Development of the scoring function for protein ligand binding affinity by implementing graph-theory based approaches

Abstract

Scoring functions are essential to computational drug discovery since they evaluate or predict a ligand's binding affinity to its target protein. These functions rank the goodness of fit of a ligand into the binding site of a protein, thereby enabling the identification of better drug candidates. For many years, a great focus has been on developing scoring functions pointing to computational in particular graph-theory methods which are complementary to descriptors based approaches. This study, therefore, investigates different graph-theory based methods in the prediction of scoring functions for protein-ligand binding affinity. The approaches will include local and global input representations, voxel-based methods, graph-based drug-target affinity (Graph DTA) approaches, and GAT (graph attention model). Each technique has some advantage, which comes in handy in describing the complex interactions between ligands and proteins. The predicted binding affinities from each graph-theory based approach were compared to the experimental values, and R² (coefficient of determination) was used as the primary metric in the analysis. Local input representation i.e., the voxel-based approach, had an R² of 0.48, the graphDTA approach- GNN yielded an R² of 0.70, and the GAT(structure integrated graph neural network) approach yielded an R² of 0.42. The graph-based methods were observed to carry out better predictions for binding affinity. The deviation of the observed binding affinity from the expected value in the docking studies done using AutoDock Vina was further assessed, and the average deviation was 0.72, showing that while docking can be a powerful technique, some variability can be alleviated by integrating more advanced graph-theory based approaches. The improvements obtained concerning R² values from local representations through advanced graph-based methods show how these methods can be used to further improve the accuracy of computational drug discovery. Future work will be oriented toward refining these methods and their integration into comprehensive pipelines for drug discovery to accelerate the identification of promising drug candidates.