PLANET: A graph neural network model with multi-training objectives for predicting protein–ligand binding affinity

Predicting protein–ligand binding affinity is a central issue in drug design, and it is therefore no surprise that various deep learning models have been developed in recent years to tackle this issue in one aspect or another. So far, most of them merely focused on reproducing the binding affinity of known binders, resulting into a tiny applicability domain. To tackle this problem, we developed a graph neural network model called PLANET (Protein-Ligand Affinity prediction NETwork). This model takes the graph-represented 3D structure of the binding pocket on the target protein and the 2D structural graph of the ligand molecule as inputs. PLANET was trained through a multi-objective process with three related tasks: deriving the protein–ligand binding affinity, protein–ligand contact map, and intra-ligand distance matrix. To serve those tasks, artificially generated decoys were used in addition to the standard PDBbind data sets normally employed for training scoring functions. When tested on the CASF-2016 benchmark, PLANET exhibited a scoring power comparable to that of other machine learning models that rely on 3D protein–ligand complex structures as inputs. It also showed notably better performance in virtual screening trials on the DUD-E and LIT-PCBA benchmarks. Compared to the popular conventional docking program Glide, PLANET took less than 1% of the computation time to finish the same virtual screening job without a significant loss in accuracy because it did not need to perform exhaustive conformational sampling. In summary, PLANET achieved decent performance in virtual screening and predicting protein–ligand binding affinity, making PLANET an attractive tool for real-world drug discovery.