SynCoGen

Generative models propose molecules far faster than chemists can evaluate them, but a large share of those proposals cannot actually be made in a laboratory. Two research lines have attacked the problem from opposite ends. Synthesis-aware generators chain commercial building blocks through known reaction templates, guaranteeing a route but working only on 2D graphs. Three-dimensional generators sample atomic coordinates directly, capturing the geometry that governs binding, but offer no assurance that the result is synthetically accessible. SynCoGen closes the gap by sampling from the joint distribution over building blocks, the reactions that join them, and the atomic coordinates of the assembled molecule.

SynCoGen was developed by researchers at the University of Toronto and The Hospital for Sick Children with collaborators at Cambridge, ETH Zürich, the Vector Institute, Mila, McGill, and Caltech, and was released as a preprint in July 2025 before publication at ICLR 2026. Training it required a purpose-built corpus, so the authors also released SynSpace, a dataset family of synthesis-annotated molecular graphs paired with low-energy 3D conformers.

The practical consequence is amortization. Because synthetic accessibility is baked into the generative prior rather than enforced by a reward or filter, one trained checkpoint handles fragment linking and pharmacophore-conditioned design across many protein targets without per-target retraining, docking-based reward models, or scoring-function fitting.

#Key Features

• Joint graph and coordinate sampling: Masked discrete diffusion over a building-block reaction graph runs in unified time with continuous flow matching over Cartesian coordinates, so chemistry and geometry are decided together rather than in sequence.
• A synthesis route with every molecule: Nodes are commercial building blocks and edges are specific reactions and reaction centers, so each sample arrives with an explicit multi-step route rather than a structure a chemist must retrosynthesize after the fact.
• Chemistry-aware constraints: Compatibility masking restricts node and edge logits to competent reagent–reaction pairs, while edge-count limits and diagonal suppression prevent impossible or macrocyclic topologies. Ablations identify these constraints as the single largest contributor to performance.
• Conditional design without retraining: Fragment linking is posed as inpainting over fixed substructures, and pharmacophore profiles are injected as massless pseudo-atoms, letting one model serve both tasks zero-shot.
• Flexible atom count: Because the prior is over building blocks rather than atoms, the number of atoms need not be specified at inference time, unlike most all-atom 3D generators.

#Technical Details

The backbone is an SE(3)-equivariant architecture adapted from SemlaFlow. The discrete component follows masked diffusion language modeling with the SUBS parametrization, operating on node and edge tensors; the continuous component is a visibility-aware conditional flow matching process over coordinates, trained with cross-entropy, masked coordinate MSE, pairwise-distance, bond-length, and smooth-LDDT losses. SynSpace was built by procedurally coupling 93 low-cost commercial building blocks through 19 high-yield reaction templates, giving 622,766 reaction graphs of two to four steps and 3,360,908 conformers; the extended SynSpace-L vocabulary of 378 blocks and 26 reactions adds 600,000 graphs and 4,223,367 conformers. Conformers were generated with ETKDG, minimized with MMFF94, re-optimized with GFN2-xTB, and filtered to a 10 kcal/mol window with 1.5 Å RMSD clustering. Training ran 100 epochs on two H100-80GB GPUs.

On unconditional 3D generation, SynCoGen reaches 96.7% RDKit validity against 93.3% for SemlaFlow, 91.1% for JODO, and 85.9% for EQGAT-diff, with retrosynthesis solve rates of 50% (AiZynthFinder) and 72% (Syntheseus) versus 38% and 48% for the strongest baselines; 87.2% of conformers pass PoseBusters and the Fréchet ChemNet Distance is 2.91. Diversity is modestly lower (0.78) as a consequence of the constrained building-block space. In pharmacophore-conditioned generation across ten PDB and LIT-PCBA targets, it achieves the best docking score on 8 of 10 targets while retaining 61%/78% retrosynthesis solve rates; docked poses were cross-checked against AlphaFold 3 predictions.

#Applications

The model targets medicinal chemistry workflows where a design must be both geometrically plausible and orderable: linking fragments that bind distinct subsites of a pocket, expanding analog series around a known ligand, hopping scaffolds, and building bifunctional molecules such as PROTACs. Because generated routes draw on in-stock reagents and robust couplings, the outputs suit automated synthesis platforms and self-driving labs. The codebase supports custom building-block and reaction vocabularies, so groups with proprietary reagent inventories can retrain against their own accessible chemical space.

#Impact

SynCoGen demonstrates that synthesizability can be a property of the generative prior rather than a post-hoc filter, and that this constraint costs little in sample quality while substantially improving retrosynthetic tractability. Trained checkpoints for unconditional and pharmacophore-conditioned sampling are released on Hugging Face under an MIT license alongside the SynSpace dataset, though the code repository itself carries no license file. The main limitation is validation: all reported results are computational, resting on docking scores, retrosynthesis solvers, and semi-empirical energies, and no generated molecule has yet been synthesized or assayed. Chemical scope is also bounded by the reaction vocabulary, and stereoisomers collapse to a single graph representation.