Reverse molecular docking and deep-learning to make predictions of receptor activity for neurotoxicology

To address the need for rapid assessment of neurotoxicity from potential exposure to molecules of unknown toxicity, we developed an in silico tool that employs reverse molecular docking to identify receptor targets for molecules and deep-learning models that predict activity on the neurological targets. A selection of human neurologic receptors were obtained from the Protein Data Bank (PDB), then curated and prepared for docking. In total we docked thousands of molecules onto multiples sites on multiple different neurological receptor structures, generating millions of docked poses and scores. With this data we identified protein and ligand interactions and compared that to previously described experimental results. The data was transformed to an image representation and used to generate 2D convolutional deep-learning models. We have generated 19 deep-learning models, of which 17 are over 90% accurate on validation data and the remaining two are 84% and 87% accurate. We have developed a reverse docking GUI and pipeline to identify potential neurological targets for toxins and predict activity of toxins with deep-learning models based on docking identified interactions as an input. As an example, we have applied this pipeline to toluene, a molecule with known toxicity, and correctly predicted it as a GABA(B) agonist. The GUI has been tested on Ubuntu 20.04LTS and Windows 10, and the code, models and GUI are available under GPLv3 on github at https://github.com/mmccarthy1/Autodock_deeplearning_toxicology_GUI.