DeepPSA: A Geometric Deep Learning Model for PROTAC Synthetic Accessibility Prediction

Proteolysis-targeting chimeras (PROTACs) have garnered significant attention in drug design due to their ability to induce the degradation of the target proteins via the ubiquitin-proteasome system. However, the synthesis of PROTACs remains a challenging process, requiring the consideration of factors such as chemical complexity and accessibility. With the rise of generative artificial intelligence, several PROTAC generation models have been introduced, but tools to evaluate the synthetic accessibility of these molecules remain underdeveloped. To address this gap, we propose a deep learning-based computational model named DeepPSA (Deep learning-based PROTAC Synthetic Accessibility) designed to predict the synthetic accessibility of PROTACs. DeepPSA offers a systematic, data-driven approach to assess the feasibility of PROTAC synthesis, providing an essential tool for the design and screening of novel compounds. DeepPSA is a graph-based model built on a graph neural network architecture, trained on an in-house dataset of 3644 PROTACs with experimental synthetic data. As the first model specifically focused on PROTAC synthetic accessibility, DeepPSA demonstrates impressive performance, achieving 92.9% prediction accuracy and an Area Under the Receiver Operating Characteristic curve (AUROC) of 0.963 on the test set, indicating its ability to capture key structural characteristics of PROTACs. Moreover, DeepPSA continues to demonstrate superior performance on the structure-based partitioned datasets, further validating its exceptional generalization ability and robustness. DeepPSA is available online at a web server (https://bailab.siais.shanghaitech.edu.cn/psa) and at GitHub repository (https://github.com/Zhang-Ran-0119/DeepPSA).