Abstract: Traditional methods like molecular docking often face high time costs or infeasibility in large-scale virtual screening tasks. To address this problem, a parallel graph Bayesian optimization framework incorporating incremental learning is proposed to efficiently handle such tasks. The method utilizes a deep graph Bayesian optimization framework for screening and employs parallelization to enable flexible deployment across multiple computational nodes on various servers, significantly improving computational efficiency. To tackle the issue of long surrogate model training times, an incremental learning strategy is introduced, along with an exponential moving average mechanism and a replay mechanism to mitigate catastrophic forgetting in incremental learning. Experimental results demonstrate that the framework can identify over 96% of the optimal molecules by docking only 6% of the molecular library. When deployed on four computational nodes, the parallel framework reduces time costs by 71% compared to the serial framework. With the incremental learning strategy, the total runtime is further reduced by 13. 8% , while still identifying 93. 7% of the optimal molecules. The proposed method significantly reduces the time cost of virtual screening while maintaining high screening performance. 

Key words: parallel, virtual screening, Bayesian optimization, incremental learning