SynFrag: Synthetic Accessibility Predictor based on Fragment Assembly Generation in Drug Discovery

AI-driven molecular generation encounters a "generation-synthesis gap": most computationally designed molecules cannot be synthesized in laboratories, limiting AI-assisted drug design (AIDD) applications. Current approaches to assess synthetic accessibility (SA) include computer-aided synthesis planning (CASP) tools that perform retrosynthetic searches, and machine learning-based SA prediction models that provide rapid scoring. CASP tools are computationally expensive for high-throughput screening, while existing SA prediction models may lack chemical synthesis logic or exhibit variable performance across different chemical spaces. We developed SynFrag, an SA prediction model using fragment assembly autoregressive generation to learn stepwise molecular construction patterns. Self-supervised pretraining on millions of unlabeled molecules enables learning of dynamic fragment assembly patterns beyond fragment occurrence statistics or reaction step annotations. This approach captures connectivity relationships relevant to synthesis difficulty cliffs, where minor structural changes substantially alter SA. Evaluation across public benchmarks, clinical drugs with intermediates, and AI-generated molecules shows consistent performance across diverse chemical spaces. The model produces sub-second predictions with attention mechanisms corresponding to key reactive sites. SynFrag provides computational efficiency suitable for large-scale screening while maintaining interpretability for detailed SA assessment in drug discovery workflows. Online platform: https://synfrag.simm.ac.cn. Code and data available: https://github.com/simmzx/SynFrag.