Fig.1

Shannon entropies based on standard tokens and characters derived from a string representation of molecules are efficient descriptors for deep neural network-based property predictions. a Comparison of network performance with the addition of different Shannon entropies in the descriptor set. IC₅₀ values of tissue factor pathway inhibitor were predicted and analyzed using MAPE, MAE and R² of fit metrics. The descriptor set containing MW, Shannon and fractional Shannon entropies extracted from SMILES showed the best performance in comparison to other descriptors in the triangular radar graph. b Cumulative enhancement of network performance using Shannon descriptors depicted in the radar graph. The target was MW normalized BEI of ligands to the tissue factor pathway inhibitor, i.e. in the form of BEI/MW. The SEF set containing MW, Shannon (SMILES) and fractional Shannon (SMILES) showed the best comparative performance in all metrics. c Comparison of direct one-pot vs tandem approach to predict IC₅₀ values of molecules to the tissue factor pathway inhibitor protein. The tandem approach first predicted BEI as an intermediate step and then predicted IC₅₀ values at higher accuracy with the BEI as an input. The model was based on MLP-based deep neural networks and all prediction metrics were averaged over at least 5 independent runs. The scaling factors of metrics were listed in Additional file 1: Table S3