Reproduce “Evaluating Molecule Synthesizability via Retrosynthetic Planning and Reaction Prediction” on OpenResearch

About this paper

Songtao Liu, Dandan Zhang, Zhengkai Tu, Hanjun Dai, Peng Liu

Researchers from The Pennsylvania State University, MIT, and Google DeepMind developed a "round-trip score" to evaluate molecule synthesizability by integrating retrosynthetic planning and forward reaction prediction. This new metric provides a more reliable assessment of synthetic route feasibility, achieving an accuracy of 78.2% and an F1 score of 84.2% in distinguishing feasible from infeasible routes, and revealed that even leading generative models produce a high proportion of challenging-to-synthesize molecules.

Key insights

• The round-trip score provides a more reliable and practical assessment of synthetic route feasibility for novel molecules, significantly outperforming the lenient 'search success rate' of retrosynthesis planners.
• There is a substantial 'synthesis gap' in state-of-the-art generative drug design models, with even the best models (e.g., Pocket2Mol) only achieving synthesizable routes for about 22% of generated molecules.
• Optimal molecular properties (e.g., binding affinity) do not always correlate with better synthesizability, highlighting a critical trade-off that generative models need to address to accelerate drug discovery.

Summary by alphaXiv.