Table 1 The best performance model by training four architectures with three featurization methods as input. Feature1 used a resolution of 20 × 20 × 20 and retained all atoms. Feature2 used a resolution of 24 × 24 × 24 and also retained all atoms. Feature3 used a resolution of 20 × 20 × 20 but ignored hydrogen and metal atoms

From: Sfcnn: a novel scoring function based on 3D convolutional neural network for accurate and stable protein–ligand affinity prediction