RAPID-Net is a deep learning-based binding pocket prediction algorithm that provides accurate binding pocket and feature predictions with seamless integration into docking pipelines. On the PoseBusters benchmark, RAPID-Net-based AutoDock Vina achieves 54.9% Top-1 poses that meet RMSD < 2 Å and PoseBusters chemical plausibility criteria, outperforming DiffBindFR (49.1%). Even on PoseBusters’ most challenging time slice for generalization ability evaluation (structures submitted after September 30, 2021), RAPID-Net-based AutoDock Vina achieves 53.1% performance, demonstrating that pose ranking is the main bottleneck for accuracy rather than sampling. It is suitable for large-scale virtual screening due to its lightweight inference, scalability, and competitive accuracy, and outperforms other pocket prediction tools such as PUResNet and Kalasanty. It demonstrates the potential for accelerating drug discovery through performance on pharmacologically relevant targets, and provides new opportunities for allosteric inhibitor design by accurately identifying remote functional sites. For the RNA-dependent RNA polymerase of SARS-CoV-2, while existing predictors mainly annotate only the honesty pocket and overlook the secondary cavity, RAPID-Net uncovered a broader set of potential binding pockets.
