Chapter 1: The Drug Discovery Process
The chapter introduction explains why drug discovery is difficult and why computational screening and machine learning are useful.
After this chapter you will be able to
- Drug discovery difficulty. Explain why developing a therapeutic is costly, slow, and prone to failure, connecting those outcomes to the vast number of possible drug-like molecules, complex biology, and rigorous safety standards.
- Computational screening value. Describe how computational approaches and machine learning can help drug discovery through rapid prototyping and screening, including assessing candidate safety.
- Interdisciplinary pipeline constraints. Explain why drug discovery teams may need to bridge different perspectives, and name factors beyond chemical, biological, and regulatory challenges that affect what a real pipeline can do.
- Search scale and computation. Explain why drug discovery cannot rely on testing every drug-like molecule and why machine learning offers a way to narrow the search.
- Chemical and biological search spaces. Describe how candidate molecules and biological targets define the drug-discovery search, and why the range of possible protein targets exceeds the gene count.
- Flexible binding and affinity. Explain why binding is not a rigid lock-and-key match and what binding affinity describes in this section.
- Eroom’s Law and its reversal. Describe Eroom’s Law as the long decline in pharmaceutical R&D efficiency and identify the change visible around 2010.
- Reversal causes and measurement limits. Explain the section’s account of the reversal and the limits of interpreting approvals per reported R&D spending as a complete measure of progress.
- Privileged structure trade-offs. Explain why medicinal chemists use privileged structures and what disadvantages can come with them.
- AI promise and approval barriers. Distinguish the section’s limited claim about AI-assisted discovery from a claim that machine learning caused the historical reversal or solves drug approval.
- FDA approval as gatekeeper. Explain why a drug candidate needs FDA approval before it can be marketed in the United States, and what that approval signifies.
- Clinical trial phases and attrition. Describe what Phases 1, 2, and 3 test, and explain why passing through successive phases leaves relatively few candidates approved.
- Molecular Representations as Model Inputs. Explain why representing a molecule to a computer is a central question for machine learning in drug discovery: the representation supplies input data to a model.
- Basic SMILES structure notation. Explain how SMILES represents molecular structure in text, including its basic notation for atoms and bonds.
- Canonical SMILES consistency. Explain why one structure can have multiple valid SMILES strings and what canonicalization accomplishes.
- Isomers and isomeric SMILES. Distinguish constitutional isomers from stereoisomers and explain why isomeric SMILES carries additional identifying information.
- Molecular graph representation. Describe how a molecular graph represents atoms and bonds and contrast it with a SMILES string.
- Machine Learning Within AI. Distinguish AI, machine learning, and deep learning, and explain how learning from examples differs from following manually coded rules.
- Training And Generalization. Trace the training and prediction phases in the dog-photo analogy, and explain why performance on unseen examples matters more than memorizing the training set.
- Representative Data For Prediction. Justify why training data must be suitable and representative, using the section’s examples of chemical diversity and cell-culture toxicity data.
- Supervised Molecular Prediction. Describe how labeled molecules and their features train a supervised model, and distinguish classification from regression using toxicity and solubility.
- Unsupervised Learning Applications. Explain what makes learning unsupervised and describe the purposes of clustering, representation learning or dimensionality reduction, compression, and generative modeling as presented here.
- RDKit molecule representations. Explain how RDKit molecule objects connect chemical structures to analysis: they store atoms, bonds, connectivity, and associated properties, and can be created from formats such as SMILES.
- Fingerprints as structural features. Describe how the example converts FDA-approved drugs into numerical ML features using molecular fingerprints, and what an ECFP6 fingerprint summarizes at the level taught here.
- PCA chemical space exploration. Explain why the example applies PCA to the fingerprint data and describe the qualified pattern visible in the resulting chemical-space plots.
- USAN stem classification boundaries. Describe how the example uses logistic regression on PCA output to classify membership in a USAN stem class and what its plotted decision boundaries and probabilities show.
- Virtual screening purpose. Explain why virtual screening is used and how it reduces the number of compounds needing experimental tests.
- Physical versus ML screening. Distinguish traditional simulation-based screening from ML-based screening in what each computes and why their screening speeds differ.
- Property prediction and selection. Trace how molecular representations and experimental property data support ML predictions and the selection of candidates with several desired properties.
- De Novo Design Approach. Explain how de novo design starts with desired properties and proposes new molecular structures, unlike virtual screening, which searches an existing library for suitable molecules.
- Scoring Function Limitations. Explain why a generative model's proposed molecules depend on the quality of its scoring function and how a weak score can produce misleading candidates.
- Learned Molecular Features. Explain why automatically learned, task-specific features can help deep learning discover novel drug candidates compared with relying only on hand-crafted chemical features.
- Forward and Retrosynthetic Planning. Distinguish forward synthesis planning from retrosynthesis by describing their starting points and the direction in which each searches for a route.
- Retrosynthesis Search Space. Explain why retrosynthesis is difficult despite its potential to find reliable or efficient routes, and describe the roles the section gives to chemists’ heuristics, rule-based tools, and deep learning.
- Protein Structure Prediction Value. Explain why predicting 3D protein structure from amino acid sequence is valuable and challenging, and what deep learning has contributed.
- Drug discovery phase sequence. Describe the phases from target identification through preclinical development, and explain why distinguishing them matters for AI optimization.
- Target selection and validation. Explain how a disease-related biomolecule becomes a proposed drug target and why its relationship to the disease must be validated.
- Hit discovery and selectivity. Explain what a hit is, how hits can be found, and why binding the intended target is not the only concern.
- Hit to lead selection. Describe how screening and assays narrow many hits to a smaller set of promising leads.
- Lead optimization and potency. Explain what is changed during lead optimization and distinguish efficacy from potency.
- Clinical trial progression. Describe how Phase I, II, and III trials differ in participants, scale, and purpose as a candidate moves toward approval.
- Expedited approval circumstances. Explain why some drugs receive an expedited timeline or conditional market approval before full registration, using the categories described here.
Section by section
Introduction
The chapter introduction explains why drug discovery is difficult and why computational screening and machine learning are useful.
- Drug discovery difficulty: Explain why developing a therapeutic is costly, slow, and prone to failure, connecting those outcomes to the vast number of possible drug-like molecules, complex biology, and rigorous safety standards.
Why it matters: These constraints explain why drug discovery needs better ways to evaluate candidates. - Computational screening value: Describe how computational approaches and machine learning can help drug discovery through rapid prototyping and screening, including assessing candidate safety.
Why it matters: It connects the difficulty of drug discovery to a concrete reason for using computational methods.
Key terms: drug-like molecules, drug candidates, clinical trial, computational approaches, machine learning, screening, drug candidate safety
1.1 The Challenge of Drug Discovery
Introduces the challenges that motivate machine learning in drug discovery and notes that practical pipelines depend on interdisciplinary collaboration.
- Interdisciplinary pipeline constraints: Explain why drug discovery teams may need to bridge different perspectives, and name factors beyond chemical, biological, and regulatory challenges that affect what a real pipeline can do.
Why it matters: A workable drug-discovery pipeline must fit the people and tools involved, not just the scientific problem.
Key terms: interdisciplinary, vocabularies, success criteria, communication, tooling
1.1.1 Needle in a Haystack
This section frames drug discovery as a vast search for molecules that act on biological targets, then explains why binding, experimental data, and other failure modes make that search difficult.
- Search scale and computation: Explain why drug discovery cannot rely on testing every drug-like molecule and why machine learning offers a way to narrow the search.
Why it matters: This explains why computational predictions are useful before committing resources to experiments. - Chemical and biological search spaces: Describe how candidate molecules and biological targets define the drug-discovery search, and why the range of possible protein targets exceeds the gene count.
Why it matters: Finding a molecule also means considering what biological forms it might affect. - Flexible binding and affinity: Explain why binding is not a rigid lock-and-key match and what binding affinity describes in this section.
Why it matters: Even when a team has chosen a target, its flexibility makes finding a suitable binder challenging. - Assay data and model quality: Describe what assays measure and how their results become training examples for machine learning models.
Why it matters: Predictions depend on the reliability of the measurements used to train the model. - Failures beyond target binding: Distinguish binding to a target from succeeding as a drug candidate by naming other ways a promising compound can fail.
Why it matters: A promising binding result is only one part of judging whether a candidate can become a drug.
Key terms: chemical search space, brute-force search, machine learning, biological search space, targets, proteoforms, conformations, free energy, binding affinity, assays
1.1.2 Eroom’s Law
The section explains Eroom’s Law, why its apparent reversal needs careful interpretation, and how risk-averse drug discovery strategies and AI fit into that picture.
- Eroom’s Law and its reversal: Describe Eroom’s Law as the long decline in pharmaceutical R&D efficiency and identify the change visible around 2010.
Why it matters: It establishes the trend that the rest of the section interprets. - Reversal causes and measurement limits: Explain the section’s account of the reversal and the limits of interpreting approvals per reported R&D spending as a complete measure of progress.
Why it matters: A better-looking approval ratio does not by itself show that researchers became better at finding drugs. - Me-too drugs and innovation: Explain why limited capital can encourage follow-on drugs, what those drugs are, and why that tendency does not describe every small firm.
Why it matters: It shows how financial risk can shape which candidate drugs receive investment. - Privileged structure trade-offs: Explain why medicinal chemists use privileged structures and what disadvantages can come with them.
Why it matters: A scaffold’s history of activity makes it attractive, but that same breadth and familiarity create risks. - AI promise and approval barriers: Distinguish the section’s limited claim about AI-assisted discovery from a claim that machine learning caused the historical reversal or solves drug approval.
Why it matters: It keeps AI’s potential contribution to early discovery separate from proof of clinical success.
Key terms: Eroom’s Law, R&D efficiency, FDA drug approvals, PhRMA member companies, oncology, rare disease, expedited pathways, me-too drugs, follow-on drugs, newer modalities
1.1.3 The Regulatory Landscape
This section explains how FDA approval and successive clinical-trial phases constrain whether a promising drug candidate can reach patients.
- FDA approval as gatekeeper: Explain why a drug candidate needs FDA approval before it can be marketed in the United States, and what that approval signifies.
Why it matters: Laboratory promise alone does not allow a drug to reach the U.S. market. - Clinical trial phases and attrition: Describe what Phases 1, 2, and 3 test, and explain why passing through successive phases leaves relatively few candidates approved.
Why it matters: Success in an early test does not mean a candidate is likely to become an approved drug.
Key terms: FDA, regulatory gatekeeper, safety, efficacy, FDA approval, preclinical testing, Phase 1, Phase 2, Phase 3, standard of care
1.2 The Language of Molecules
This introduction frames molecular representation as the bridge between small molecules and machine-learning models in drug discovery.
- Molecular Representations as Model Inputs: Explain why representing a molecule to a computer is a central question for machine learning in drug discovery: the representation supplies input data to a model.
Why it matters: This connects the chemistry being studied to the data a model can use. - Small Molecules as Starting Point: Justify the section's choice to begin with small molecules and their representations: they are concrete, widely supported, and a practical starting point for machine learning.
Why it matters: It explains why the book starts with small molecules rather than the broader biological setting.
Key terms: molecule, representations, input data, ML models, small molecules, practical on-ramp
1.2.1 What is a Molecule? The Joy of SMILES
This section introduces molecules and the two computational representations used throughout the book: SMILES strings and molecular graphs.
- Molecules and chemical bonds: Distinguish molecules from compounds and explain how covalent and noncovalent interactions differ.
Why it matters: These distinctions establish what the later representations depict. - Basic SMILES structure notation: Explain how SMILES represents molecular structure in text, including its basic notation for atoms and bonds.
Why it matters: SMILES gives the book a compact way to store and work with molecular structures. - Canonical SMILES consistency: Explain why one structure can have multiple valid SMILES strings and what canonicalization accomplishes.
Why it matters: A standardized representation makes equivalent structures easier to compare, index, and store. - Isomers and isomeric SMILES: Distinguish constitutional isomers from stereoisomers and explain why isomeric SMILES carries additional identifying information.
Why it matters: Compounds with the same formula may need to be distinguished in screening and drug design. - Chirality and biological effects: Explain chirality and why different spatial forms of a molecule can matter biologically, using the section's qualification of the thalidomide example.
Why it matters: It shows why a molecular representation may need to preserve spatial distinctions, not just a formula. - Molecular graph representation: Describe how a molecular graph represents atoms and bonds and contrast it with a SMILES string.
Why it matters: It establishes the second way the book represents molecules for computational work.
Key terms: molecule, compound, covalent bonds, noncovalent bonds, SMILES, atomic symbols, brackets, bond symbols, implicit bonds, canonical SMILES
1.2.2 What is Machine Learning?
The section explains how machine learning learns from examples, why it must generalize to new molecules, and how supervised and unsupervised learning differ.
- Machine Learning Within AI: Distinguish AI, machine learning, and deep learning, and explain how learning from examples differs from following manually coded rules.
Why it matters: This distinction explains why examples, rather than an exhaustive set of chemical rules, are central to the models in the book. - Training And Generalization: Trace the training and prediction phases in the dog-photo analogy, and explain why performance on unseen examples matters more than memorizing the training set.
Why it matters: Drug-discovery predictions need to work on molecules the model has not already seen. - Representative Data For Prediction: Justify why training data must be suitable and representative, using the section’s examples of chemical diversity and cell-culture toxicity data.
Why it matters: A prediction about a new compound is only useful if the model’s examples support that kind of prediction. - Supervised Molecular Prediction: Describe how labeled molecules and their features train a supervised model, and distinguish classification from regression using toxicity and solubility.
Why it matters: The kind of label determines whether a molecular prediction is treated as classification or regression. - Unsupervised Learning Applications: Explain what makes learning unsupervised and describe the purposes of clustering, representation learning or dimensionality reduction, compression, and generative modeling as presented here.
Why it matters: Not every useful molecular-data task requires a known answer for each example.
Key terms: artificial intelligence, machine learning, deep learning, artificial neural networks, training set, training instance, training phase, inference, prediction phase, test set
1.2.3 An Example Application with USAN Stems & RDKit
The section uses RDKit to turn drug structures into numerical features, then illustrates unsupervised exploration and supervised classification of drugs by USAN stem.
- RDKit molecule representations: Explain how RDKit molecule objects connect chemical structures to analysis: they store atoms, bonds, connectivity, and associated properties, and can be created from formats such as SMILES.
Why it matters: The example starts by turning drug structures into objects that software can analyze. - Fingerprints as structural features: Describe how the example converts FDA-approved drugs into numerical ML features using molecular fingerprints, and what an ECFP6 fingerprint summarizes at the level taught here.
Why it matters: Fingerprints provide the numerical bridge between chemical structures and the ML examples. - PCA chemical space exploration: Explain why the example applies PCA to the fingerprint data and describe the qualified pattern visible in the resulting chemical-space plots.
Why it matters: The plots show what structural patterns can—and cannot—be readily distinguished by eye in reduced chemical space. - USAN stem classification boundaries: Describe how the example uses logistic regression on PCA output to classify membership in a USAN stem class and what its plotted decision boundaries and probabilities show.
Why it matters: This turns the exploratory chemical-space view into a concrete example of supervised classification.
Key terms: RDKit, molecule object, Mol, ROMol, SMILES, molecular fingerprint, ECFP6, bit vector, substructures, PCA
1.3 Deep Learning’s Value Proposition
This introduction previews drug discovery applications of deep learning that the following sections will discuss.
1.3.1 Virtual Screening & Property Prediction
The section explains how virtual screening prioritizes drug candidates, contrasts physics-based and ML-based approaches, and describes the benefits and limits of screening at scale.
- Virtual screening purpose: Explain why virtual screening is used and how it reduces the number of compounds needing experimental tests.
Why it matters: It explains why computation is placed before experimental testing in the discovery pipeline. - Physical versus ML screening: Distinguish traditional simulation-based screening from ML-based screening in what each computes and why their screening speeds differ.
Why it matters: The distinction clarifies why a trained predictor can help explore more candidates than repeated physical simulations. - Property prediction and selection: Trace how molecular representations and experimental property data support ML predictions and the selection of candidates with several desired properties.
Why it matters: It connects molecular data to the practical decision of which predicted candidates deserve further investigation. - Screening scale and limits: Explain both the practical advantages of fast virtual screening and why its scale still does not cover drug-like chemical space.
Why it matters: Speed improves candidate triage without making an exhaustive search of possible drugs feasible.
Key terms: virtual screening, drug candidates, experimental screening, molecular docking, ligand, binding affinity, ML-based virtual screening, molecular representations, SMILES, molecular graph
1.3.2 Generative Chemistry
The section contrasts de novo molecular design with virtual screening and explains both the promise and limitations of AI-generated drug candidates.
- De Novo Design Approach: Explain how de novo design starts with desired properties and proposes new molecular structures, unlike virtual screening, which searches an existing library for suitable molecules.
Why it matters: This distinction clarifies whether an AI method is finding candidates in a library or proposing new structures. - Scoring Function Limitations: Explain why a generative model's proposed molecules depend on the quality of its scoring function and how a weak score can produce misleading candidates.
Why it matters: A high model score alone does not establish that a proposed molecule is a viable drug candidate. - Learned Molecular Features: Explain why automatically learned, task-specific features can help deep learning discover novel drug candidates compared with relying only on hand-crafted chemical features.
Why it matters: Learned features may help a search move beyond familiar chemical patterns when relevant design rules are not yet known.
Key terms: de novo design, generative chemistry, virtual screening, functional space, chemical space, scoring function, property-prediction models, reward hacking, hand-crafted chemical features, feature engineering
1.3.3 Chemical Reaction Prediction & Retrosynthesis
This section explains why novel drug designs need synthesis plans, how retrosynthesis works backward from a product, and why its large search space motivates automated planning.
- Synthetic Access to Novel Compounds: Explain why de novo drug design creates a synthesis challenge: it can propose compounds beyond those commercially available, but those compounds must be made before they can be used.
Why it matters: A promising proposed molecule is useful only if it can be made. - Forward and Retrosynthetic Planning: Distinguish forward synthesis planning from retrosynthesis by describing their starting points and the direction in which each searches for a route.
Why it matters: The direction of the search changes which synthesis routes chemists can consider. - Retrosynthesis Search Space: Explain why retrosynthesis is difficult despite its potential to find reliable or efficient routes, and describe the roles the section gives to chemists’ heuristics, rule-based tools, and deep learning.
Why it matters: The number of possible reaction steps explains why synthesis planning is a candidate for AI assistance. - Uses Beyond Novel Drugs: Describe how retrosynthesis can also help make existing drugs more rapidly or cost-effectively, and relate that use to process chemistry’s manufacturing concerns.
Why it matters: Synthesis planning matters not just for making a candidate once, but also for making an existing drug practical to manufacture.
Key terms: de novo drug design, synthetic chemistry, forward synthesis, retrosynthesis, reactants, product, chemical search space, transformations, heuristics, rule-based synthesis planning tools
1.3.4 Protein Folding & Simulations
This section explains why protein structure matters, how deep learning helps predict it, and how molecular dynamics adds information about protein motion.
- Amino Acid Sequence and Structure: Explain how a protein's amino acid sequence and interactions between amino acid sidechains guide its structure.
Why it matters: This connects a protein's sequence to the shape that helps determine what it does. - Protein Structure Prediction Value: Explain why predicting 3D protein structure from amino acid sequence is valuable and challenging, and what deep learning has contributed.
Why it matters: Prediction helps address the gap between the many available sequences and the fewer known structures. - Static Structures and Protein Motion: Distinguish a predicted static structure from the changing behavior modeled by molecular dynamics simulations, and describe how the two approaches complement each other.
Why it matters: A single predicted shape cannot show all the ways a protein moves and binds.
Key terms: amino acids, sidechains, protein structure, 3D structure, protein folding problem, AlphaFold2, deep learning, predicted structure, molecular dynamics (MD) simulations, static snapshot
1.4 Introducing Drug Discovery
This section maps the phases of drug discovery and distinguishes them from the clinical and regulatory steps of drug development.
- Drug discovery phase sequence: Describe the phases from target identification through preclinical development, and explain why distinguishing them matters for AI optimization.
Why it matters: The phase map helps you place later AI applications in the part of discovery they aim to improve. - Discovery and development boundary: Distinguish the discovery phases from what happens after a candidate clears them.
Why it matters: This boundary keeps preclinical discovery distinct from clinical testing and regulatory review.
Key terms: target identification or selection, target-to-hit, hit discovery, hit-to-lead, lead identification, lead optimization, drug discovery, drug development, clinical trials, FDA
1.4.1 Target Identification & Hit Discovery
The section explains how researchers choose and validate a disease-related target, find compounds that bind it, and consider how those compounds act and where else they might act.
- Target selection and validation: Explain how a disease-related biomolecule becomes a proposed drug target and why its relationship to the disease must be validated.
Why it matters: The choice of target gives compound discovery a biological direction. - Candidate target prioritization: Describe the supporting role computational tools can play when teams consider candidate targets.
Why it matters: Computational rankings can help organize uncertain evidence without replacing expert judgment. - Hit discovery and selectivity: Explain what a hit is, how hits can be found, and why binding the intended target is not the only concern.
Why it matters: Finding binding compounds is a starting point, not proof that they will act safely. - Enzyme inhibition and kinases: Explain how an enzyme inhibitor affects a reaction and why similar kinases pose a selectivity problem.
Why it matters: The kinase example shows why the effects of a compound depend on more than its intended target. - Ligand actions and treatment mechanisms: Distinguish agonists from antagonists and explain why the section’s inflammation and depression examples should not be treated as the same receptor mechanism.
Why it matters: Naming the molecule a drug acts on helps distinguish mechanisms that might otherwise sound alike.
Key terms: target identification, biomolecule target, validate, computational tools, large language model agents, candidate targets, hit, binding affinity, compound library, virtual screening
1.4.2 Lead Identification & Lead Optimization
This section explains how hits become leads, how leads are optimized for therapeutic properties, and what happens before clinical testing.
- Hit to lead selection: Describe how screening and assays narrow many hits to a smaller set of promising leads.
Why it matters: A promising hit is a starting point, not yet a lead ready for further development. - PK and PD distinction: Distinguish pharmacokinetic properties from pharmacodynamic properties using the examples given here.
Why it matters: The distinction organizes the properties considered when judging and improving a lead. - Lead optimization and potency: Explain what is changed during lead optimization and distinguish efficacy from potency.
Why it matters: Optimization aims to improve a lead, and a stronger maximum effect is not the same thing as needing a lower dose. - Safety and drug availability: Describe why selectivity, therapeutic index, and bioavailability matter when improving a lead.
Why it matters: A lead must not only act on its target but also avoid unwanted effects and reach an effective concentration. - Optimized leads before clinical trials: Describe what happens to optimized leads in preclinical study and what that study helps determine.
Why it matters: Preclinical evaluation is the bridge between an optimized candidate and decisions about testing it in patients. - De novo design motivation: Explain why the effort of moving from hits to preclinical candidates motivates interest in de novo design.
Why it matters: It connects the work of selecting and improving compounds to the appeal of generating candidates for desired properties.
Key terms: hits, leads, lead identification, ADMET screening, pharmacokinetic (PK), pharmacodynamic (PD), lead optimization, core molecular structure, efficacy, activity
1.4.3 Drug Development
This section traces drug candidates through clinical trials and expedited approval, then connects earlier modeling examples to the book’s broader drug-discovery workflow.
- Clinical trial progression: Describe how Phase I, II, and III trials differ in participants, scale, and purpose as a candidate moves toward approval.
Why it matters: The phases show how evidence about a candidate grows from initial human safety testing to a broader judgment of benefits and risks. - Expedited approval circumstances: Explain why some drugs receive an expedited timeline or conditional market approval before full registration, using the categories described here.
Why it matters: An expedited route explains why the usual approval timeline is not identical for every candidate. - Illustrative modeling workflow: Distinguish the limited teaching examples from clinically meaningful applications while stating the workflow they share with later applications.
Why it matters: The examples provide a reusable modeling pattern without implying that their results are ready for clinical use.
Key terms: Phase I, Phase II, Phase III, safety, efficacy, risk-benefit profile, expedited approval, conditional market approval, first-in-class drugs, orphan drugs
1.5 Summary
Where this is used later
- What is Virtual Screening? (chapter 2) · Develops virtual screening as a way to evaluate candidate molecules.
- Property-based Filters (chapter 2) · the drug-discovery context for pharmacokinetic properties
- (chapter intro) (chapter 10) · The difficulty of searching a vast space for a useful drug candidate
- Protein-Ligand Docking (chapter 9) · Examines protein-ligand docking, where the physical association of a molecule and target matters.
- Data Acquisition, Exploration, & Curation (chapter 3) · Addresses acquiring, exploring, and curating data for a machine learning task.
- Strategy: Similarity Searching (chapter 2) · Provides context for searching for compounds related to known active structures.
- Validating & Standardizing SMILES (chapter 3) · Assumes that a SMILES string represents a molecular structure when validating and standardizing strings.
- Virtual Screening Taxonomy (chapter 2) · Provides context for why precise molecular identity matters when screening compounds.
Concepts explained in this chapter
- machine learningMachine learning is a computational approach that learns patterns or predictive functions from data rather than relying solely on explicitly programmed rules.
- deep learningA deep neural network is a neural network composed of multiple computational layers that learn hierarchical representations from data.
- generative modelingA generative model learns patterns or a probability distribution from data and can produce new samples with similar characteristics.
- language modelA language model assigns probabilities to sequences or predicts the next token from preceding tokens.
- ligandA ligand is a molecule that binds to a biomolecular target and can alter its activity or signaling.
- SMILESSMILES is a line notation that encodes molecular atoms, bonds, and connectivity as a text string using defined symbols and grammar rules.
- molecular graphA molecular graph represents a molecule as a graph whose nodes are atoms and whose edges are chemical bonds.
- RDKitRDKit is an open-source cheminformatics toolkit for representing, manipulating, visualizing, and analyzing chemical structures.
- preclinical developmentA preclinical study evaluates a drug candidate's safety and efficacy before testing in human clinical trials, typically using animals or model organisms.
- artificial intelligenceArtificial intelligence is the field of computational methods for performing tasks that ordinarily require aspects of human intelligence, such as learning and reasoning.
- classifierClassification is a supervised learning task that predicts a discrete class label for each input.
- clusteringClustering is an unsupervised learning technique that partitions data instances into groups of similar instances.
- dimensionality reductionDimensionality reduction is the transformation of data into fewer dimensions while retaining important information.
- evaluation datasetA validation set is a held-out collection of examples used during development to assess model performance and guide configuration choices.
- feature vectorA feature vector is an ordered collection of numerical or categorical features representing one input instance.
- generalizationGeneralization is a model's ability to maintain predictive performance on previously unseen data drawn from the relevant distribution.
- generationInference is the process of applying a trained model to input data to produce predictions.
- linear regressionLinear regression is a regression method that models a linear relationship between input features and a target variable.
- neural networkAn artificial neural network is a parameterized computational model composed of interconnected units arranged in layers.
- overfittingOverfitting is the condition in which a model fits training examples too closely and performs poorly on new data.
- representation learningRepresentation learning is a technique in which a model learns useful encoded features from raw data.
- supervised learningSupervised learning trains a model from examples paired with target labels or values.
- trainingModel training is the process of fitting a model to learn a mapping from input features to target labels or values.
- training datasetA training set is the collection of examples provided to a learning algorithm for fitting a model.
- unsupervised learningUnsupervised learning trains a model from data without provided target labels.
- moleculeA molecule is an electrically neutral group of atoms held together by chemical bonds.
- drug discoveryThe process of identifying and characterizing chemical or biological entities that could become medicines.
- atomAn atom is the basic unit of a chemical element that participates in chemical bonding.
- canonical SMILESA unique SMILES is a canonicalized SMILES representation selected so that equivalent molecular structures receive the same textual representation.
- chiralityChirality is the property of an entity that cannot be superimposed on its mirror image.
- covalent bondA covalent bond is a chemical bond formed when atoms share electrons.
- enantiomerAn enantiomer is one of a pair of stereoisomers that are non-superimposable mirror images of each other.
- isomerAn isomer is one of multiple compounds with the same molecular formula but different structural arrangement or spatial orientation.
- isomeric SMILESIsomeric SMILES is a SMILES representation that explicitly encodes stereochemical and other isomer-specific structural information.
- stereochemistryStereochemistry is the study and representation of the three-dimensional spatial arrangement of atoms in molecules.
- stereoisomerA stereoisomer is a compound with the same molecular formula and connectivity as another compound but a different spatial arrangement of atoms.
- pharmacokineticsPharmacokinetics is the study of how an administered drug is absorbed, distributed, metabolized, and eliminated by the body.
- molecular conformationA conformation, or conformer, is a particular three-dimensional arrangement of a molecule accessible through internal structural flexibility without changing its connectivity.
- hitA hit is a compound that shows measurable binding or activity against a selected biological target during discovery screening.
- generative chemistryGenerative molecular design is the computational creation of molecular structures intended to satisfy specified biological, chemical, or medicinal requirements.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.
