Complementary PLS and KNN algorithms for improved 3D-QSDAR consensus modeling of AhR binding

Determination of structural factors affecting binding to mu, kappa and delta opioid receptors

Addiction is a complex behavioral phenomenon in which naturally occurring or synthetic chemicals modulate the response of the reward system through their binding to a variety of neuroreceptors, resulting in compulsive substance-seeking and use despite harmful consequences to the individual. Among these, the opioid receptor (OR) family and more specifically, the mu-opioid receptor (MOR) subtype plays a critical role in the addiction to powerful prescription and illicit drugs such as hydrocodone, oxycodone, fentanyl, cocaine, and methamphetamine (Contet et al. in Curr Opin Neurobiol 14(3):370-378, 2004). Conversely, agonists binding to kappa (KOR) and antagonists binding to delta opioid receptors (DOR) have been reported to induce negative reinforcing effects. As more than 700 new psychoactive substances were illegally sold between 2009 and 2016 (DEA-DCT-DIR-032-18), most of them lacking basic toxicological and pharmacological profiles, molecular modeling approaches that could quickly and reliably fill the gaps in our knowledge would be highly desirable tools for determining the effects of these synthetics. Here, we report accurate 3D-spectrometric data-activity relationship classification models for large and diverse datasets of MOR, KOR and DOR binders with areas under the receiver operating characteristic curve for the "blind" prediction sets exceeding 0.88. Structural features associated with (selective) binding to MOR, KOR and/or DOR were identified. These models could assist regulatory agencies in evaluating the health risks associated with the use of unprofiled substances as well as to help the pharmaceutical industry in its search for new drugs to combat addiction.

Computational prediction models for assessing endocrine disrupting potential of chemicals

Endocrine disrupting chemicals (EDCs) mimic natural hormones and disrupt endocrine function. Humans and wildlife are exposed to EDCs might alter endocrine functions through various mechanisms and lead to an adverse effects. Hence, EDCs identification is important to protect the ecosystem and to promote the public health. Leveraging in-vitro and in-vivo experiments to identify potential EDCs is time consuming and expensive. Hence, quantitative structure-activity relationship is applied to screen the potential EDCs. Here, we summarize the predictive models developed using various algorithms to forecast the binding activity of chemicals to the estrogen and androgen receptors, alpha-fetoprotein, and sex hormone binding globulin.

Why are most phospholipidosis inducers also hERG blockers?

Recent reports have noted that a number of compounds that block the human Ether-à-go-go related gene (hERG) ion channel also induce phospholipidosis (PLD). To explore a hypothesis explaining why most PLD inducers are also hERG inhibitors, a modeling approach was undertaken with data sets comprised of 4096 compounds assayed for hERG inhibition and 5490 compounds assayed for PLD induction. To eliminate the chemical domain effect, a filtered data set of 567 compounds tested in quantitative high-throughput screening (qHTS) format for both hERG inhibition and PLD induction was constructed. Partial least squares (PLS) modeling followed by 3D-SDAR mapping of the most frequently occurring bins and projection on to the chemical structure suggested that both adverse effects are driven by similar structural features, namely two aromatic rings and an amino group forming a three-center toxicophore. Non-parametric U-tests performed on the original 3D-SDAR bins indicated that the distance between the two aromatic rings is the main factor determining the differences in activity; at distances of up to about 5.5 Å, a phospholipidotic compound would also inhibit hERG, while at longer distances, a sharp reduction of the PLD-inducing potential leaves only a well-pronounced hERG blocking effect. The hERG activity itself diminishes after the distance between the centroids of the two aromatic rings exceeds 12.5 Å. Further comparison of the two toxicophores revealed that the almost identical aromatic rings to amino group distances play no significant role in distinguishing between PLD and hERG activity. The hypothesis that the PLD toxicophore appears to be a subset of the hERG toxicophore explains why about 80% of all phospholipidotic chemicals (the remaining 20% are thought to act via a different mechanism) also inhibit the hERG ion channel. These models were further validated in large-scale qHTS assays testing 1085 chemicals for their PLD-inducing potential and 1570 compounds for hERG inhibition. After removal of the modeling and experimental inconclusive compounds, the area under the receiver-operating characteristic (ROC) curve was 0.92 for the PLD model and 0.87 for the hERG model. Due to the exceptional ability of these models to recognize safe compounds (negative predictive values of 0.99 for PLD and 0.94 for hERG were achieved), their use in regulatory settings might be particularly useful.

Rigorous 3-dimensional spectral data activity relationship approach modeling strategy for ToxCast estrogen receptor data classification, validation, and feature extraction

The estrogenic potential (expressed as a score composite of 18 high throughput screening bioassays) of 1528 compounds from the ToxCast database was modeled by a 3-dimensional spectral data activity relationship approach (3D-SDAR). Due to a lack of ¹⁷ O nuclear magnetic resonance (NMR) simulation software, the most informative carbon-carbon 3D-SDAR fingerprints were augmented with indicator variables representing oxygen atoms from carbonyl and carboxamide, ester, sulfonyl, nitro, aliphatic hydroxyl, and phenolic hydroxyl groups. To evaluate the true predictive performance of the authors' model the United States Environmental Protection Agency provided them with a blind test set consisting of 2008 compounds. Of these, 543 had available literature data-their binding affinity served to estimate the external classification accuracy of the developed model: predictive accuracy of 0.62, sensitivity of 0.71, and specificity of 0.53 were obtained. Compared with alternative modeling techniques, the authors' model displayed very little reduction in performance between the modeling and the prediction set. A 3D-SDAR mapping technique allowed identification of structural features essential for estrogenicity: 1) the presence of a phenolic OH group or cyclohexenone, 2) a second aromatic or phenolic ring at a distance of 6 Å to 8 Å from the oxygen of the first phenol ring, 3) the presence of a methyl group approximately 6 Å away from the centroid of a phenol ring, and 4) a carbonyl group in close proximity (∼4 Å measured to the centroid) to 1 of the phenol rings. Environ Toxicol Chem 2017;36:823-830. Published 2016 Wiley Periodicals Inc. on behalf of SETAC. This article is a US government work and, as such, is in the public domain in the United States of America.

An automated curation procedure for addressing chemical errors and inconsistencies in public datasets used in QSAR modelling

The increasing availability of large collections of chemical structures and associated experimental data provides an opportunity to build robust QSAR models for applications in different fields. One common concern is the quality of both the chemical structure information and associated experimental data. Here we describe the development of an automated KNIME workflow to curate and correct errors in the structure and identity of chemicals using the publicly available PHYSPROP physicochemical properties and environmental fate datasets. The workflow first assembles structure-identity pairs using up to four provided chemical identifiers, including chemical name, CASRNs, SMILES, and MolBlock. Problems detected included errors and mismatches in chemical structure formats, identifiers and various structure validation issues, including hypervalency and stereochemistry descriptions. Subsequently, a machine learning procedure was applied to evaluate the impact of this curation process. The performance of QSAR models built on only the highest-quality subset of the original dataset was compared with the larger curated and corrected dataset. The latter showed statistically improved predictive performance. The final workflow was used to curate the full list of PHYSPROP datasets, and is being made publicly available for further usage and integration by the scientific community.

CERAPP: Collaborative Estrogen Receptor Activity Prediction Project

BACKGROUND

Humans are exposed to thousands of man-made chemicals in the environment. Some chemicals mimic natural endocrine hormones and, thus, have the potential to be endocrine disruptors. Most of these chemicals have never been tested for their ability to interact with the estrogen receptor (ER). Risk assessors need tools to prioritize chemicals for evaluation in costly in vivo tests, for instance, within the U.S. EPA Endocrine Disruptor Screening Program.

OBJECTIVES

We describe a large-scale modeling project called CERAPP (Collaborative Estrogen Receptor Activity Prediction Project) and demonstrate the efficacy of using predictive computational models trained on high-throughput screening data to evaluate thousands of chemicals for ER-related activity and prioritize them for further testing.

METHODS

CERAPP combined multiple models developed in collaboration with 17 groups in the United States and Europe to predict ER activity of a common set of 32,464 chemical structures. Quantitative structure-activity relationship models and docking approaches were employed, mostly using a common training set of 1,677 chemical structures provided by the U.S. EPA, to build a total of 40 categorical and 8 continuous models for binding, agonist, and antagonist ER activity. All predictions were evaluated on a set of 7,522 chemicals curated from the literature. To overcome the limitations of single models, a consensus was built by weighting models on scores based on their evaluated accuracies.

RESULTS

Individual model scores ranged from 0.69 to 0.85, showing high prediction reliabilities. Out of the 32,464 chemicals, the consensus model predicted 4,001 chemicals (12.3%) as high priority actives and 6,742 potential actives (20.8%) to be considered for further testing.

CONCLUSION

This project demonstrated the possibility to screen large libraries of chemicals using a consensus of different in silico approaches. This concept will be applied in future projects related to other end points.

CITATION

Mansouri K, Abdelaziz A, Rybacka A, Roncaglioni A, Tropsha A, Varnek A, Zakharov A, Worth A, Richard AM, Grulke CM, Trisciuzzi D, Fourches D, Horvath D, Benfenati E, Muratov E, Wedebye EB, Grisoni F, Mangiatordi GF, Incisivo GM, Hong H, Ng HW, Tetko IV, Balabin I, Kancherla J, Shen J, Burton J, Nicklaus M, Cassotti M, Nikolov NG, Nicolotti O, Andersson PL, Zang Q, Politi R, Beger RD, Todeschini R, Huang R, Farag S, Rosenberg SA, Slavov S, Hu X, Judson RS. 2016.

CERAPP

Collaborative Estrogen Receptor Activity Prediction Project. Environ Health Perspect 124:1023-1033; http://dx.doi.org/10.1289/ehp.1510267.

Alignment-independent technique for 3D QSAR analysis

Molecular biochemistry is controlled by 3D phenomena but structure-activity models based on 3D descriptors are infrequently used for large data sets because of the computational overhead for determining molecular conformations. A diverse dataset of 146 androgen receptor binders was used to investigate how different methods for defining molecular conformations affect the performance of 3D-quantitative spectral data activity relationship models. Molecular conformations tested: (1) global minimum of molecules' potential energy surface; (2) alignment-to-templates using equal electronic and steric force field contributions; (3) alignment using contributions "Best-for-Each" template; (4) non-energy optimized, non-aligned (2D > 3D). Aggregate predictions from models were compared. Highest average coefficients of determination ranged from R Test (2) = 0.56 to 0.61. The best model using 2D > 3D (imported directly from ChemSpider) produced R Test (2) = 0.61. It was superior to energy-minimized and conformation-aligned models and was achieved in only 3-7 % of the time required using the other conformation strategies. Predictions averaged from models built on different conformations achieved a consensus R Test (2) = 0.65. The best 2D > 3D model was analyzed for underlying structure-activity relationships. For the compound strongest binding to the androgen receptor, 10 substructural features contributing to binding were flagged. Utility of 2D > 3D was compared for two other activity endpoints, each modeling a medium sized data set. Results suggested that large scale, accurate predictions using 2D > 3D SDAR descriptors may be produced for interactions involving endocrine system nuclear receptors and other data sets in which strongest activities are produced by fairly inflexible substrates.

Partial least square and k-nearest neighbor algorithms for improved 3D quantitative spectral data-activity relationship consensus modeling of acute toxicity

A diverse set of 154 chemicals that included US Food and Drug Administration-regulated compounds tested for their aquatic toxicity in Daphnia magna were modeled by a 3-dimensional quantitative spectral data-activity relationship (3D-QSDAR). Two distinct algorithms, partial least squares (PLS) and Tanimoto similarity-based k-nearest neighbors (KNN), were used to process bin occupancy descriptor matrices obtained after tessellation of the 3D-QSDAR space into regularly sized bins. The performance of models utilizing bins ranging in size from 2 ppm × 2 ppm × 0.5 Å to 20 ppm × 20 ppm × 2.5 Å was explored. Rigorous quality-control criteria were imposed: 1) 100 randomized 20% hold-out test sets were generated and the average R(2) test of the respective models was used as a measure of their performance, and 2) a Y-scrambling procedure was used to identify chance correlations. A consensus between the best-performing composite PLS model using 0.5 Å × 14 ppm × 14 ppm bins and 10 latent variables (average R(2) test  = 0.770) and the best composite KNN model using 0.5 Å × 8 ppm × 8 ppm and 2 neighbors (average R(2) test  = 0.801) offered an improvement of about 7.5% (R(2) test consensus  = 0.845). Projection of the most frequently occurring bins on the standard coordinate space indicated that the presence of a primary or secondary amino group-substituted aromatic systems-would result in an increased toxic effect in Daphnia. The presence of a second aromatic ring with highly electronegative substituents 5 Å to 7 Å apart from the first ring would lead to a further increase in toxicity.