Combining docking pose rank and structure with deep learning improves protein-ligand binding mode prediction