We are happy to introduce SmartDock, a prototype system to discriminate binders and non-binders from docking results based on Artificial Intelligence.
The pharmacologic impact of drugs depends on their ability to engage and occupy physiologically relevant target receptor binding sites. However, this binding is not static but dynamic and the pharmacologic action is determined by the temporal stability of the drug-target complex. Very often, drug-target residence time is directly linked to the ability of the drug to establish specific molecular interactions with the target protein.
In this regard, computational docking is widely used in drug discovery for the study of these interactions and also as a tool for virtual screening and profiling. However, despite recent improvements in docking and scoring methods, docking calculations are still challenged by the generation of enormous amounts of false positives. That is because very often only ligand-target binding energies are used to determine positive or negative results. As an example, Figure 1 shows that there are not real significant differences in terms of energy between active and decoys in PARP1 model.
In order to discriminate binders (true positives) and non-binders (false positives) in docking experiments, SmartDock uses information from the tons of structural data deposited on the PDB to develop an intelligent system able to learn the interactions pattern leading to experimental ligand-protein binding and, thus, to the desired biological effect. With this data, models for 700 protein targets have been built based on active and decoy compounds, extracting an interaction model for active compounds. Three decoys have been used for each active compound.
SmartDock has been validated with excellent results on PARP1, a protein involved in DNA damage repair and with an important role in both cancer and aging. Our software outperforms traditional enrichment processes, promoting thus the use of ligand docking methods as reliable virtual screening/profiling tools. Thanks to this intelligent system we build ROC curves with AUC higher than 0.9, and in the case of PARP1, AUC=9.4 (in red). This contrasts the AUC=0.75 (in blue) obtained if you only use the docking results.