GFlowNets for molecule generation with guaranteed synthesis pathways

Recent breakthroughs in generative modelling have led to a number of works proposing molecular generation models for drug discovery. While these models perform well at capturing drug-like motifs such as rings, they are known to produce synthetically inaccessible molecules. This is because these models are trained to compose atoms or fragments in a way that approximates the training distribution, but these models are not explicitly aware of the synthesis constraints what come with making molecules in the laboratory. To help tackle this, we introduce Synthesis Flow Networks (SynFlowNet), a GFlowNet model whose action space is to build molecules from reactions and reactants. We find that SynFlowNet consistently samples synthetically feasible molecules, while still being able to find diverse and high-utility candidates. Furthermore, we found that molecules designed with SynFlowNet were 85% more synthesisable than those sampled from a GFlowNet which builds molecules using molecular fragments, according to SA score.