Recently much effort has been invested in using convolutional neural network (CNN) models trained on 3D structural images of protein-ligand complexes to distinguish binding from non-binding ligands for virtual screening. However, the dearth of reliable protein-ligand x-ray structures and binding affinity data has required the use of constructed datasets for the training and evaluation of CNN molecular recognition models. Here, we outline various sources of bias in one such widely-used dataset, the Directory of Useful Decoys: Enhanced (DUD-E). We have constructed and performed tests to investigate whether CNN models developed using DUD-E are properly learning the underlying physics of molecular recognition, as intended, or are instead learning biases inherent in the dataset itself. We find that superior enrichment efficiency in CNN models can be attributed to the analogue and decoy bias hidden in the DUD-E dataset rather than successful generalization of the pattern of protein-ligand interactions. Comparing additional deep learning models trained on PDBbind datasets, we found that their enrichment performances using DUD-E are not superior to the performance of the docking program AutoDock Vina. Together, these results suggest that biases that could be present in constructed datasets should be thoroughly evaluated before applying them to machine learning based methodology development.

<p>Recently much effort has been invested in using convolutional neural network (CNN) models trained on 3D structural images of protein-ligand complexes to distinguish binding from non-binding ligands for virtual screening. However, the dearth of reliable protein-ligand x-ray structures and binding affinity data has required the use of constructed datasets for the training and evaluation of CNN molecular recognition models. Here, we outline various sources of bias in one such widely-used dataset, the Directory of Useful Decoys: Enhanced (DUD-E). We have constructed and performed tests to investigate whether CNN models developed using DUD-E are properly learning the underlying physics of molecular recognition, as intended, or are instead learning biases inherent in the dataset itself. We find that superior enrichment efficiency in CNN models can be attributed to the analogue and decoy bias hidden in the DUD-E dataset rather than successful generalization of the pattern of protein-ligand interactions. Comparing additional deep learning models trained on PDBbind datasets, we found that their enrichment performances using DUD-E are not superior to the performance of the docking program AutoDock Vina. Together, these results suggest that biases that could be present in constructed datasets should be thoroughly evaluated before applying them to machine learning based methodology development. </p>

Grid Inhomogeneous Solvation Theory (GIST) maps out solvation thermodynamic properties on a fine meshed grid and provides a statistical mechanical formalism for thermodynamic end-state calculations. However, differences in how long-range non-bonded interactions are calculated in molecular dynamics engines and in the current implementation of GIST have prevented precise comparisons between free energies estimated using GIST and those from other free energy methods such as thermodynamic integration (TI). Here, we address this by presenting PME-GIST, a formalism by which particle mesh Ewald (PME) based electrostatic energies and long-range Lennard-Jones (LJ) energies are decomposed and assigned to individual atoms and the corresponding voxels they occupy in a manner consistent with the GIST approach. PME-GIST yields potential energy calculations that are precisely consistent with modern simulation engines and performs these calculations at a dramatically faster speed than prior implementations. Here, we apply PME-GIST end-states analyses to 32 small molecules whose solvation free energies are close to evenly distributed from 2 kcal/mol to -17 kcal/mol and obtain solvation energies consistent with TI calculations (R2 = 0.99, mean unsigned difference 0.8 kcal/mol). We also estimate the entropy contribution from the 2nd and higher order entropy terms that are truncated in GIST by the differences between entropies calculated in TI and GIST. With a simple correction for the high order entropy terms, PME-GIST obtains solvation free energies that are highly consistent with TI calculations (R2 = 0.99, mean unsigned difference = 0.4 kcal/mol) and experimental results (R2 = 0.88, mean unsigned difference = 1.4 kcal/mol). The precision of PME-GIST also enables us to show that the solvation free energy of small hydrophobic and hydrophilic molecules can be largely understood based on perturbations of the solvent in a region extending a few solvation shells from the solute. We have integrated PME-GIST into the open-source molecular dynamics analysis software CPPTRAJ.Thermodynamic decomposition of solvation free energies with particle mesh Ewald and long-range Lennard-Jones interactions in Grid Inhomogeneous Solvation Theory