# NeuroDeRisk Models¶

The NeuroDeRisk in silico toolbox includes 3D-models developed for profiling chemical structures for prediction of the potential risk of neurotoxic adverse outcomes such as, convulsant and seizure-inducing effects, psychological/psychiatric effects and peripheral neuropathies. They are included in the NeuroDeRisk IL Profiler node.

Info

The models are evolving and new model development is ongoing during the period of the NeuroDeRisk project. As new information about the models evolves and can be made available, the documentation will be updated.

## Model Names¶

The names of the 3D-pharmacophore models in the NeuroDeRisk IL profiler contain information about the model. For example, the partner involved in developing the model, the target or adverse outcome that the model has been designed to predict if macromolecular target information (structure-based) or only ligand information (ligand-based) was used to create the model. For structure-based models the protein data bank code (PDB ID) has been included.

Additional codes, both letters and numbers at the end of the name are for internal use and version tracking for the model database. Text included in the names of the models can be used in the NeuroDeRisk IL Profiler node to search (using the text box) for subsets of models to use for your profiling/prediction. They also give guidance on the expected outcomes they have been designed to predict.

Examples and explanations of abbreviations are given below.

### Examples¶

#### NDR-IL-GABA-A-gs-Agonist-6huj-4¶

NeuroDeRisk-Inte:Ligand-GABA-A Receptor-GABA-A orthosteric binding site-Agonist outcome-[PDB ID: 6huj]-[internal database code:4]


The model has been developed to predict chemical structures with likelihood to interact with the GABA-A receptor and exhibit selectivity for the orthosteric GABA binding site with agonist pharmacological outcomes. The model is a structure-based model with information on binding interactions derived from a structure reported in the protein data bank with the PDB ID: 6huj.

#### NDR-IL- GABA-A-gs-Antag-LB¶

NeuroDeRisk-Inte:Ligand-GABA-A Receptor-GABA-A orthosteric binding site-Antagonist-outcome-Ligand-Based.


This model has been developed to predict chemical structures with likelihood to interact with the GABA-A receptor and exhibit selectivity for the orthosteric GABA binding site with antagonist pharmacological outcomes. The model is a ligand-based model (LB) that was developed using GABA-A GABA site antagonist ligands.

#### NDR-UV-PNS-Vinca-LB¶

NeuroDeRisk-UniVie-Peripheral-Vinca-Alkaloids-LB


This model has been developed to predict chemical structures with likelihood to produce PNS neuropathies. The model is a ligand-based model (LB) that was developed using vinca alkaloid ligand structures

### Used Abbreviations¶

Method Description
GABA-A GABA-A receptor
gs GABA site (GABA-A receptor gamma-aminobutyric acid binding site)
IL Inte:Ligand
LB Ligand-based
NAM Negative allosteric modulator
NDR NeuroDeRisk
NSteroid Neurosteroid
PAM Positive allosteric modulator
UV University of Vienna
PNS Peripheral Nervous System
VincaA Vinca alkaloid

## General Methodology¶

While the methodology for developing each model varies in terms of detail, concepts underlying the creation of the models are similar and will be included herein. Details involving the specific methodology applied towards the development of each model will be reported in research publications. Following publication, we will update the information in this documentation accordingly.

### Software¶

LigandScout 4.4 Expert 1 and the IL Expert KNIME Extensions 2 were used for the generation of 3D-structure-based (SB) and ligand-based (LB) chemical feature-based pharmacophore models as well as for virtual screening, activity profiling, calculation of the receiver operating characteristic (ROC) curves, extraction of data from public data sources, creation and calculation of compound databases and creation of datasets for model training and testing.

#### LigandScout Chemical Feature Based Pharmacophores¶

LigandScout 3D-pharmacophore models represent an ensemble of chemical features responsible for blocking and/or eliciting a biological response. Chemical feature types include, hydrogen bond donors and acceptors (including directional vectors and non-directional spheres), aromatic (cation-pi, pi-pi (orthogonal and parallel), hydrophobic interactions, positive and negative ionizable features, halogen bond donor, covalent bond interactions and metal ion coordination. In addition exclusion volume features, are included in the models that define restricted areas in a binding site or restricted areas around a set of ligands. Exclusion volumes are important for increasing the sensitivity of pharmacophore models by reducing false positive hit rates and increasing enrichment in virtual screening and activity profiling experiments. For additional theory see reference 3.

The 3D-pharmacophore models developed in the NeuroDeRisk project consist of various combinations of these chemical feature types and have been determined using curated datasets of chemical structures exhibiting the target and/or neurotoxic adverse outcomes being modeled. In addition, each model has been manually refined, including chemical feature adjustments to achieve the desired predictive properties.

#### Alignments and pharmacophore fit scores¶

LigandScout uses a unique pattern-matching algorithm for alignment of chemical structures based on pharmacophore features rather than scaffolds 4. It is used for rapid and accurate virtual screening. However, it is also enables the clustering of molecules based on pharmacophore space that is useful for the generation of LB models especially applicable for associating chemical features from a set of molecules with adverse outcomes. Furthermore, the algorithm is used for profiling of chemical structures in the NeuroDeRisk IL Profiler tool. The algorithm gives a relative pharmacophore fit score (0 -1) that indicates how well the chemical structure fits to the pharmacopohore model. The higher the fit score the better the fit to the model.

### GABA-A Modeling¶

Chemical structures annotated with GABA-A experimental outcomes were extracted from the ChEMBL database 5 and manually from literature. The data was curated and separated into datasets to be used for model creation for each model type, model assessment and testing. For SB-models, structural data involving the GABAA receptor was downloaded from the RCSB Protein Data Bank 6. Both LB and SB GABA-A models developed and incorporated in the NeuroDeRisk Computational Toolbox thus far are described in the Models section.

### Suicidality Modeling¶

Chemical structures of drugs associated with reports of suicality related psychological/psychiatric effects were collected from several sources including literature articles and publically available databases, such as FAERS 7 and METADEDB 8. The data was curated and annotated with chemical structure information and clustered by pharmacophore space. A large number of LB models were generated based on pharmacophore clustering, chemical properties, frequency of reports, black box warning labels and other properties and assessed based on overall hit rates, profiles and drugs retrieved. In addition, datasets were assembled based on 17 interferons like RNA editing drugs based on the Editox® assay developed by Alcediag team 9. The resulting selection of best performing models have been included in the NeuroDeRisk Profiler and are listed in the Models section.

### PNS Modeling¶

In the case of PNS models, data sets were generated from compounds designated as exerting peripheral neuropathies. Chemical structures were extracted manually from literature. The data was curated and separated into datasets by LigandScout pharmacophore-based clustering algorithms. Model validation was performed using ROC curve calculation based on retrieval rates of true actives versus false positives.

## GABA-A Models¶

GABA-A receptors have been major drug targets for treating CNS disorders, including anxiety, epilepsy, and insomnia. Widely prescribed drugs, like benzodiazepines, barbiturates, anesthetics, neurosteroids, and convulsants bind to distinct allosteric sites to exert their physiological effects. Orthosteric agonists of GABA-A receptors exhibit antiepileptic, sleep-inducing, hypnotic, and analgesic properties while competitive antagonists, such as, bicuculline and gabazine as well as channel blockers like picrotoxin antagonize GABA-A receptors leading to convulsions or epileptic seizures. Though GABA-A receptor drugs have been clinically beneficial, they also pose serious risks of addiction, physical dependence and life-threatening events due to misuse, overdose and withdrawal effects. Enormous efforts have been made to unravel the complexity and variable subunit composition of GABA-A receptors and the underlying pharmacology of ligands associated with their various binding pockets in order to avoid undesirable effects and improve safety profiles.

In the NeuroDeRisk project we are developing a panel of predictive GABA-A pharmacophore models to support efforts to understand the underlying pharmacology of chemical structures and their interactions with different GABA-A receptor binding sites and derisk compounds in preclinical drug discovery for adverse effects associated with GABA-A interactions.

The following GABA-A models have been developed so far and integrated into the NeuroDeRisk IL Profiler for use by partners.

NDR-IL-GABA-A-gs-Agonist-6huj-4
Structure-based model to predict chemical structures with likelihood to interact with the GABA-A receptor and exhibit selectivity for the orthosteric GABA binding site with agonist pharmacological outcomes.
NDR-IL-GABA-A-gs-Agonist-LB
Ligand-based model to predict chemical structures with likelihood to interact with the GABA-A receptor and exhibit selectivity for the orthosteric GABA binding site with agonist pharmacological outcomes.
NDR-IL-GABA-A-gs-Antag-6huk-3
Structure-based model to predict chemical structures with likelihood to interact with the GABA-A receptor and exhibit selectivity for the orthosteric GABA binding site with antagonist pharmacological outcomes.
NDR-IL-GABA-A-gs-Antag-LB
Ligand-based model to predict chemical structures with likelihood to interact with the GABA-A receptor and exhibit selectivity for the orthosteric GABA binding site with antagonist pharmacological outcomes.
NDR-IL-GABA-A-Channel-LB-5
Prediction of chemical structures with likelihood to block the GABA-A receptor channel.
NDR-IL-GABA-NAM-Flumazenil-6d6t
Prediction of chemical structures with likelihood to exhibit negative allosteric modulation of the GABA-A receptor similar to flumazenil.
NDR-IL-GABA-PAM-Diazepam-6hup
Prediction of chemical structures with likelihood to exhibit positive allosteric modulation of the GABA-A receptor similar to diazepam.
NDR-IL-GABA-PAM-Flurazepam-2yoe
Prediction of chemical structures with likelihood to exhibit positive allosteric modulation of the GABA-A receptor similar to flurazepam.
NDR-IL-GABA-NSteroid-Pregnenolone-5o8f
Prediction of chemical structures with likelihood to exhibit neurosteroid-like modulation of the GABA-A receptor similar to pregnenolone.

## Suicidality Models¶

Adverse drug reactions and safety issues related to central nervous system (CNS) drugs have resulted in terminated clinical development of numerous compounds. Treatment emergent suicidal ideation has emerged as a serious adverse outcome and few tools exist to assess or predict it. To further complicate matters, anti-depressant SSRIs with black box warning labels required by the FDA since 2004 are not the only class of drugs associated with suicidal ideation. Published case reports related to treatment emergent suicidal ideation and behavior have involved beta-blockers, asthma, smoking cessation, weight loss drugs and others.

In the NeuroDeRisk project we are developing a panel of pharmacophore models to support efforts to understand and predict psychological and psychiatric side effects (P/P-Tox). The following set of pharmacophore models have been developed so far and are included in the NeuroDeRisk IL Profiler.

NDR-IL-Suicidality-SE-ed-1
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes. Selective model.
NDR-IL-Suicidality-2
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes.
NDR-IL-Suicidality-3
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes.
NDR-IL-Suicidality-3v1
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes.
NDR-IL-Suicidality-4
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes.
NDR-IL-Suicidality-5
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes.
NDR-IL-Suicidality-SE-sd-6
Prediction of chemical structures with likelihood to be associated with suicidal ideation adverse outcomes. Selective model.

## PNS Models¶

NDR-UV-PNS-Bendam-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to bendamustine.
NDR-UV-PNS-CarfilTacrol-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to carfilzomib or tacrolimus.
NDR-UV-PNS-Conazols-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to antimycotics of the conazole type.
NDR-UV-PNS-EribEto-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to eribuline-type macrocycles or podophyllotoxin analogs.
NDR-UV-PNS-Ixabepil-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to epothilone-type macrocycles.
NDR-UV-PNS-M9-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to thalidomide derivatives.
NDR-UV-PNS-M6-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to amiodarone.
NDR-UV-PNS-M18-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to antiviral and anticancer nucleoside analogs.
NDR-UV-PNS-Mefloq-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to mefloquine.
NDR-UV-PNS-Omibs-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to bortezomib.
NDR-UV-PNS-Procarb-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to procarbazine.
NDR-UV-PNS-Snibs-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to PNS toxic kinase inhibitors.
NDR-UV-PNS-Taxels-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to taxol-derived anticancer agents
NDR-UV-PNS-VincaA-LB
Prediction of chemical structures with likelihood to be associated with inducing peripheral neuropathies similar to vinca alkaloids derived anticancer agents.

1. a) LigandScout 4.4 Expert. Available from Inte:Ligand GmbH, Vienna, Austria. http://www.inteligand.com/ligandscout.
b) Wolber, G.; Langer, T. LigandScout: 3-D Pharmacophores Derived from Protein-Bound Ligands and Their Use as Virtual Screening Filters. J. Chem. Inf. Model. 2005, 45 (1), 160–169. https://doi.org/10.1021/ci049885e

2. LigandScout Extensions for the KNIME Workbench. Available from Inte:Ligand GmbH, Vienna, Austria.
https://hub.knime.com/inteligand/extensions/com.inteligand.knime/latest

3. Seidel, T.; Bryant, S. D.; Ibis, G.; Poli, G.; Langer, T. 3D Pharmacophore Modeling Techniques in Computer-Aided Molecular Design Using LigandScout. Tutorials in Chemoinformatics 2017, 279–309. https://doi.org/10.1002/9781119161110.ch20

4. Wolber, G.; Seidel, T.; Bendix, F.; Langer, T. Molecule-Pharmacophore Superpositioning and Pattern Matching in Computational Drug Design. Drug Discovery Today. 2008, 13 (1-2), 23–29. https://doi.org/10.1016/j.drudis.2007.09.007

5. a) ChEMBL Database https://www.ebi.ac.uk/chembl/;
b) Gaulton, A.; Hersey, A.; Nowotka, M.; Bento, A. P.; Chambers, J.; Mendez, D.; Mutowo, P.; Atkinson, F.; Bellis, L. J.; Cibrián-Uhalte, E., et al., The ChEMBL Database in 2017. Nucleic Acids Res. 2017, 45 (D1), D945–D954. https://doi.org/10.1093/nar/gkw1074

6. a) RCSB Protein Databank. https://www.rcsb.org/;
b). Berman, H. M. The Protein Data Bank. Nucleic Acids Res. 2000, 28 (1), 235–242. https://doi.org/10.1093/nar/28.1.235

7. Food and Drug Administration, FDA AEs reporting system (FAERS) public dashboard. https://www.fda.gov/Drugs/GuidanceComplianceRegulatoryInformation/Surveillance/AdverseDrugEffects/ucm070093.htm

8. a). Zhuohang Yu, Zengrui Wu*, Weihua Li, Guixia Liu, Yun Tang*. MetaADEDB 2.0: a comprehensive database on adverse drug events. Bioinformatics. Under Review;
b). Feixiong Cheng, Weihua Li, Xichuan Wang, Yadi Zhou, Zengrui Wu, Jie Shen, Yun Tang*. Adverse Drug Events: Database Construction and in Silico Prediction. J. Chem. Inf. Model., 2013, 53 (4): 744-752.

9. a) Alcediag EDITOX®. Available at: http://www.alcediag-alcen.com/fr
b) Van der Laan, S., Salvetat, N., Weissmann, D., & Molina, F. (2017). Emerging RNA editing biomarkers will foster drug development. Drug Discovery Today, 22(7), 1056 1063. doi:10.1016/j.drudis.2017.01.017 b). Cavarec, L. et al. (2013) In vitro screening for drug-induced depression and/or suicidal adverse effects: a new toxicogenomic assay based on CE-SSCP analysis of HTR2C mRNA editing in SH- SY5Y cells. Neurotox. Res. 23, 49–62.