Machine Learning and Deep Learning in Computational Toxicology

Preface
Contents
Editor and Contributors
1 Machine Learning and Deep Learning Promote Computational Toxicology for Risk Assessment of Chemicals
	1.1 Risk Assessment of Chemicals
	1.2 Computational Toxicology
	1.3 Machine Learning in Computational Toxicology
	1.4 Deep Learning in Toxicology
	1.5 Perspectives
	References
Part I Machine Learning and Deep Learning Methods for Computational Toxicology
2 Assessment of the Xenobiotics Toxicity Taking into Account Their Metabolism
	2.1 Introduction
	2.2 Computational Methods of Studying Metabolism
		2.2.1 Databases Containing Xenobiotic Metabolism Information
		2.2.2 Descriptors/Notation Used for Metabolism Prediction
		2.2.3 Prediction of Biotransformation Sites
		2.2.4 Generation of the Structures of Probable Metabolites
		2.2.5 Reactive Metabolite Formation Prediction
	2.3 Integral Computational Assessment of Xenobiotic Toxicity
	2.4 Future Directions in Xenobiotic Toxicity Assessment
	References
3 Emerging Machine Learning Techniques in Predicting Adverse Drug Reactions
	3.1 Introduction
	3.2 Feature Generation for Machine Learning
		3.2.1 Structure-Based Features
		3.2.2 Interactions and Associations
		3.2.3 Data Sources for Feature Generation
	3.3 Conventional Methods for ADR Prediction
	3.4 Emerging Methods for ADR Prediction
		3.4.1 Molecule-Based Methods
		3.4.2 Similarity-Based Methods
		3.4.3 Network- and Graph-Based Methods
	3.5 ADR Prediction Future Directions
	References
4 Drug Effect Deep Learner Based on Graphical Convolutional Network
	4.1 Introduction
	4.2 Results
		4.2.1 Gene Vector: Generation and Evaluation
		4.2.2 Molecular Feature and Vector Generation
		4.2.3 Cell Vector: Generation and Evaluation
		4.2.4 Deep Drug Effect Predictor: Training and Validation
		4.2.5 Application of DDEP to Predict the Effects of Anti-cancer Drugs Against Breast Adenocarcinoma
		4.2.6 Insights into Drug Classification
	4.3 Discussion
	4.4 Methods
		4.4.1 Capture Contextual Information of Genes from Their Interaction Networks
		4.4.2 Generating Gene Vectors and Cell Vectors
		4.4.3 GCN-Based Pre-models
		4.4.4 Deep Drug Effect Predictor
	References
5 AOP-Based Machine Learning for Toxicity Prediction
	5.1 Introduction
	5.2 Research Status and Existing Problems for ML
	5.3 General Overview of AOP
		5.3.1 The Generation of AOP
		5.3.2 The Framework of AOP
		5.3.3 Qualitative AOP and Quantitative AOP
	5.4 Research Progress of Toxicity Prediction by AOP and ML
	5.5 Perspectives and Future Prospects of AOP
	References
6 Graph Kernel Learning for Predictive Toxicity Models
	6.1 Introduction
	6.2 A Brief Introduction of Graph Concepts
		6.2.1 Graph Theory Definitions
		6.2.2 Graph Kernels Fundamentals
	6.3 Graph Kernel Learning for Molecular Representations
	6.4 Applications of GKL Methods on Chemical Toxicity
		6.4.1 Benchmark Data Sets and Methods About Chemical Toxicity
		6.4.2 Applications of Graph Kernel-Based Methods
		6.4.3 Applications of Graph Neural Networks
		6.4.4 Applications of Learnable Graph Embeddings
		6.4.5 Applications of Learnable Graph Kernels
	6.5 Challenges and Perspectives of Graph Kernel Learning on Toxicity-Related Problems
	6.6 Conclusion
	References
7 Optimize and Strengthen Machine Learning Models Based on in Vitro Assays with Mechanistic Knowledge and Real-World Data
	7.1 Introduction
	7.2 Incorporating AOPs to Construct Parsimonious Machine Learning Models
		7.2.1 AOPs and AOP Networks
		7.2.2 Using AOPs to Facilitate Building Parsimonious Machine Learning Models
	7.3 Utilize Spontaneous Reporting Databases to Corroborate Findings of Machine Learning Models
		7.3.1 Statistical Methods for Safety Signal Mining Using Spontaneous Reporting Databases
		7.3.2 Obtain Data from FAERS
		7.3.3 Poisson Regression Model for Report Counts
		7.3.4 Incorporating Host Factors in Testing
		7.3.5 Utilize FAERS Data to Corroborate Models Based on in Vitro Assays
	7.4 Conclusions
	References
8 Multitask Learning for Quantitative Structure–Activity Relationships: A Tutorial
	8.1 Introduction
	8.2 QSAR and Multitask Learning
		8.2.1 Definition of MTL Problem
		8.2.2 Task Relatedness
		8.2.3 Multitask Neural Networks
		8.2.4 Performance Evaluation
	8.3 Case Study: NURA Dataset
	8.4 Hands-On Tutorial
		8.4.1 Getting Started
	8.5 Conclusions
	References
Part II Tools and Approaches Facilitating Machine Learning and Deep Learning Methods in Computational Toxicology
9 Isalos Predictive Analytics Platform: Cheminformatics, Nanoinformatics, and Data Mining Applications
	9.1 Introduction
	9.2 Isalos Platform
	9.3 Data Input
	9.4 Data Transformation
		9.4.1 Normalizers
		9.4.2 Data Manipulation
		9.4.3 Dataset Splitting
	9.5 Analytics
		9.5.1 Modelling Methodologies
		9.5.2 Feature Selection
		9.5.3 Existing Model Utilization
	9.6 Statistics
		9.6.1 Domain—APD
		9.6.2 Model Metrics
	9.7 Development of Predictive Models with Isalos
		9.7.1 Ecotox Models
		9.7.2 Molecular, Size, and Surface-Based Safe by Design (MS3bD, MSzeta) Model
		9.7.3 Cell Viability Model
	9.8 Conclusions
	References
10 ED Profiler: Machine Learning Tool for Screening Potential Endocrine-Disrupting Chemicals
	10.1 Introduction
	10.2 Materials and Methods
		10.2.1 Data Sets
		10.2.2 Molecular Descriptor Calculation
		10.2.3 (Q)SAR Modeling
		10.2.4 Applicability Domain and Reliability Evaluation
		10.2.5 Software Development
	10.3 Development of Predictive Models
		10.3.1 Proposed Predictive Model System
		10.3.2 Development of SAR Models
		10.3.3 Development of QSAR Models
	10.4 Development of Software
		10.4.1 Features and Overview of the Software
		10.4.2 Examples
	10.5 Conclusions
	References
11 Quantitative Target-specific Toxicity Prediction Modeling (QTTPM): Coupling Machine Learning with Dynamic Protein–Ligand Interaction Descriptors (DyPLIDs) to Predict Androgen Receptor-mediated Toxicity
	11.1 Introduction
	11.2 Materials and Methods
		11.2.1 Study Design
		11.2.2 Dataset Curation, Preprocessing, and Chemical Preparation
		11.2.3 Molecular Docking
		11.2.4 Molecular Dynamics (MD) Simulations
		11.2.5 Dynamic Protein–Ligand Interaction Descriptors (DyPLIDs) Calculation
		11.2.6 Feature Selection: Down-Selection of Descriptors
		11.2.7 Dataset Splitting
		11.2.8 Machine Learning for Quantitative AR Activity Prediction Modeling
	11.3 Results and Discussion
		11.3.1 Conformational Ensemble of AR-Ligand Interactions
		11.3.2 Comparison of 6 ns Versus 100 ns Simulations
		11.3.3 Fingerprint Chemical Diversity
		11.3.4 Predictive QSAR Model
		11.3.5 Feature Importance
		11.3.6 Model Limitation
	11.4 Conclusion
	References
12 Mold2 Descriptors Facilitate Development of Machine Learning and Deep Learning Models for Predicting Toxicity of Chemicals
	12.1 Introduction
	12.2 Mold2 Descriptors
		12.2.1 The Descriptors
		12.2.2 The Software
	12.3 Information Content of Mold2 Descriptors
	12.4 Applications in Machine Learning
		12.4.1 Predicting Estrogenic Activity
		12.4.2 Predicting Androgenic Activity
		12.4.3 Predicting Kinase Inhibitors
	12.5 Applications in Deep Learning
		12.5.1 Predicting DILI
		12.5.2 Predicting Drug-Likeness
	12.6 Summary
	References
13 Applicability Domain Characterization for Machine Learning QSAR Models
	13.1 An Outline of Quantitative Structure–Activity Relationship (QSAR) Models
		13.1.1 Core Elements of QSAR Models
		13.1.2 Validity of QSAR Models
	13.2 Concepts and Understandings of AD
		13.2.1 Physicochemical, Structural, Mechanistic, and Metabolic Aspects
		13.2.2 Interpolation, Distance/Similarity, and Boundary
		13.2.3 AD Metrics Evaluating Prediction Performance on Individual Chemicals
	13.3 AD Characterization Methods
		13.3.1 Descriptor Domain
		13.3.2 Structural (Similarity) Domain
		13.3.3 Clustering-Based Methods
		13.3.4 First-Class AD Metrics
		13.3.5 Second-Class AD Metrics
		13.3.6 Visualization of AD
	13.4 Impacts on the QSAR Modeling Scenario from Machine Learning Algorithms
	13.5 Toward Broader AD for Machine Learning QSAR Models
	References
14 Controlling for Confounding in Complex Survey Machine Learning Models to Assess Drug Safety and Risk
	14.1 Introduction and Background
	14.2 The Propensity Score Method
	14.3 Software and Assessment of Balance
	14.4 A Drug Safety Model for Prescription NSAIDs
	14.5 Propensity Score Weighting
		14.5.1 Adjusting for Confounding with Propensity Score ATE Weighting
		14.5.2 Adjusting for Confounding with Propensity Score ATT Weighting
	14.6 Propensity Score Stratification
		14.6.1 Adjusting for Confounding with Propensity Score Stratification
	14.7 Conclusion
	References
15 Multivariate Curve Resolution for Analysis of Heterogeneous System in Toxicogenomics
	15.1 Introduction
	15.2 Basic Conceptions and Application Scenarios
		15.2.1 The Definition of MCR in Heterogeneous Systems
		15.2.2 The Ultimate Goal of Applying MCR in TGx
	15.3 Method Categories
		15.3.1 Determining or Estimating k
		15.3.2 Determining or Estimating E and W
		15.3.3 Jointly Utilization and Alternate Estimation
	15.4 Available Resources
		15.4.1 Databases of TGx Data
		15.4.2 Functions of MCR
		15.4.3 Tools of Deconvolution by Using MCR
	15.5 Perspective and Future Directions
	References
Part III Machine Learning and Deep Learning for Chemical Toxicity Prediction
16 The Use of Machine Learning to Support Drug Safety Prediction
	16.1 Introduction
	16.2 Chemical-Based Safety Machine Learning
		16.2.1 Overview
		16.2.2 Databases
		16.2.3 Machine Learning Algorithms
	16.3 Case Study—Assessment of Pharmaceutical Impurities
	16.4 Conclusions
	References
17 Machine Learning-Based QSAR Models and Structural Alerts for Prediction of Mitochondrial Dysfunction
	17.1 Introduction
	17.2 Datasets and Methods
		17.2.1 Data on Mitochondrial Dysfunction
		17.2.2 Machine Learning Methods Used for Model Construction
		17.2.3 Model Evaluation
		17.2.4 Methods to Identify Structural Alerts
	17.3 Mitochondrial Dysfunction QSAR Models and Structural Alerts
		17.3.1 Mitochondrial Dysfunction QSAR Models
		17.3.2 Structural Alerts for Mitochondrial Dysfunction
	17.4 Conclusions and Future Directions
	References
18 Machine Learning and Deep Learning Applications to Evaluate Mutagenicity
	18.1 In Silico Methods to Predict Bacterial Mutagenicity
	18.2 Data for Modeling Mutagenicity
	18.3 Traditional Machine Learning for Mutagenicity Prediction
	18.4 Deep Learning for Mutagenicity Prediction
	18.5 Discussion and Perspective
	References
19 Modeling Tox21 Data for Toxicity Prediction and Mechanism Deconvolution
	19.1 Introduction
	19.2 Tox21 10K Compound Library and Assay Data
		19.2.1 Tox21 Compound Collection
		19.2.2 Tox21 qHTS Process
	19.3 Modeling Tox21 Data for Toxicity Prediction
		19.3.1 Multiple Species In Vivo Toxicity
		19.3.2 Human In Vivo Toxicity
		19.3.3 In Vitro Toxicity
	19.4 Toxicity Pathways and Mechanisms
	19.5 Conclusions and Moving Forward
	References
20 Identification of Structural Alerts by Machine Learning and Their Applications in Toxicology
	20.1 Introduction
	20.2 Approaches for Identification of Structural Alerts
		20.2.1 Expert Systems
		20.2.2 Computational Approaches
		20.2.3 Comparison of Data-Driven Structural Alerts with Expert Systems
	20.3 Application of Structural Alerts in Toxicology
		20.3.1 Toxicity Prediction
		20.3.2 Explanation of QSAR Models
		20.3.3 Molecular Optimization
		20.3.4 Exploring New Mechanisms
	20.4 Perspectives and Outlook
	References
21 Machine Learning in Prediction of Nanotoxicology
	21.1 Introduction
	21.2 Toxicity of Nanomaterials
		21.2.1 Toxicity of Carbon Nanomaterials
		21.2.2 Toxicity of Transition Metal Dichalcogenides
		21.2.3 Toxicity of MOFs
	21.3 Prediction of Nanotoxicity by Machine Learning
		21.3.1 Prediction of Carbon Nanomaterials Toxicity by Machine Learning
		21.3.2 Prediction of Nanometal Toxicity by Machine Learning
		21.3.3 Prediction of Nanometal Oxide Toxicity by Machine Learning
		21.3.4 Prediction of Other Nanomaterials Toxicity by Machine Learning
	21.4 Future Directions of Machine Learning in Nanotoxicology Prediction
	References
22 Machine Learning for Predicting Organ Toxicity
	22.1 Introduction
	22.2 Machine Learning Algorithms
		22.2.1 Classification and Regression Tree
		22.2.2 k-Nearest Neighbors (kNN)
		22.2.3 Naïve Bayes (NB)
		22.2.4 Random Forest
		22.2.5 Support Vector Machine
	22.3 Organ Toxicity Prediction
		22.3.1 Liver Toxicity
		22.3.2 Kidney Toxicity
		22.3.3 Heart Toxicity
	22.4 A Case Study for Organ Toxicity Prediction
		22.4.1 Data Sources
		22.4.2 Supervised Machine Learning
		22.4.3 Results
	22.5 Summary
	References
Part IV The Progress of Machine Learning and Deep Learning in New Areas
23 Computational Modeling for the Prediction of Hepatotoxicity Caused by Drugs and Chemicals
	23.1 Introduction
	23.2 Machine Learning Methods for Predicting Hepatotoxicity
		23.2.1 Toxicity Dataset for Machine Learning
		23.2.2 Metrics for Evaluating Model Performance
		23.2.3 Machine Learning Algorithms
	23.3 A Case Study: Machine Learning Modeling for Hepatotoxicity Prediction
		23.3.1 Data Sources
		23.3.2 Modeling by Machine Learning Approaches
		23.3.3 Results
	23.4 Summary and Future Direction
	References
24 Artificial Intelligence for Risk Assessment of Cancer Therapy-Related Cardiotoxicity and Precision Cardio-Oncology
	24.1 Introduction
	24.2 Methods and Materials
		24.2.1 Data Resources
		24.2.2 Molecular Feature and Vector Generation
		24.2.3 Defining Biological Endpoints and Clinical Outcomes
		24.2.4 AI/ML Algorithm and Model Selection
		24.2.5 Evaluating Model Performance
	24.3 Variable Network Construction
	24.4 Case Studies
		24.4.1 In Silico Pharmacoepidemiologic Evaluation of Drug-Induced Cardiovascular Complications Using Combined Classifiers
		24.4.2 Machine Learning-Based Risk Assessment for Cancer Therapy-Related Cardiac Dysfunction in 4300 Longitudinal Oncology Patients
		24.4.3 Cardiac Risk Stratification in Cancer Patients: A Longitudinal Patient-Patient Network Analysis
	24.5 Future Directions and Conclusion
	References
25 Deep Learning Model for Prediction of Compound Activities Over a Panel of Major Toxicity-Related Proteins
	25.1 Introduction
	25.2 Methods
		25.2.1 Dataset
		25.2.2 Chemical Diversity Analysis
		25.2.3 Prediction Models
		25.2.4 Evaluation Metrics
	25.3 Results and Discussion
		25.3.1 Data Collection and Analysis
		25.3.2 Drug and Target Representations Selection
		25.3.3 Model Performance
		25.3.4 Comparison with Conventional per Protein Models
		25.3.5 External Validation
	25.4 Conclusions
	References
26 Machine Learning for Analyzing Drug Safety in Electronic Health Records
	26.1 Introduction
	26.2 Drug Safety Problems to Solve with ML
		26.2.1 Prescription Error
		26.2.2 Medication Misuse
		26.2.3 Drug-Drug Interactions
	26.3 Recent Trends of NLP and ML Methods in Pharmacovigilance
		26.3.1 The Existing of NLP Approaches
		26.3.2 Machine Learning Methods
	26.4 Discussions
	References
27 Powering Toxicogenomic Studies by Applying Machine Learning to Genomic Sequencing and Variant Detection
	27.1 Introduction
	27.2 Machine Learning in Genomic Variant Detections
		27.2.1 Machine Learning Algorithms in Germline Variant Detection
		27.2.2 Challenges in Somatic Mutation Calling
		27.2.3 Machine Learning to Improve Accuracy of Somatic Mutation Detection
	27.3 Training Data for Machine Learning-Based Variant Callers
	27.4 Conclusion
	References
28 Machine Learning for Predicting Gas Adsorption Capacities of Metal Organic Framework
	28.1 Introduction
	28.2 Data Sources
	28.3 Descriptors of MOFs
	28.4 ML Algorithms
	28.5 ML Models for Predicting Gas Adsorption of MOFs
		28.5.1 ML Models for CH4 Adsorption
		28.5.2 ML Models for H2 Adsorption
		28.5.3 ML Models for CO2 Adsorption
		28.5.4 ML Models for Xe/Kr Selective Adsorption
	28.6 Conclusion Remarks and Future Perspective
	References