Computational Prediction of Metabolic α-Carbon Hydroxylation Potential of N-Nitrosamines: Overcoming Data Limitations for Carcinogenicity Assessment.

Recent withdrawal of several drugs from the market due to elevated levels of N-nitrosamine impurities underscores the need for computational approaches to assess the carcinogenicity risk of nitrosamines. However, current approaches are limited because robust animal carcinogenicity data are only available for a few simple nitrosamines, which do not represent the structural diversity of the many possible nitrosamine drug substance related impurities (NDSRIs). In this paper, we present a novel method that uses data on CYP-mediated metabolic hydroxylation of CH2 groups in non-nitrosamine xenobiotics to identify structural features that may also help in predicting the likelihood of metabolic α-carbon hydroxylation in N-nitrosamines. Our approach offers a new avenue for tapping into potentially large experimental data sets on xenobiotic metabolism to improve risk assessment of nitrosamines. As α-carbon hydroxylation is the crucial rate-limiting step in nitrosamine metabolic activation, identifying and quantifying the influence of various structural features on this step can provide valuable insights into their carcinogenic potential. This is especially important considering the scarce information available on factors that affect NDSRI metabolic activation. We have identified hundreds of structural features and calculated their impact on hydroxylation, a significant advancement compared to the limited findings from the small nitrosamine carcinogenicity data set. While relying solely on α-carbon hydroxylation prediction is insufficient for forecasting carcinogenic potency, the identified features can help in the selection of relevant structural analogues in read across studies and assist experts who, after considering other factors such as the reactivity of the resulting electrophilic diazonium species, can establish the acceptable intake (AI) limits for nitrosamine impurities.

Augmenting Expert Knowledge-Based Toxicity Alerts by Statistically Mined Molecular Fragments.

Structural alerts are molecular substructures assumed to be associated with molecular initiating events in various toxic effects and an integral part of in silico toxicology. However, alerts derived using the knowledge of human experts often suffer from a lack of predictivity, specificity, and satisfactory coverage. In this work, we present a method to build hybrid QSAR models by combining expert knowledge-based alerts and statistically mined molecular fragments. Our objective was to find out if the combination is better than the individual systems. Lasso regularization-based variable selection was applied on combined sets of knowledge-based alerts and molecular fragments, but the variable elimination was only allowed to happen on the molecular fragments. We tested the concept on three toxicity end points, i.e., skin sensitization, acute Daphnia toxicity, and Ames mutagenicity, which covered both classification and regression problems. Results showed the predictive performance of such hybrid models is, indeed, better than the models based solely on expert alerts or statistically mined fragments alone. The method also enables the discovery of activating and mitigating/deactivating features for toxicity alerts and the identification of new alerts, thereby reducing false positive and false negative outcomes commonly associated with generic alerts and alerts with poor coverage, respectively.

Development of QSAR models to predict blood-brain barrier permeability.

Assessing drug permeability across the blood-brain barrier (BBB) is important when evaluating the abuse potential of new pharmaceuticals as well as developing novel therapeutics that target central nervous system disorders. One of the gold-standard in vivo methods for determining BBB permeability is rodent log BB; however, like most in vivo methods, it is time-consuming and expensive. In the present study, two statistical-based quantitative structure-activity relationship (QSAR) models were developed to predict BBB permeability of drugs based on their chemical structure. The in vivo BBB permeability data were harvested for 921 compounds from publicly available literature, non-proprietary drug approval packages, and University of Washington's Drug Interaction Database. The cross-validation performance statistics for the BBB models ranged from 82 to 85% in sensitivity and 80-83% in negative predictivity. Additionally, the performance of newly developed models was assessed using an external validation set comprised of 83 chemicals. Overall, performance of individual models ranged from 70 to 75% in sensitivity, 70-72% in negative predictivity, and 78-86% in coverage. The predictive performance was further improved to 93% in coverage by combining predictions across the two software programs. These new models can be rapidly deployed to predict blood brain barrier permeability of pharmaceutical candidates and reduce the use of experimental animals.

MultiCASE Platform for In Silico Toxicology.

Predictive and computational toxicology, a highly scientific and research-based field, is rapidly progressing with wider acceptance by regulatory agencies around the world. Almost every aspect of the field has seen fundamental changes during the last decade due to the availability of more data, usage, and acceptance of a variety of predictive tools and an increase in the overall awareness. Also, the influence from the recent explosive developments in the field of artificial intelligence has been significant. However, the need for sophisticated, easy to use and well-maintained software platforms for in silico toxicological assessments remains very high. The MultiCASE suite of software is one such platform that consists of an integrated collection of software programs, tools, and databases. While providing easy-to-use and highly useful tools that are relevant at present, it has always remained at the forefront of research and development by inventing new technologies and discovering new insights in the area of QSAR, artificial intelligence, and machine learning. This chapter gives the background, an overview of the software and databases involved, and a brief description of the usage methodology with the aid of examples.

Reason Vectors: Abstract Representation of Chemistry-Biology Interaction Outcomes, for Reasoning and Prediction.

Many traditional quantitative structure-activity relationship (QSAR) models are based on correlation with high-dimensional, highly variable molecular features in their raw form, limiting their generalizing capabilities despite the use of large training sets. They also lack elements of causality and reasoning. With these issues in mind, we developed a method for learning higher-level abstract representations of the effects of the interactions between molecular features and biology. We named the representations as the reason vectors. They are composed of a series of computed activity of substructures obtained from stepwise reconstruction of the molecule. This representation is very different from fingerprints, which are composed of molecular features directly. These vectors capture reasons of bioactivity of chemicals (or absence thereof) in an abstract form, uncover causality in interactions between chemical features, and generalize beyond specific chemical classes or bioactivity. Reason vectors contain only a few key attributes and are much smaller than molecular fingerprints. They allow vague and conceptual similarity searches, less susceptible to failure on novel combinations of query molecule features and more likely to identify reasons of activity in chemical classes that are absent in training data. Reason vectors can be compared with each other and their activity can be computed by matching with vectors from molecules with known bioactivity. A single molecule produces as many reason vectors as heavy atoms in it, and a simple count of these vectors in a series of activity ranges is all what is needed to predict its bioactivity. Thus, the prediction method is devoid of gradient optimization or statistical fitting.

Transitioning to composite bacterial mutagenicity models in ICH M7 (Q)SAR analyses.

The International Council on Harmonisation (ICH) M7(R1) guideline describes the use of complementary (quantitative) structure-activity relationship ((Q)SAR) models to assess the mutagenic potential of drug impurities in new and generic drugs. Historically, the CASE Ultra and Leadscope software platforms used two different statistical-based models to predict mutations at G-C (guanine-cytosine) and A-T (adenine-thymine) sites, to comprehensively assess bacterial mutagenesis. In the present study, composite bacterial mutagenicity models covering multiple mutation types were developed. These new models contain more than double the number of chemicals (n = 9,254 and n = 13,514) than the corresponding non-composite models and show better toxicophore coverage. Additionally, the use of a single composite bacterial mutagenicity model simplifies impurity analysis in an ICH M7 (Q)SAR workflow by reducing the number of model outputs requiring review. An external validation set of 388 drug impurities representing proprietary pharmaceutical chemical space showed performance statistics ranging from of 66-82% in sensitivity, 91-95% in negative predictivity and 96% in coverage. This effort represents a major enhancement to these (Q)SAR models and their use under ICH M7(R1), leading to improved patient safety through greater predictive accuracy, applicability, and efficiency when assessing the bacterial mutagenic potential of drug impurities.

Computing similarity between structural environments of mutagenicity alerts.

This article describes a method to generate molecular fingerprints from structural environments of mutagenicity alerts and calculate similarity between them. This approach was used to improve classification accuracy of alerts and for searching structurally similar analogues of an alerting chemical. It builds fingerprints using molecular fragments from the vicinity of the alerts and automatically accounts for the activating and deactivating/mitigating features of alerts needed for accurate predictions. This study also demonstrates the usefulness of transfer learning in which a distributed representation of chemical fragments was first trained on millions of unlabelled chemicals and then used for generating fingerprints and similarity search on smaller data sets labelled with Ames test outcomes. The distributed fingerprints gave better prediction performance and increased coverage compared to traditional binary fingerprints. The methodology was applied to four common mutagenic functionalities-primary aromatic amine, aromatic nitro, epoxide and alkyl chloride. Effects of various hyperparameters on prediction accuracy and test coverage for the k-nearest neighbours prediction method are also described, e.g. similarity thresholds, number of neighbours and size of the alert environment.

Improvement of quantitative structure-activity relationship (QSAR) tools for predicting Ames mutagenicity: outcomes of the Ames/QSAR International Challenge Project.

The International Conference on Harmonization (ICH) M7 guideline allows the use of in silico approaches for predicting Ames mutagenicity for the initial assessment of impurities in pharmaceuticals. This is the first international guideline that addresses the use of quantitative structure-activity relationship (QSAR) models in lieu of actual toxicological studies for human health assessment. Therefore, QSAR models for Ames mutagenicity now require higher predictive power for identifying mutagenic chemicals. To increase the predictive power of QSAR models, larger experimental datasets from reliable sources are required. The Division of Genetics and Mutagenesis, National Institute of Health Sciences (DGM/NIHS) of Japan recently established a unique proprietary Ames mutagenicity database containing 12140 new chemicals that have not been previously used for developing QSAR models. The DGM/NIHS provided this Ames database to QSAR vendors to validate and improve their QSAR tools. The Ames/QSAR International Challenge Project was initiated in 2014 with 12 QSAR vendors testing 17 QSAR tools against these compounds in three phases. We now present the final results. All tools were considerably improved by participation in this project. Most tools achieved >50% sensitivity (positive prediction among all Ames positives) and predictive power (accuracy) was as high as 80%, almost equivalent to the inter-laboratory reproducibility of Ames tests. To further increase the predictive power of QSAR tools, accumulation of additional Ames test data is required as well as re-evaluation of some previous Ames test results. Indeed, some Ames-positive or Ames-negative chemicals may have previously been incorrectly classified because of methodological weakness, resulting in false-positive or false-negative predictions by QSAR tools. These incorrect data hamper prediction and are a source of noise in the development of QSAR models. It is thus essential to establish a large benchmark database consisting only of well-validated Ames test results to build more accurate QSAR models.

Descriptor Free QSAR Modeling Using Deep Learning With Long Short-Term Memory Neural Networks.

Current practice of building QSAR models usually involves computing a set of descriptors for the training set compounds, applying a descriptor selection algorithm and finally using a statistical fitting method to build the model. In this study, we explored the prospects of building good quality interpretable QSARs for big and diverse datasets, without using any pre-calculated descriptors. We have used different forms of Long Short-Term Memory (LSTM) neural networks to achieve this, trained directly using either traditional SMILES codes or a new linear molecular notation developed as part of this work. Three endpoints were modeled: Ames mutagenicity, inhibition of P. falciparum Dd2 and inhibition of Hepatitis C Virus, with training sets ranging from 7,866 to 31,919 compounds. To boost the interpretability of the prediction results, attention-based machine learning mechanism, jointly with a bidirectional LSTM was used to detect structural alerts for the mutagenicity data set. Traditional fragment descriptor-based models were used for comparison. As per the results of the external and cross-validation experiments, overall prediction accuracies of the LSTM models were close to the fragment-based models. However, LSTM models were superior in predicting test chemicals that are dissimilar to the training set compounds, a coveted quality of QSAR models in real world applications. In summary, it is possible to build QSAR models using LSTMs without using pre-computed traditional descriptors, and models are far from being "black box." We wish that this study will be helpful in bringing large, descriptor-less QSARs to mainstream use.

Distributed Representation of Chemical Fragments.

This article describes an unsupervised machine learning method for computing distributed vector representation of molecular fragments. These vectors encode fragment features in a continuous high-dimensional space and enable similarity computation between individual fragments, even for small fragments with only two heavy atoms. The method is based on a word embedding algorithm borrowed from natural language processing field, and approximately 6 million unlabeled PubChem chemicals were used for training. The resulting dense fragment vectors are in contrast to the traditional sparse "one-hot" fragment representation and capture rich relational structure in the fragment space. The vectors of small linear fragments were averaged to yield distributed vectors of bigger fragments and molecules, which were used for different tasks, e.g., clustering, ligand recall, and quantitative structure-activity relationship modeling. The distributed vectors were found to be better at clustering ring systems and recall of kinase ligands as compared to standard binary fingerprints. This work demonstrates unsupervised learning of fragment chemistry from large sets of unlabeled chemical structures and subsequent application to supervised training on relatively small data sets of labeled chemicals.

Critical evaluation of human oral bioavailability for pharmaceutical drugs by using various cheminformatics approaches.

PURPOSE: Oral bioavailability (%F) is a key factor that determines the fate of a new drug in clinical trials. Traditionally, %F is measured using costly and time-consuming experimental tests. Developing computational models to evaluate the %F of new drugs before they are synthesized would be beneficial in the drug discovery process. METHODS: We employed Combinatorial Quantitative Structure-Activity Relationship approach to develop several computational %F models. We compiled a %F dataset of 995 drugs from public sources. After generating chemical descriptors for each compound, we used random forest, support vector machine, k nearest neighbor, and CASE Ultra to develop the relevant QSAR models. The resulting models were validated using five-fold cross-validation. RESULTS: The external predictivity of %F values was poor (R(2) = 0.28, n = 995, MAE = 24), but was improved (R(2) = 0.40, n = 362, MAE = 21) by filtering unreliable predictions that had a high probability of interacting with MDR1 and MRP2 transporters. Furthermore, classifying the compounds according to the %F values (%F < 50% as "low", %F ≥ 50% as 'high") and developing category QSAR models resulted in an external accuracy of 76%. CONCLUSIONS: In this study, we developed predictive %F QSAR models that could be used to evaluate new drug compounds, and integrating drug-transporter interactions data greatly benefits the resulting models.

Effectiveness of CASE Ultra Expert System in Evaluating Adverse Effects of Drugs.

Purpose of this pilot study is to test the QSAR expert system CASE Ultra for adverse effect prediction of drugs. 870 drugs from the SIDER adverse effect dataset were tested using CASE Ultra for carcinogenicity, genetic, liver, cardiac, renal and reproductive toxicity. 47 drugs that were withdrawn from market since the 1950s were also evaluated for potential risks using CASE Ultra and compared them with the actual reasons for which the drugs were recalled. For the whole SIDER test set (n=870), sensitivity and specificity of the carcinogenicity predictions are 66.67 % and 82.17 % respectively; for liver toxicity: 78.95 %, 78.50 %; cardiotoxicity: 69.07 %, 57.57 %; renal toxicity: 46.88 %, 67.90 %; and reproductive toxicity: 100.00 %, 61.10 %. For the SIDER test chemicals not present in the training sets of the models, sensitivity and specificity of carcinogenicity predictions are 100.00 % and 88.89 % respectively (n=404); for liver toxicity: 100.00 %, 51.33 % (n=115); cardiotoxicity: 100.00 %, 20.45 % (n=94); renal toxicity: 100.00 %, 45.54 % (n=115); and reproductive toxicity: 100.00 %, 48.57 % (n=246). CASE Ultra correctly recognized the relevant toxic effects in 43 out of the 47 withdrawn drugs. It predicted all 9 drugs that were not part of the training set of the models, as unsafe.

Optimizing predictive performance of CASE Ultra expert system models using the applicability domains of individual toxicity alerts.

Fragment based expert system models of toxicological end points are primarily comprised of a set of substructures that are statistically related to the toxic property in question. These special substructures are often referred to as toxicity alerts, toxicophores, or biophores. They are the main building blocks/classifying units of the model, and it is important to define the chemical structural space within which the alerts are expected to produce reliable predictions. Furthermore, defining an appropriate applicability domain is required as part of the OECD guidelines for the validation of quantitative structure-activity relationships (QSARs). In this respect, this paper describes a method to construct applicability domains for individual toxicity alerts that are part of the CASE Ultra expert system models. Defining applicability domain for individual alerts was necessary because each CASE Ultra model is comprised of multiple alerts, and different alerts of a model usually represent different toxicity mechanisms and cover different structural space; the use of an applicability domain for the overall model is often not adequate. The domain for each alert was constructed using a set of fragments that were found to be statistically related to the end point in question as opposed to using overall structural similarity or physicochemical properties. Use of the applicability domains in reducing false positive predictions is demonstrated. It is now possible to obtain ROC (receiver operating characteristic) profiles of CASE Ultra models by applying domain adherence cutoffs on the alerts identified in test chemicals. This helps in optimizing the performance of a model based on their true positive-false positive prediction trade-offs and reduce drastic effects on the predictive performance caused by the active/inactive ratio of the model's training set. None of the major currently available commercial expert systems for toxicity prediction offer the possibility to explore a model's full range of sensitivity-specificity spectrum, and therefore, the methodology developed in this study can be of benefit in improving the predictive ability of the alert based expert systems.

A structural analysis of the differential cytotoxicity of chemicals in the NCI-60 cancer cell lines.

The primary functions of cancer chemotherapeutic agents are not only to inhibit the growth or kill the cancer cells, but to do so without eliciting unreasonable cytotoxic effects on the healthy cells and to withstand the ability of the cancer cells to develop resistance against it. This has unfortunately been proven so far to be a very difficult objective. In this perspective, the ability of small molecules (anti-tumor agents) to 'see' different cell types differently can be a key attribute. Thus the term 'differential cytotoxicity' is normally used to describe the drug's specificity. In the present paper, we have quantified differential cytotoxicity from a study of the chemicals tested in the National Cancer Institute's Developmental Therapeutics Program. The MULTICASE (Multiple Computer Automated Structure Evaluation) methodology was used to discover statistically significant structural fragments (biophores) related to the differential cytotoxicity of the compounds. We found that even small structural features often become important in this regard which is evident from the biophores that were found in structurally diverse chemicals. By utilizing the difference between the raw and normalized differential cytotoxicity indices, we found that the alpha,beta-unsaturated carbonyl group (O=C-C=CH(2)) is the major biophore associated with compounds with essentially parallel concentration profiles in the cell lines in question. These compounds have high non-normalized differential cytotoxicity but considerably low normalized differential cytotoxocity. The models developed were cross validated for their predictive ability.

ESP: a method to predict toxicity and pharmacological properties of chemicals using multiple MCASE databases.

We describe here the development of a computer program which uses a new method called Expert System Prediction (ESP), to predict toxic end points and pharmacological properties of chemicals based on multiple modules created by the MCASE artificial intelligence system. The modules are generally based on different biological models measuring related end points. The purpose is to improve the decision making process regarding the overall activity or inactivity of the chemicals and also to enable rapid in silico screening. ESP evaluates the significance of the biophores from a different viewpoint and uses this information for predicting the activity of new chemicals. We have used a unique encoding system to represent relevant features of a chemical in the form of a pattern vector and a genetic artificial neural network (GA-ANN) to gain knowledge of the relationship between these patterns and the overall pharmacological property. The effectiveness of ESP is illustrated in the prediction of general carcinogenicity of a diverse set of chemicals using MCASE male/female rats and mice carcinogenicity modules.

Structure-activity relationship study of a diverse set of estrogen receptor ligands (I) using MultiCASE expert system.

The MultiCASE expert system was used to construct a quantitative structure-activity relationship model to screen chemicals with estrogen receptor (ER) binding potential. Structures and ER binding data of 313 chemicals were used as inputs to train the expert system. The training data set covers inactive, weak as well as very powerful ER binders and represents a variety of chemical compounds. Substructural features associated with ER binding activity (biophores) and features that prevent receptor binding (biophobes) were identified. Although a single phenolic hydroxyl group was found to be the most important biophore responsible for the estrogenic activity of most of the chemicals, MultiCASE also identified other biophores and structural features that modulate the activity of the chemicals. Furthermore, the findings supported our previous hypothesis that a 6 A distant descriptor may describe a ligand-binding site on an ER. Quantitative structure-activity relationship models for the chemicals associated with each biophore were constructed as part of the expert system and can be used to predict the activity of new chemicals. The model was cross validated via 10 x 10%-off tests, giving an average concordance of 84.04%.

Screening of high production volume chemicals for estrogen receptor binding activity (II) by the MultiCASE expert system.

A structurally and functionally diverse and cross-validated quantitative structure-activity knowledge database generated by the MultiCASE expert system was used to screen 2526 high production volume chemicals (HPVCs) for their estrogen receptor binding activity. 73 HPVCs were found to contain structural features or biophores that have been documented as having the ability to bind to the estrogen receptor. Potential chemicals were ranked according to their quantitatively predicted ER binding potential and the details of the biophores found in them are discussed.