SynCoGen: Synthesizable 3D Molecule Generation via Joint Reaction and Coordinate Modeling

Andrei Rekesh, Miruna Cretu, Dmytro Shevchuk, Vignesh Ram Somnath, Pietro Lio, Robert A. Batey, Mike Tyers, Michał Koziarski, Cheng-Hao Liu

Keywords: diffusion, flow matching, small molecules, conformers, synthesis, cogeneration

Abstract: Ensuring synthesizability remains a major challenge in generative small molecule design. While recent developments in synthesizable molecule generation have demonstrated promising results, they are largely confined to 2D molecular graph space, limiting their ability to perform geometry-based conditional generation. In this work, we present SynCoGen, a single framework with simultaneous masked graph diffusion and flow matching for synthesizable molecule generation. SynCoGen samples from the joint distribution of molecular building blocks, chemical reactions, and atomic coordinates. To train the model, we curated SynSpace, a dataset containing over 600K synthesis‑aware building block graphs and 3.3M conformers. We show that SynCoGen achieves state-of-the-art performance on unconditional small molecule graph and conformer generation, and the model delivers competitive performance in zero-shot for linker design. Overall, our multimodal formulation represents a foundation for future applications enabled by non-autoregressive molecular generation, including analogue expansion, lead optimization, and direct structure conditioning.

Submission Number: 114