S2: Quantitative Target-specific Toxicity Prediction Modelling

Quantitative Target-specific Toxicity Prediction Modelling, OpenTox USA 2018
PRESENTING AUTHOR: 

Ping Gong

AUTHOR(S): 

Ping Gong, Sundar Thangapandian, Gabriel Idakwo, Nan Wang, and Chaoyang Zhang

REFERENCES: 
1* Environmental Laboratory, US Army Engineer Research and Development Center, Vicksburg, MS 39180
2 School of Computing, University of Southern Mississippi, Hattiesburg, MS 39406
3 Department of Computer Science, New Jersey City University, Jersey City, NJ 07305
ABSTRACT CONTENT / DETAILS: 

Background: Quantitative structure-activity relationship (QSAR) modelling is a chemical descriptors-based approach for quantitative prediction of biological activity, potency or toxicity of a chemical. QSAR modelling may suffer low prediction accuracy in the absence of information on chemical-biomacromolecule interactions. In order to mitigate this problem, we developed a novel Quantitative Target-specific Toxicity Prediction Model (QTTPM) approach that integrated molecular dynamics (MD) simulation and machine learning. As a proof-of-concept study, we chose androgen receptor (AR) as the toxicant-targeted biomacromolecule because AR is a nuclear receptor playing crucial roles in the development of male reproductive system and tumors in prostate, bladder, liver, kidney and lung. Molecular docking and MD simulations were employed to generate a new set of dynamic protein-ligand interaction descriptors (dyPLIDs) used for developing QTTPMs. We selected 274 chemicals (154 agonists/120 antagonists) with quantitative AR assay outcomes from Tox21 datasets. First, we performed five 100-ns MD simulations of AR crystal structures in its un-bound (apo), two agonist-bound (testosterone and dihydrotestostrone), and two antagonist-bound (R-bicalutamide and cyproterone acetate) forms and identified key interaction patterns leading to >400 dyPLIDs. Second, 6-ns MD simulations of 274 AR-ligand docked complexes were performed to calculate dyPLIDs. Third, Random Forest (RF) algorithm was deployed to identify key descriptors (including both conventional 1D/2D/3D descriptors and dyPLIDs). Fourth, QTTPMs were built using the key descriptors and AR assay data.

Results: QTTPMs demonstrated superior accuracy than QSAR models constructed with conventional chemical descriptors. In addition, QTTPMs provided insights of key protein structural changes upon ligand binding that modulated the activity of the AR.

Conclusions: The novel QTTPM approach was developed using a small dataset of 274 AR agonists/antagonists. Although more biomacromolecular targets and chemicals warrant further investigations, this study demonstrates the superiority of QTTPM over QSAR and that QTTPM is a promising new tool for computational predictive toxicology.