Artificial intelligence in drug discovery

Title
Copyright
Contents
Section 1: Introduction to Artificial Intelligence and Chemistry
	Chapter 1 Introduction
		1.1 Introduction
Section 2: Chemical Data
	Chapter 2 The History of Artificial Intelligence and Chemistry
		2.1 Artificial Intelligence in History
		2.2 The Winters of Artificial Intelligence
		2.3 Chemistry Finding Artificial Intelligence
		2.4 Synthesis Planning
		2.5 Predictive Modelling of Properties
		2.6 Summary
		References
	Chapter 3 Chemical Topic Modeling – An Unsupervised Approach Originating from Text-mining to Organize Chemical Data
		3.1 Introduction
		3.2 Topic Modeling and LDA
			3.2.1 The Mathematical Framework of LDA
			3.2.2 Advanced Topic Modeling Extensions
			3.2.3 Topic Modeling and Its Relation to Other Machine Learning Methods
			3.2.4 Topic Modeling in Different Scientific Disciplines
		3.3 Chemical Topic Modeling
			3.3.1 Feature Representation for Chemical Topic Modeling
			3.3.2 Creating and Interpreting a Chemical Topic Model
			3.3.3 Evaluation of a Chemical Topic Model
		3.4 Exploring Large Data Sets with Chemical Topic Modeling
			3.4.1 Hierarchical Topics
		3.5 Combining Text and Chemical Information
		3.6 Conclusions, Limitations and Future Work
		References
	Chapter 4 Deep Learning and Chemical Data
		4.1 Introduction
		4.2 Background
			4.2.1 Deep Learning
			4.2.2 Evaluation Methods
			4.2.3 Natural-language Processing
		4.3 Case Study 1: Spectroscopic Analysis
			4.3.1 Background
			4.3.2 Worked NMR Example
		4.4 Case Study 2: Natural Language Processing Experiments
			4.4.1 Introduction
			4.4.2 Chemical Entity Mentions in Patents
			4.4.3 Deep Learning vs. Feature Engineering for Relationship Extraction
		4.5 Conclusions and Future Work
		References
Section 3: Ligand-based Predictive Modelling
	Chapter 5 Concepts and Applications of Conformal Prediction in Computational Drug Discovery
		5.1 Introduction
		5.2 Conformal Prediction Modalities Commonly Used in Computer-aided Drug Design
			5.2.1 Inductive Conformal Prediction (ICP)
		5.3 Handling Imbalanced Datasets: Mondrian Conformal Prediction (MCP)
			5.3.1 ICP for Regression
			5.3.2 Conformal Prediction Using All Labelled Data for Learning
		5.4 Conformal Prediction Methods for Deep Learning
		5.5 Open-source Implementations of Conformal Prediction
		5.6 Current Limitations of Conformal Prediction and Future Perspectives
		Conflicts of Interest
		References
	Chapter 6 Non-applicability Domain. The Benefits of Defining “I Don't Know” in Artificial Intelligence
		6.1 Introduction
		6.2 Predictive Models
		6.3 Defining NotAvailable Predictions
		6.4 All Leave One Out Models
		6.5 Benefits of Defining NotAvailable Predictions
		6.6 Simulation Study
			6.6.1 Design of the Experiment
			6.6.2 Results of the Experiment
			6.6.3 Discussion
		6.7 Questions and Criticism
			6.7.1 Question 1
			6.7.2 Question 2
			6.7.3 Question 3
			6.7.4 Question 4
			6.7.5 Question 5
			6.7.6 Question 6
			6.7.7 Question 7
			6.7.8 Question 8
		6.8 Final Remarks
		Abbreviations
		References
Section 4: Structure-based Predictive Modelling
	Chapter 7 Predicting Protein-ligand Binding Affinities
		7.1 Introduction
		7.2 A Brief Background on Classical Methodologies
			7.2.1 Potential-based
			7.2.2 Simulation-based
			7.2.3 Data-based
		7.3 Modern Machine-learning Scoring Functions
			7.3.1 Domain Applicability
			7.3.2 Descriptors
			7.3.3 Models
			7.3.4 Interpretability
			7.3.5 Implementation and Availability
		7.4 Available Data and Evaluation
			7.4.1 Scope and Databases
			7.4.2 Evaluation
		7.5 Discussion
		References
	Chapter 8 Virtual Screening with Convolutional Neural Networks
		8.1 Introduction
			8.1.1 Virtual Screening
			8.1.2 Traditional Approaches to Virtual Screening
			8.1.3 Machine Learning Scoring Functions
			8.1.4 Rationale for Deep Learning Approaches
		8.2 Virtual Screening
			8.2.1 Data Sets for Structure-based Virtual Screening
			8.2.2 Appropriate Train/Test Splits for SBVS
			8.2.3 Evaluation Measures
		8.3 Convolutional Neural Networks
			8.3.1 CNNs: A Primer
			8.3.2 ImageNet
			8.3.3 Modern CNN Architectures
		8.4 CNN Applications for Virtual Screening
			8.4.1 Input Format for CNN Structure-based Virtual Screening
			8.4.2 Outline and Performance of CNN-based Methods
		8.5 Other Closely Related Tasks
			8.5.1 Pose Prediction
			8.5.2 Binding Affinity Prediction
		8.6 Visualisation
		8.7 Outlook
		References
	Chapter 9 Machine Learning in the Area of Molecular Dynamics Simulations
		9.1 Introduction
			9.1.1 Basics of Molecular Dynamics
			9.1.2 Machine-learning Applications
			9.1.3 MD and ML
		9.2 Using Machine Learning to Improve Force Fields
			9.2.1 Multi-variate Linear Regression
			9.2.2 Bayesian Inference
			9.2.3 Genetic Algorithm
			9.2.4 Random Forest Regression
			9.2.5 Artificial Neural Network
			9.2.6 Remarks
		9.3 Improving Sampling in MD Simulations
			9.3.1 General Sampling Enhancement
			9.3.2 Estimating the Biasing Potential for a Given Reaction Coordinate
			9.3.3 Estimating Optimal Collective Variables
		9.4 Learning from MD Trajectories
			9.4.1 Application to Clustering
			9.4.2 Application to Property Prediction
			9.4.3 Application to Kinetic Models
		9.5 Perspectives and Challenges
			9.5.1 Datasets on Dynamics Information
			9.5.2 Benchmarking
			9.5.3 Open-source Implementation
			9.5.4 Concluding Remarks
		References
Section 5: Molecular Design
	Chapter 10 Compound Design Using Generative Neural Networks
		10.1 Introduction
		10.2 Principles of Deep Learning
		10.3 De Novo Design via Deep Learning
			10.3.1 Molecular Representation
			10.3.2 Recurrent Neural Networks
			10.3.3 Autoencoder Variants
			10.3.4 Graph-based Neural Networks
		10.4 Property Prediction through Deep Learning
		10.5 Conclusions and Outlook
		References
	Chapter 11 Junction Tree Variational Autoencoder for Molecular Graph Generation
		11.1 Introduction
		11.2 Neural Generation of Molecular Graphs
			11.2.1 Junction Tree
			11.2.2 Tree and Graph Encoder
			11.2.3 Junction Tree Decoder
			11.2.4 Graph Decoder
		11.3 Application to Molecular Design
			11.3.1 Molecular Generative Model
			11.3.2 Molecule-to-Molecule Translation
		11.4 Experiments
			11.4.1 Molecular Variational Autoencoder
			11.4.2 Molecular Translation
		11.5 Conclusion
		References
	Chapter 12 AI via Matched Molecular Pair Analysis
		12.1 Introduction
		12.2 Essential Features of Artificial Intelligence
		12.3 Matched Molecular Pair Analysis
			12.3.1 Generic Issues in Identifying Matched Molecular Pairs
			12.3.2 Automation
			12.3.3 Other Matched Pair Technologies
			12.3.4 Fuzzy Matched Pairs
			12.3.5 Matched Molecular Series
			12.3.6 MMPA Enhanced by Protein Structural Data
		12.4 Future Developments
		12.5 Summary
		References
	Chapter 13 Molecular De Novo Design Through Deep Generative Models
		13.1 Introduction
		13.2 Sequence-based Methods for De Novo Generation of Small Molecules
			13.2.1 Embeddings and Tokenization
			13.2.2 Recurrent Neural Networks
			13.2.3 Sampling SMILES from RNNs
			13.2.4 Properties and Synthesizability
			13.2.5 Advanced Neural Architectures
		13.3 Graph-based De Novo Structure Generation
		13.4 Benchmarking Generative Molecular De Novo Design Models
			13.4.1 Benchmarking Explorative Models
			13.4.2 Benchmarking Exploitative Models
			13.4.3 Benchmarking Models During Training
			13.4.4 Comparing Model Architectures
		13.5 Conclusions
		References
	Chapter 14 Active Learning for Drug Discovery and Automated Data Curation
		14.1 Introduction
		14.2 Active Learning for Drug Discovery, Chemistry, and Material Science
			14.2.1 Exploitation vs. Exploration
			14.2.2 Balancing Different Objectives
			14.2.3 When to Stop – Say When!
			14.2.4 Batch Selection
			14.2.5 Benchmarking the Learning
		14.3 Active Learning for Data Curation
			14.3.1 Reduced Redundancy and Balanced Data
			14.3.2 Reactive Learning
		14.4 Conclusions
		References
	Section 6: Synthesis Planning
	Chapter 15 Data-driven Prediction of Organic Reaction Outcomes
		15.1 Introduction
			15.1.1 The Role of Reaction Prediction
			15.1.2 Non-data Driven Heuristic Systems
		15.2 Data-driven Approaches
			15.2.1 Focused Analyses of Specific Reaction Classes
			15.2.2 At the Mechanistic Level
			15.2.3 Via Reaction Templates
			15.2.4 Without Reaction Templates: Graphs
			15.2.5 Without Reaction Templates: Sequences
		15.3 Conclusion
			15.3.1 Data Availability
			15.3.2 Evaluation
			15.3.3 Breadth versus Accuracy
		15.4 Model Types
		15.5 Conclusion
		References
Section 7: Future Outlook
	Chapter 16 ChemOS: An Orchestration Software to Democratize Autonomous Discovery
		16.1 Introduction
		16.2 Automated Approaches to Scientific Discovery
			16.2.1 Algorithmic Strategies to Screen the Parameter Space
			16.2.2 Examples of Automation in Key Industrial Sectors and Academia
			16.2.3 Limitations of Automated Approaches
		16.3 Autonomous Approaches to Scientific Discovery
			16.3.1 Algorithmic Strategies to Experiment Planning
			16.3.2 Roadmap for Deploying and Orchestrating the Self-driving Laboratories
			16.3.3 Early Realization of Self-driving Laboratories
		16.4 ChemOS to Orchestrate Next-generation Experimentation
			16.4.1 AI-aided Experiment Planning
			16.4.2 Intuitive Human–Robot Interactions
			16.4.3 Online Analysis of Experimental Results
			16.4.4 Databases Management and Storage Solutions
			16.4.5 Connection to Automated Solutions
		16.5 Successful Applications of ChemOS to Scientific Challenges
			16.5.1 Calibration of an Automated Setup for Real-time Reaction Monitoring
			16.5.2 Discovery of Conductive Thin-film Materials
			16.5.3 Formulation of Polymer Blends for Photo-stable Solar Cells
		16.6 Application of ChemOS to Drug Discovery
		16.7 Conclusion and Outlook
		References
Section 8: Summary and Outlook
	Chapter 17 Summary and Outlook
		17.1 Introduction
		17.2 Challenges
			17.2.1 Data
			17.2.2 Compute
			17.2.3 Culture
		17.3 Summary
Subject Index