Chapter 3: Ligand-based Screening: Machine Learning
Introduces the stages and vocabulary of a machine-learning workflow for cardiotoxicity prediction and explains why the workflow is iterative.
After this chapter you will be able to
- Machine Learning Workflow Stages. Describe how a machine-learning project moves from defining the problem through preparing data, training and evaluating a model, deploying it, and checking whether it remains useful.
- Features Labels and Training Instances. Explain what features, labels, and training instances are and how model training relates them.
- Iterative Workflow Decisions. Explain why a machine-learning workflow may return to earlier stages, using the section's examples of poor evaluation, unmet production objectives, or new data.
- Screening Project Requirements. Explain why the screening model must be designed around the project’s requirements, constraints, and end goal, not accuracy alone. Describe its place in further filtering hits found through similarity searching.
- hERG Screening Decisions. Describe why hERG blockers are filtered out, what the two prediction labels mean, and why a non-blocker prediction does not establish safety. Explain how uncertain predictions and the two kinds of mistaken decision are treated as distinct concerns.
- Conditions for useful machine learning. Explain why machine learning may add value when labeled data contains learnable patterns, enough examples support stable relationships, and predictions can be made quickly at scale for similar new molecules.
- hERG classification task definition. Describe the hERG prediction task as assigning each compound to blocking or non-blocking classes, and explain why it is both binary classification and supervised learning.
- Molecular Dataset Sources. Describe where molecular property prediction datasets can be found and explain why a dataset aggregator is useful when the relevant publications are unfamiliar.
- Blocker Label and Potency. Explain how the binary hERG blocker label relates to IC50 and pIC50, including what higher pIC50 means.
- Potency Exclusion From Features. Justify why pIC50 should not be used as a model-training feature despite its clear separation of the classes.
- Exploratory Data Quality Checks. Explain why dataset curation and exploratory data analysis are needed, and describe what an unexpected pIC50 distribution could reveal.
- Purpose of molecular standardization. Distinguish molecular standardization from validation and explain why aligning compound representations supports comparison and reproducibility.
- Standardization operations and limits. Describe the main operations in the RDKit standardization pipeline and explain why its canonical tautomer is reproducible but not necessarily the biologically dominant form.
- Morgan fingerprint representations. Describe how the 587 standardized molecules are converted into 2048-dimensional Morgan fingerprints that give each molecule a numerical representation.
- Fingerprint bit distributions. Describe the sparse pattern of set bits across the hERG fingerprints, including how bit frequencies differ across compounds and how many bits a typical molecule sets. Explain the section’s suggested interpretation of unusually frequent bits.
- Feature to property mapping. Explain that, during training, an ML model uses data encoded as features to learn a mapping to the property it is meant to predict.
- hERG Input Output Mapping. Explain how the hERG prediction task maps a compound’s numerical features to a yes-or-no hERG-blocking label, using experimentally labeled compounds as examples.
- Ground Truth Approximation. Explain why the model must approximate an unknown target function from sampled training examples and why correct training predictions alone are not enough.
- Training Weights and Errors. Describe how training adjusts model weights using a loss for individual prediction errors, a cost that aggregates errors, and a routine that minimizes that cost.
- Fingerprint to hERG predictions. Explain that the linear classifier learns a correspondence between molecular fingerprints and whether a molecule blocks the hERG channel, using the training portion of the predefined split and Scikit-Learn’s SGDClassifier.
- Training accuracy and its limits. State that the trained classifier can generate predictions and that its training-set accuracy is 96.4%, but explain that this result alone does not establish that hERG blocker prediction is solved.
- Protected final test set. Explain why the final test set should be reserved until a model has been selected, and why validation should instead use the training data.
- K-fold validation process. Describe how k-fold cross-validation trains and validates models on the training set, how results are combined, and what changes when k increases.
- Imbalanced accuracy and confusion matrix. Explain why accuracy alone is misleading for this dataset and describe what a confusion matrix adds.
- Precision recall and F1. Distinguish precision from recall using positive predictions and actual positives, and explain why F1 considers both.
- Training validation performance gap. Interpret the reported training-versus-validation scores and state what they imply about the current hERG blocker model.
- Data-first model improvement. Explain why the reader should investigate the data before trying fancier models, and describe the kinds of checks that may reveal performance problems.
- Large coefficients and overfitting. Explain why coefficients much larger than the model’s +1 and −1 outputs raise concern about overfitting, and how fitting training-set noise can increase generalization error.
- Regularization penalties and strength. Describe how regularization constrains coefficients through a cost-function penalty, distinguish L1, L2, and elastic net, and state what increasing λ does.
- Purpose of Hyperparameter Tuning. Explain why hyperparameters are tuned and identify the kinds of settings varied in this molecular classification example.
- Pipeline Structure and Leakage. Describe how transformers and an estimator form the pipeline, and explain why preprocessing belongs inside cross-validation rather than being fitted on all data beforehand.
- Grid Search Pipeline Results. Trace what the molecular pipeline and GridSearchCV do, and state what the best cross-validated result represents.
- Grid Versus Randomized Search. Distinguish exhaustive grid search from randomized search and explain when randomized search is useful.
- Held-Out Model Evaluation. Explain why the selected model is evaluated on the test set and interpret its test F1 score alongside its cross-validated F1 score.
- Fingerprint Weight Interpretation Limits. Explain how weighted fingerprint bits can suggest structural relationships while limiting what can be inferred about any one structure.
- Performance Benchmark Comparisons. Describe the comparison points for the final model and interpret its reported MCC against published work.
- Chemical Applicability Domain. Distinguish performance on held-out dataset molecules from reliable predictions for structurally different molecules.
- Saved model application. Explain that a Scikit-Learn model can be saved, loaded, and used to predict on additional data, such as the top 1000 Malaria Box hits identified in chapter 2. State that loading a model with joblib also requires importing any custom classes or functions it uses.
What it builds on
- Similarity Searching (chapter 2) · Provides the earlier ligand-based screening context.
- What is Machine Learning? (chapter 1) · The introductory concept of machine learning.
- Understanding the Dataset as a Structure Data File (chapter 2) · Working with a dataset of molecular structures.
- Molecule Sanitization (chapter 2) · Earlier preparation of molecules before using them computationally.
- Hashed Fingerprints (chapter 2) · Morgan fingerprints were introduced as a molecular representation.
- Fingerprinting our Library (chapter 2) · The use of molecular fingerprints as features.
- Fingerprints: Representing Molecules as Numbers (chapter 2) · Fingerprints represent molecules using bits.
- Searching against a Query (chapter 2) · The search that identified compounds to which the model could be applied.
Section by section
Introduction
Introduces the stages and vocabulary of a machine-learning workflow for cardiotoxicity prediction and explains why the workflow is iterative.
- Machine Learning Workflow Stages: Describe how a machine-learning project moves from defining the problem through preparing data, training and evaluating a model, deploying it, and checking whether it remains useful.
Why it matters: The stages give you a way to place each modeling task in the larger project. - Features Labels and Training Instances: Explain what features, labels, and training instances are and how model training relates them.
Why it matters: These terms clarify what the model receives and what it is trained to predict. - Iterative Workflow Decisions: Explain why a machine-learning workflow may return to earlier stages, using the section's examples of poor evaluation, unmet production objectives, or new data.
Why it matters: A disappointing result may call for revisiting an earlier decision rather than merely proceeding to the next stage.
Key terms: problem definition, data acquisition, data exploration, pre-processing, featurization, model training, features, training instance, labels, iterative
3.1 Problem Understanding
This section frames hERG blocker screening as a project-defined filtering task and explains the predictions, risks, and benchmarks that matter for judging a model.
- Screening Project Requirements: Explain why the screening model must be designed around the project’s requirements, constraints, and end goal, not accuracy alone. Describe its place in further filtering hits found through similarity searching.
Why it matters: A high accuracy score alone cannot tell you whether the model serves the screening project. - hERG Screening Decisions: Describe why hERG blockers are filtered out, what the two prediction labels mean, and why a non-blocker prediction does not establish safety. Explain how uncertain predictions and the two kinds of mistaken decision are treated as distinct concerns.
Why it matters: The meaning and consequences of each prediction determine how the filter should be used. - Model Success Benchmarks: Describe the proposed ways to benchmark the model, including comparison with other models and a non-ML structural-alert filter. Explain, at the level given here, how that filter could use structures associated with arrhythmia endpoints as a proxy for hERG blockage.
Why it matters: A comparison gives the model’s performance meaning beyond its score in isolation.
Key terms: virtual screening pipeline, project requirements, Malaria Box compounds, antitarget, off-target, hERG potassium channel, hERG non-blocker, hERG blocker, low-confidence compounds, benchmark
3.1.1 Your Machine Learning Task
This section explains when machine learning may be useful for screening compounds and defines hERG blocking prediction as a supervised binary classification task.
- Conditions for useful machine learning: Explain why machine learning may add value when labeled data contains learnable patterns, enough examples support stable relationships, and predictions can be made quickly at scale for similar new molecules.
Why it matters: These conditions help you judge whether a predictive model is worth trying for a screening problem. - hERG classification task definition: Describe the hERG prediction task as assigning each compound to blocking or non-blocking classes, and explain why it is both binary classification and supervised learning.
Why it matters: Defining the classes and labels makes clear what the model must learn and predict.
Key terms: capacity, training data, generalize, unseen molecules, positive class, negative class, binary classification, multiclass classification, supervised learning, labeled dataset
3.2 Data Acquisition, Exploration, & Curation
This section identifies sources of molecular property data and explains when a dataset aggregator helps readers find relevant datasets.
- Molecular Dataset Sources: Describe where molecular property prediction datasets can be found and explain why a dataset aggregator is useful when the relevant publications are unfamiliar.
Why it matters: Knowing where to look is the first step toward finding data for a molecular prediction task.
Key terms: PubChem, ChEMBL, benchmark datasets, dataset aggregator, Therapeutics Data Commons
3.2.1 Loading and Exploring the hERG Blockers Dataset
This section introduces the hERG blockers dataset, identifies the target and split columns, and uses exploratory analysis to explain why data quality matters before modeling.
- Dataset Rows and Columns: Describe what a row represents in the hERG blockers dataset and identify the roles of its structure, potency, class, and split columns.
Why it matters: Keeping the columns’ roles straight prevents the model from using the answer as an input. - Blocker Label and Potency: Explain how the binary hERG blocker label relates to IC50 and pIC50, including what higher pIC50 means.
Why it matters: The potency threshold explains where the label comes from and why potency and class are closely related. - Potency Exclusion From Features: Justify why pIC50 should not be used as a model-training feature despite its clear separation of the classes.
Why it matters: A feature that gives away the label would not support the intended pre-experiment prediction. - Exploratory Data Quality Checks: Explain why dataset curation and exploratory data analysis are needed, and describe what an unexpected pIC50 distribution could reveal.
Why it matters: Checking the data before modeling can catch errors that would otherwise undermine training or evaluation.
Key terms: SMILES, pIC50, Class, Set I, Set II, IC50, hERG blockers, feature, class variable, curation
3.2.2 Validating & Standardizing SMILES
This section uses visual inspection to motivate standardizing molecular structures, then describes the standardization operations and their limits.
- Structural patterns and outliers: Explain how viewing compounds at the extremes of the dataset can reveal structural patterns associated with hERG blockage and possible structure errors, while leaving complex pattern detection to the ML model.
Why it matters: Inspecting structures helps catch questionable data before modeling without requiring you to define every predictive pattern by hand. - Purpose of molecular standardization: Distinguish molecular standardization from validation and explain why aligning compound representations supports comparison and reproducibility.
Why it matters: Consistent representations make comparisons across collected compound data more meaningful. - Standardization operations and limits: Describe the main operations in the RDKit standardization pipeline and explain why its canonical tautomer is reproducible but not necessarily the biologically dominant form.
Why it matters: The pipeline produces comparable structures, but its chosen form should not be mistaken for the form a molecule necessarily takes in an assay.
Key terms: hERG blockers, structural patterns, outliers, structural alerts, molecular validation, molecular standardization, reproducibility, rdMolStandardize, largest parent fragment, neutralize
3.2.3 Feature Generation & Exploration
The section represents standardized hERG molecules with Morgan fingerprints and explores how frequently their bits are set before modeling.
- Morgan fingerprint representations: Describe how the 587 standardized molecules are converted into 2048-dimensional Morgan fingerprints that give each molecule a numerical representation.
Why it matters: The fingerprints are the molecular features this chapter will use for modeling. - Fingerprint bit distributions: Describe the sparse pattern of set bits across the hERG fingerprints, including how bit frequencies differ across compounds and how many bits a typical molecule sets. Explain the section’s suggested interpretation of unusually frequent bits.
Why it matters: Looking at bit frequencies reveals what the numerical representation is like before it is used in a model. - Fingerprint size tuning: Justify the section’s advice to tune fingerprint length and Morgan radius, and describe its suggested starting lengths for a dataset of 587 molecules.
Why it matters: The number of fingerprint features should be considered alongside the amount of available training data.
Key terms: Morgan fingerprints, fingerprint bits, 2048-dimensional representations, sparse, outlier bits, bits set per molecule, fingerprint length, nBits, Morgan radius, overfitting
3.3 Application of Linear Models
This section introduces modeling as learning a mapping from molecular features to a property to predict.
- Feature to property mapping: Explain that, during training, an ML model uses data encoded as features to learn a mapping to the property it is meant to predict.
Why it matters: This connects the features prepared earlier to the predictions the model will make.
Key terms: training, features, property
3.3.1 Learning from Data
This section explains how labeled hERG data supports a supervised model, what the model learns during training, and how it is used for inference.
- hERG Input Output Mapping: Explain how the hERG prediction task maps a compound’s numerical features to a yes-or-no hERG-blocking label, using experimentally labeled compounds as examples.
Why it matters: This identifies exactly what the model receives and what it is meant to predict. - Ground Truth Approximation: Explain why the model must approximate an unknown target function from sampled training examples and why correct training predictions alone are not enough.
Why it matters: A useful screening model must work beyond the compounds whose results were already measured. - Training Weights and Errors: Describe how training adjusts model weights using a loss for individual prediction errors, a cost that aggregates errors, and a routine that minimizes that cost.
Why it matters: These are the roles of the quantities involved when the model learns from labeled examples. - Frozen Inference and Parametric Models: Distinguish inference from training, and describe what makes the chapter’s linear and logistic regression models parametric.
Why it matters: It clarifies what remains fixed when the trained model is applied to new compounds.
Key terms: input-output examples, molecular descriptors, structural fingerprints, hERG blocking, target function, ground truth, training dataset, generalizing, weights, parameters
3.3.2 Training our Linear Model
This section trains a linear classifier to predict hERG channel blocking from molecular fingerprints, then cautions against treating high training-set accuracy as a solved prediction problem.
- Fingerprint to hERG predictions: Explain that the linear classifier learns a correspondence between molecular fingerprints and whether a molecule blocks the hERG channel, using the training portion of the predefined split and Scikit-Learn’s SGDClassifier.
Why it matters: This connects the molecule representation to the prediction the model is trained to make. - Training accuracy and its limits: State that the trained classifier can generate predictions and that its training-set accuracy is 96.4%, but explain that this result alone does not establish that hERG blocker prediction is solved.
Why it matters: It prevents a strong result on the training data from being mistaken for proof that the prediction problem is solved.
Key terms: fingerprints, hERG channel, linear classifier, training set, test set, SGDClassifier, accuracy, binary classifier
3.3.3 Evaluating our Model
This section evaluates a hERG blocker classifier with cross-validation and classification metrics, revealing that its strong training performance does not carry over to unseen data.
- Protected final test set: Explain why the final test set should be reserved until a model has been selected, and why validation should instead use the training data.
Why it matters: Keeping the test set untouched makes the final evaluation more credible. - K-fold validation process: Describe how k-fold cross-validation trains and validates models on the training set, how results are combined, and what changes when k increases.
Why it matters: Cross-validation supports model comparison while making use of limited training data. - Similar analog validation risk: Explain why randomly assigned folds can make cross-validated performance look better than real-world performance for molecular data.
Why it matters: A reassuring validation score can be misleading if validation molecules closely resemble training molecules. - Imbalanced accuracy and confusion matrix: Explain why accuracy alone is misleading for this dataset and describe what a confusion matrix adds.
Why it matters: A single accuracy number can hide which classification errors the model makes. - Precision recall and F1: Distinguish precision from recall using positive predictions and actual positives, and explain why F1 considers both.
Why it matters: These metrics let you discuss false alarms and missed positives separately rather than relying on accuracy alone. - Training validation performance gap: Interpret the reported training-versus-validation scores and state what they imply about the current hERG blocker model.
Why it matters: Strong training results do not mean the hERG blocker prediction challenge has been solved.
Key terms: test set, generalization error, validation set, k-fold cross validation, folds, validation fold, near-identical analogs, training and validation folds, splitting strategy, accuracy
3.4 Improving our Model
This section argues for checking the data for sources of poor performance before trying model-centric improvements.
- Data-first model improvement: Explain why the reader should investigate the data before trying fancier models, and describe the kinds of checks that may reveal performance problems.
Why it matters: A data problem may limit performance even if you spend time improving the model itself.
Key terms: class balance, duplicate or near-duplicate compounds, chemical scaffolds, cross-validation fold, false positives, false negatives
3.4.1 Regularization
This section uses unusually large model coefficients to motivate regularization, then explains how common penalties constrain weights.
- Large coefficients and overfitting: Explain why coefficients much larger than the model’s +1 and −1 outputs raise concern about overfitting, and how fitting training-set noise can increase generalization error.
Why it matters: This explains why the model’s coefficient plot motivates a change to the model. - Correlated fingerprint features: Explain why fingerprint bits may be correlated and how that correlation can make coefficients unstable and hard to interpret.
Why it matters: Large weights are not solely a sign of fitting noise; correlated inputs can also complicate their interpretation. - Regularization penalties and strength: Describe how regularization constrains coefficients through a cost-function penalty, distinguish L1, L2, and elastic net, and state what increasing λ does.
Why it matters: The penalty provides a way to limit the large weights identified in the model.
Key terms: coefficient weights, overfitting, generalization error, multicollinearity, fingerprint bits, features, regularization, cost function, L1 regularization, L2 regularization
3.4.2 Non-linear Transformation
3.4.3 Hyperparameter Tuning
This section shows how to tune a molecular classification pipeline with grid and randomized search, while keeping evaluation leakage-free and interpreting the results.
- Purpose of Hyperparameter Tuning: Explain why hyperparameters are tuned and identify the kinds of settings varied in this molecular classification example.
Why it matters: The point of tuning is a model that works on new molecules, not merely one that fits its training examples. - Pipeline Structure and Leakage: Describe how transformers and an estimator form the pipeline, and explain why preprocessing belongs inside cross-validation rather than being fitted on all data beforehand.
Why it matters: A pipeline makes the search easier to rerun without giving the model information it would not have when predicting new molecules. - Grid Search Pipeline Results: Trace what the molecular pipeline and GridSearchCV do, and state what the best cross-validated result represents.
Why it matters: The selected score comes from cross-validation, while the refitted pipeline is what can subsequently make test-set predictions. - Grid Versus Randomized Search: Distinguish exhaustive grid search from randomized search and explain when randomized search is useful.
Why it matters: The search method determines how much of the setting space can be explored for the available computation. - Polynomial Fingerprint Search Outcome: Explain why the randomized-search variant restricted fingerprint size and why its polynomial features performed worse than the linear grid-search model.
Why it matters: Adding more elaborate features does not necessarily improve predictions when the starting representation already captures useful structure. - Ensemble Learning Tradeoff: Describe the basic ensemble-learning idea and its stated benefit and cost.
Why it matters: Ensembling offers another possible performance improvement, with a computational tradeoff.
Key terms: hyperparameter tuning, underfits, overfits, Morgan fingerprint, regularization, pipeline, transformer, estimator, cross-validation, data leakage
3.4.4 Evaluating the Best Model
The section evaluates the selected hERG-blocker model on held-out data, interprets its fingerprint weights, compares its performance with benchmarks, and limits claims to an appropriate applicability domain.
- Held-Out Model Evaluation: Explain why the selected model is evaluated on the test set and interpret its test F1 score alongside its cross-validated F1 score.
Why it matters: Held-out performance is a better check of the selected model than relying on its cross-validation score alone. - Polynomial Fingerprint Limitations: Explain why polynomial features added little to this bit-vector fingerprint model.
Why it matters: More generated features do not necessarily add useful chemical information. - Regularized Coefficient Interpretation: Describe what the controlled magnitude of the final model’s coefficients suggests about generalization.
Why it matters: The coefficient plot gives a way to relate regularization to the model’s improved generalization. - Fingerprint Weight Interpretation Limits: Explain how weighted fingerprint bits can suggest structural relationships while limiting what can be inferred about any one structure.
Why it matters: Interpreting influential bits is useful only if their structural ambiguity is kept in view. - Performance Benchmark Comparisons: Describe the comparison points for the final model and interpret its reported MCC against published work.
Why it matters: Benchmarks put a model score in context rather than treating it as meaningful in isolation. - Chemical Applicability Domain: Distinguish performance on held-out dataset molecules from reliable predictions for structurally different molecules.
Why it matters: A good test score does not by itself make predictions trustworthy across all of chemical space.
Key terms: test set, cross-validated F1-score, grid search, experimental variability, polynomial features, bit vector, molecular fingerprints, sparsity, regularization, coefficients
3.4.5 Saving and Applying our Model
This subsection explains how to save and reload a trained model for predictions on additional compounds.
- Saved model application: Explain that a Scikit-Learn model can be saved, loaded, and used to predict on additional data, such as the top 1000 Malaria Box hits identified in chapter 2. State that loading a model with joblib also requires importing any custom classes or functions it uses.
Why it matters: Saving the model lets you apply it to new compounds without rebuilding it for each prediction task.
Key terms: Scikit-Learn, joblib, custom classes or functions, Malaria Box compounds
3.5 Summary
Where this is used later
- Protein Kinases (chapter 8) · Assumes familiarity with acquiring and curating bioactivity data.
- Target Variable Distribution (chapter 4) · The earlier approach of standardizing molecules by removing fragments.
- Models as Computational Graphs (chapter 8) · Provides context for learning from supplied features.
- Solubility with Linear Regression (chapter 4) · The earlier training of a linear classification model.
- Load the Data (chapter 4) · The earlier use of random splitting to validate classification models.
- Logistic Regression in Practice (chapter 5) · Assumes prior experience evaluating a trained model.
- (Mis)calibration: Questioning Probabilistic Output Assumptions (chapter 5) · Precision and recall as classification performance measures.
- Dealing with Data Set Imbalance (chapter 5) · The distinction between accuracy and class-specific errors remains important when dealing with dataset imbalance.
Concepts explained in this chapter
- f-scoreThe F1 score is the harmonic mean of precision and recall for a classification system.
- data leakageData leakage is the inadvertent use of information outside a training partition during model fitting, producing biased performance estimates.
- cross-validationCross-validation estimates model performance or selects configurations by repeatedly training and evaluating on different partitions of the training data.
- grid searchGrid search is a hyperparameter optimization method that evaluates models across a specified Cartesian product of candidate parameter values.
- RandomizedSearchCVRandomized search is a hyperparameter optimization method that evaluates configurations sampled from specified parameter distributions or candidate ranges.
- hyperparameter tuningHyperparameter tuning is the systematic selection of model configuration values that are not learned directly from the training examples.
- data explorationExploratory data analysis is the examination and summarization of dataset attributes, distributions, and relationships to identify patterns and inconsistencies.
- regularizationRegularization is a modeling technique that constrains learning to reduce overreliance on particular features and improve performance on unseen data.
- polynomial feature interactionA polynomial feature is a feature formed by applying powers or products of original input features to represent nonlinear interactions.
- hERG inhibitionhERG blockage is inhibition of the human ether-à-go-go-related gene potassium channel by a chemical compound.
- hERG potassium channelThe hERG potassium channel is a cardiac ion channel whose drug blockade can disrupt electrical repolarization and cause cardiotoxicity.
- linear regressionLinear regression is a regression method that models a linear relationship between input features and a target variable.
- generalizationGeneralization is a model's ability to maintain predictive performance on previously unseen data drawn from the relevant distribution.
- accuracyAccuracy is the proportion of all classified instances that receive the correct predicted class.
- machine learningMachine learning is a computational approach that learns patterns or predictive functions from data rather than relying solely on explicitly programmed rules.
- problem definitionProblem definition is the specification of a predictive or analytical task, its success criteria, and whether a chosen method is appropriate.
- Scikit-Learn pipelineScikit-Learn Pipeline is an API object that chains data transformations and a final estimator into a single fitted workflow.
- estimatorAn estimator is a machine-learning component that fits a predictive model from input data and target values.
- hyperparameterA hyperparameter is a configuration value selected before or outside model fitting that controls an algorithm, representation, preprocessing, or model complexity.
- transformer architectureA transformer architecture is a neural-network architecture that models relationships among sequence elements using attention and position-wise transformations, potentially with encoder and decoder stacks.
- evaluation datasetA validation set is a held-out collection of examples used during development to assess model performance and guide configuration choices.
- confusion matrixA confusion matrix is a tabulation of predicted versus actual class counts, including true positives, true negatives, false positives, and false negatives.
- precisionPrecision is the proportion of predicted positive instances that are truly positive.
- evaluationModel evaluation is the assessment and comparison of model predictions, including analysis of predictive errors.
- trainingModel training is the process of fitting a model to learn a mapping from input features to target labels or values.
- applicability domainAn applicability domain is the region of input chemical or feature space in which a predictive model is expected to make reliable predictions.
- interpretabilityModel interpretability is the degree to which a model's predictions, representations, or decision process can be understood by humans.
- matthews correlation coefficientThe Matthews correlation coefficient is a binary-classification metric computed from all four entries of the confusion matrix and ranging from -1 to +1.
- IC50IC50, or the half-maximal inhibitory concentration, is the concentration of a compound required to reduce a specified biological activity by 50 percent under defined experimental conditions.
- data curationData curation is the systematic review, correction, normalization, and organization of data to improve its quality and consistency.
- pIC50pIC50 is the negative base-ten logarithm of a compound's half-maximal inhibitory concentration expressed in molar units.
- random splitA random split partitions dataset examples into subsets by randomly assigning examples to each subset.
- supervised learningSupervised learning trains a model from examples paired with target labels or values.
- linear classifierLinear classification assigns discrete classes using a threshold on a linear combination of input features.
- classifierClassification is a supervised learning task that predicts a discrete class label for each input.
- molecular standardizationMolecular standardization is the systematic normalization of molecular representations and structural information into a consistent form.
- objective functionA cost function measures aggregate prediction error across multiple data instances, with lower values representing better performance.
- parametric modelA parametric model assumes a specified functional form and represents it with a fixed number of parameters.
- target functionA target function is the unknown input-output mapping that defines the true outcome of a prediction task.
- ligand-based virtual screeningLigand-based virtual screening prioritizes compounds by using information from known ligands rather than an explicit target structure.
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.
