existing molecular optimization frameworks, generates synthesizable molecules with desirable properties. Through comprehensive benchmarking on various molecular datasets, we demonstrate that SynTwins effectively bridges the gap between computational design and experimental synthesis, making it a practical solution for accelerating the discovery of synthesizable molecules with desired properties for a wide range of applications.
