DeepSA: A Deep-learning Driven Predictor of Compound Synthesis Accessibility
Talk, World Artificial Intelligence Conference 2023, Shanghai, China; Hangzhou, China
The difficulty of synthesizing new molecules generated by molecule generation models, i.e., the synthetic accessibility of compounds, is a key factor affecting the cost of drug development. We propose a deep learning-based chemical language model called DeepSA, which provides a useful tool for drug developers to select target synthetic molecules in real-world studies.DeepSA has a significant advantage in identifying difficult-to-synthesize molecules, with an AUROC of 89.6%, significantly better than existing methods, and with some interpretability. Meanwhile, the model was provided with only the SMILES feature information of molecules during the training process, reflecting the efficient training strategy of the model.