Chapter 6: Case Study: Small Molecule Binding to an RNA Target
The introduction explains why small-molecule binding to RNA is therapeutically promising but difficult to predict, and frames QSAR as a way to relate molecular descriptors to measured binding.
After this chapter you will be able to
- RNA Targeting Challenges. Explain why selective RNA binding and predictive modeling are difficult, connecting similar RNA motifs, conformational flexibility, and limited experimental binding data.
- QSAR Structure Binding Relationship. Describe the relationship the proposed QSAR model seeks to establish between molecular descriptors and experimentally observed RNA-binding profiles, and what an interpretable model could reveal.
- TAR as a model system. Explain what a model system is and why TAR RNA serves that role in this QSAR case study.
- Importance of targeting TAR. Explain why disrupting TAR could matter medically and why the work may be relevant beyond HIV.
- TAR RNA antiviral rationale. Explain why HIV-1 TAR RNA is a potential antiviral target: the viral Tat protein must bind this RNA stem-loop to enable efficient transcription, so blocking their interaction is a validated strategy.
- Small ligand dataset diversity. Describe the experimental ligand set used for modeling: it contains 48 compounds, including 29 known TAR ligands and 19 known RNA-targeted scaffolds, drawn from five categories to represent varied structures and binding characteristics.
- Protomer States and pH. Explain what protomers are and why the workflow generates protonation and tautomerization states over a pH range suited to the binding experiment.
- Conformer Sampling and Geometry. Describe why a ligand needs multiple 3D conformers and how RDKit generates plausible starting geometries with distance geometry and ETKDG.
- Conformer Energy Selection. Trace how the example refines generated conformers and selects those used for descriptor calculation.
- Weighted Ligand Descriptors. Explain how descriptors from multiple states become one ligand vector and qualify the cross-protomer weighting used here.
- RNA binding response variables. Explain what kon, koff, and KD say about a ligand binding to HIV-1 TAR, including how KD is calculated.
- SPR binding measurement. Trace how SPR detects ligand binding at the sensor surface and how its measurements yield association and dissociation rates.
- Descriptor Feature Design. Explain why designing molecular features is an important part of the ML pipeline, including how aggregating descriptors across a ligand’s protomers or conformers can better reflect its real properties.
- Log-Transformed Binding Responses. Explain why the binding response variables are log-transformed before QSAR modeling, including the effects on skew, variance, and relationships a linear model can capture.
- Redundant Descriptor Filtering. Describe which near-constant and highly correlated descriptors are removed and why redundancy is a problem for the model.
- Correlated Descriptor Selection. Trace how the section chooses among highly correlated descriptors and explain why using the full dataset's response correlations risks data leakage.
- Representative split and applicability. Explain why the 48-ligand dataset needs a representative split and why training samples should span both the range and intermediate variation of binding activity.
- Kennard-Stone sample selection. Trace how the Kennard-Stone algorithm selects a diverse training subset using distances between samples.
- High-dimensional distance limitations. Explain why distance-based selection becomes statistically less reliable with 48 ligands in a 191-dimensional descriptor space, and distinguish that issue from computational cost.
- Dimensionality Reduction Rationale. Explain why adding data alone cannot readily overcome the curse of dimensionality and why focusing on informative dimensions can help.
- Principal Component Construction. Describe how PCA creates ranked, orthogonal components and projects samples onto a selected number of them.
- Retained Component Criteria. Explain the trade-off in choosing k and describe the three criteria presented for selecting it.
- Split Design Versus Model Leakage. Distinguish fitting PCA for split design from fitting it as part of a predictive modeling pipeline.
- Interpretable Model and Feature Selection. Explain why multiple linear regression (MLR) is a starting point for this QSAR task and why Lasso feature selection precedes it.
- Candidate MLR Model Selection. Trace how the authors search for and choose an MLR model after feature selection, including the complexity limit and evaluation criteria.
- Small Dataset Reliability Limits. Explain why the reported single-split QSAR performance should be treated cautiously and what stronger validation would report.
- Descriptor-Based Activity Hypotheses. Explain how a QSAR equation can suggest which compound descriptors influence activity and how changing a compound might alter a predicted outcome. Distinguish an easily inspected descriptor from one that must be recalculated and is harder to explain.
- Regulatory Interpretation Boundaries. Describe the qualities the section identifies for regulatory use of a QSAR model and explain what limits the validity of its interpretations.
- Statistical and Expert Validation. Explain why assessing QSAR performance statistically is not enough and how expert review can challenge an apparently important feature.
- RNA Binding Descriptor Differences. Explain why a model that distinguishes RNA-binding from protein-binding compounds could help identify molecular descriptors associated with RNA-binding activity.
- Gradient Boosting Comparison Results. Describe how the XGBClassifier was evaluated and how its reported performance compared with the published LASSO and neural network results.
- Parallel and Sequential Ensembles. Distinguish random forests from boosting in how they train and combine models, and explain the different errors their ensembles primarily reduce.
- Residual Based Gradient Boosting. Trace how the decision-tree gradient-boosting example calculates residuals, fits a new tree, and updates the ensemble.
- XGBoost Importance and Scope. Describe XGBoost's three feature-importance measures and distinguish their global interpretation from the desired interpretation of an individual compound's prediction.
- Model-agnostic interpretation benefits. Explain why model-agnostic interpretation is useful when comparing models and choosing a predictive method.
- Sensitivity analysis and derivative limits. Describe how descriptor perturbations reveal sensitivity, and distinguish the illustrated local finite-difference analysis from global transformations. Explain why finite-difference derivatives are fragile for tree ensembles.
- Shapley descriptor contributions. Explain why a descriptor’s Shapley value averages its marginal contributions across coalitions rather than measuring only the effect of removing it from the full descriptor set.
- SHAP explanations and computation. Describe how SHAP uses descriptor Shapley values for an additive explanation, and why computing such explanations calls for practical choices as descriptor counts grow.
What it builds on
- Molecular Descriptors (chapter 2) · Introduces molecular descriptors as representations of compounds.
- Linear Models (chapter 4) · The use of linear models to relate features to a response.
- Feature Generation & Exploration (chapter 3) · Representing molecules with computed features.
- Random Forests: A Strong Collective of Weak Decision Trees (chapter 5) · The earlier treatment of random forests as ensembles of decision trees.
- Gradient Descent (chapter 4) · The earlier introduction to gradient-based optimization.
- Decision Trees (chapter 5) · The earlier introduction to decision trees.
Section by section
Introduction
The introduction explains why small-molecule binding to RNA is therapeutically promising but difficult to predict, and frames QSAR as a way to relate molecular descriptors to measured binding.
- RNA Binding Therapeutic Potential: Explain why binding an RNA target with a small molecule can be useful in drug discovery, giving an outcome described in the section.
Why it matters: This explains why RNA binding is worth investigating despite the difficulty of designing selective compounds. - RNA Targeting Challenges: Explain why selective RNA binding and predictive modeling are difficult, connecting similar RNA motifs, conformational flexibility, and limited experimental binding data.
Why it matters: These constraints explain why RNA-binding models must contend with flexibility and limited measurements. - QSAR Structure Binding Relationship: Describe the relationship the proposed QSAR model seeks to establish between molecular descriptors and experimentally observed RNA-binding profiles, and what an interpretable model could reveal.
Why it matters: It states what the case study aims to connect and why interpreting the resulting model would be valuable.
Key terms: RNA target, small molecule binding, selectivity, conformational flexibility, RNA-ligand structures, QSAR, descriptors, binding profile, interpretability
6.1 Small Molecule Binding to an RNA Target
Introduces HIV-1 TAR RNA as a simplified model system for a QSAR case study and explains why targeting it matters.
- TAR as a model system: Explain what a model system is and why TAR RNA serves that role in this QSAR case study.
Why it matters: It clarifies what the TAR case study can help researchers investigate. - Importance of targeting TAR: Explain why disrupting TAR could matter medically and why the work may be relevant beyond HIV.
Why it matters: It connects the case study to a potential treatment goal and to other RNA targets.
Key terms: HIV-1 transactivation response (TAR), model system, QSAR workflow, TAR element, HIV viral replication, antiretroviral treatments, RNA structures
6.1.1 The HIV-1 Transactivation Response (TAR) RNA Model System
This section introduces HIV-1 TAR RNA as an antiviral target and describes the small, chemically varied ligand set used for modeling.
- TAR RNA antiviral rationale: Explain why HIV-1 TAR RNA is a potential antiviral target: the viral Tat protein must bind this RNA stem-loop to enable efficient transcription, so blocking their interaction is a validated strategy.
Why it matters: It explains why the case study investigates molecules that bind TAR RNA. - Small ligand dataset diversity: Describe the experimental ligand set used for modeling: it contains 48 compounds, including 29 known TAR ligands and 19 known RNA-targeted scaffolds, drawn from five categories to represent varied structures and binding characteristics.
Why it matters: The dataset’s small size and chemical variety frame the modeling choices in the case study.
Key terms: TAR RNA, Tat protein, TAR-Tat interaction, transcription, ligands, TAR-ligands, RNA-targeted scaffolds, binding characteristics
6.1.2 Structure: Computing Descriptors
This section expands a small RNA-binding ligand dataset into plausible protomer and conformer states, then combines their descriptors into one vector per ligand.
- Protomer States and pH: Explain what protomers are and why the workflow generates protonation and tautomerization states over a pH range suited to the binding experiment.
Why it matters: The ligand’s form under the experiment’s conditions affects which structures the descriptors represent. - Conformer Sampling and Geometry: Describe why a ligand needs multiple 3D conformers and how RDKit generates plausible starting geometries with distance geometry and ETKDG.
Why it matters: One connectivity pattern does not imply one 3D shape or one set of 3D descriptors. - Conformer Energy Selection: Trace how the example refines generated conformers and selects those used for descriptor calculation.
Why it matters: The energy filter keeps the descriptor calculation focused on less energetically strained sampled shapes. - Weighted Ligand Descriptors: Explain how descriptors from multiple states become one ligand vector and qualify the cross-protomer weighting used here.
Why it matters: The final modeling input summarizes many possible structures without claiming that their cross-state energy weights are rigorous. - Descriptor Precision and Cost: Justify when computing 3D descriptors across conformers may be worthwhile and why the section suggests trying 2D descriptors first.
Why it matters: Descriptor detail has a computational cost that should be weighed against its effect on modeling performance.
Key terms: protomer, protonation, tautomerization, pH range, conformer, distance geometry, ETKDG, distance bounds matrix, Universal Force Field (UFF), relative conformer energies
6.1.3 Activity: Experimentally Measuring Binding Profiles
This section explains how SPR measurements provide binding-activity responses for HIV-1 TAR ligands and why the range of those responses matters for model evaluation.
- RNA binding response variables: Explain what kon, koff, and KD say about a ligand binding to HIV-1 TAR, including how KD is calculated.
Why it matters: These are the binding measurements the molecular descriptors are used to predict. - SPR binding measurement: Trace how SPR detects ligand binding at the sensor surface and how its measurements yield association and dissociation rates.
Why it matters: SPR connects the observed signal to the binding rates used as model responses. - Binding range and representation: Justify why the broad range of measured binding profiles makes representative training and held-out test sets important.
Why it matters: A model tested on an unrepresentative slice may not be meaningfully evaluated across the binding profiles of interest.
Key terms: kon, koff, KD, binding kinetics, response variables, surface plasmon resonance (SPR), immobilized target, association, dissociation, refractive index
6.2 Representative Data Splitting & Dimensionality Reduction
This section explains why descriptor choices matter before modeling and why the case study uses the paper’s published descriptor values.
- Descriptor Feature Design: Explain why designing molecular features is an important part of the ML pipeline, including how aggregating descriptors across a ligand’s protomers or conformers can better reflect its real properties.
Why it matters: The features chosen for a model can require as much care as the modeling step. - Published Descriptor Values: Justify the choice to use the paper’s published descriptor values in this case study.
Why it matters: Using the same descriptor values makes the case study’s results easier to compare with the paper’s.
Key terms: feature design, descriptors, protomers, conformers, published descriptor values
6.2.1 Data Refinement
This section refines QSAR data by log-transforming binding responses and removing redundant descriptors, while identifying a potential source of data leakage.
- Log-Transformed Binding Responses: Explain why the binding response variables are log-transformed before QSAR modeling, including the effects on skew, variance, and relationships a linear model can capture.
Why it matters: The transformed responses are better suited to the linear QSAR model and its later statistical interpretation. - Redundant Descriptor Filtering: Describe which near-constant and highly correlated descriptors are removed and why redundancy is a problem for the model.
Why it matters: Filtering keeps largely repetitive inputs from complicating the QSAR model and its coefficients. - Correlated Descriptor Selection: Trace how the section chooses among highly correlated descriptors and explain why using the full dataset's response correlations risks data leakage.
Why it matters: The choice of which descriptor survives affects the model, and consulting responses outside the training set can leak information into that choice.
Key terms: log-transform, response variables, residuals, pKD, descriptors, Pearson correlation coefficient, multicollinearity, intercorrelation, response variable, training set
6.2.2 Representative Data Splitting: Kennard-Stone Algorithm
Explains why a small QSAR dataset needs a representative train/test split, how Kennard-Stone selects diverse training samples, and why high-dimensional descriptors complicate that selection.
- Representative split and applicability: Explain why the 48-ligand dataset needs a representative split and why training samples should span both the range and intermediate variation of binding activity.
Why it matters: What the training set covers affects which new ligands the model can reasonably predict. - Kennard-Stone sample selection: Trace how the Kennard-Stone algorithm selects a diverse training subset using distances between samples.
Why it matters: The selection rule explains why the chosen training samples can span more of the dataset than an unlucky random split. - High-dimensional distance limitations: Explain why distance-based selection becomes statistically less reliable with 48 ligands in a 191-dimensional descriptor space, and distinguish that issue from computational cost.
Why it matters: Distance calculations may be feasible even when too few samples make their results hard to interpret as similarity. - High-dimensional interpretation risks: Describe the additional difficulties high dimensionality creates for visualizing the split and for model generalization when dimensions are spurious.
Why it matters: Descriptor count affects not just sample selection but also how readily the split can be inspected and how well a model may generalize.
Key terms: representative data split, applicability domain, binding activity, interpolate, Kennard-Stone algorithm, KSA, pairwise distances, training subset, descriptor space, curse of dimensionality
6.2.3 Dimensionality Reduction: Principal Component Analysis
Explains how PCA reduces descriptor dimensionality, how to choose retained components, and why its use for split design differs from its use in a predictive pipeline.
- Dimensionality Reduction Rationale: Explain why adding data alone cannot readily overcome the curse of dimensionality and why focusing on informative dimensions can help.
Why it matters: This explains why the section reduces descriptors rather than simply seeking more data. - Principal Component Construction: Describe how PCA creates ranked, orthogonal components and projects samples onto a selected number of them.
Why it matters: The component ranking explains both what PCA preserves and what may be lost when components are dropped. - Retained Component Criteria: Explain the trade-off in choosing k and describe the three criteria presented for selecting it.
Why it matters: Choosing k determines whether the compressed data still represents the original data adequately. - PCA In Representative Splitting: Describe how the case study used PCA for the Kennard-Stone split and why its two-component plot does not represent the full retained dataset.
Why it matters: A readable two-dimensional visualization should not be mistaken for the data used to design the split. - Split Design Versus Model Leakage: Distinguish fitting PCA for split design from fitting it as part of a predictive modeling pipeline.
Why it matters: The distinction prevents a split-design choice from being copied into a predictive pipeline where it would cause leakage.
Key terms: curse of dimensionality, blessing of non-uniformity, principal components, orthogonal, variance, covariance matrix, eigenvectors, eigenvalues, k, eigenvalue criterion
6.3 QSAR Modeling: Mapping Descriptors to Measurements
This introduction previews the QSAR modeling workflow, its intended application, and a comparison with gradient boosting machines.
6.3.1 Exemplary QSAR Modeling Workflow
This section builds interpretable QSAR models for HIV TAR1 binding, explains how candidate models are selected, and cautions against overinterpreting results from a small dataset.
- Interpretable Model and Feature Selection: Explain why multiple linear regression (MLR) is a starting point for this QSAR task and why Lasso feature selection precedes it.
Why it matters: The choice of model and feature-selection step follows from both the scientific need for interpretation and the size of the dataset. - Candidate MLR Model Selection: Trace how the authors search for and choose an MLR model after feature selection, including the complexity limit and evaluation criteria.
Why it matters: A good training fit alone does not establish that a compact QSAR model predicts unseen ligands well. - Small Dataset Reliability Limits: Explain why the reported single-split QSAR performance should be treated cautiously and what stronger validation would report.
Why it matters: It prevents a promising score from being mistaken for dependable predictive evidence. - Williams Plot Applicability Domain: Describe what a Williams plot compares and how this section uses it to identify a ligand outside the model’s applicability domain.
Why it matters: An applicability-domain check flags points for which the fitted model may be less representative.
Key terms: QSAR, multiple linear regression, MLR, Lasso, descriptors, Topliss rule, R2, Q2, Occam’s razor, small-data regime
6.3.2 QSAR Model Interpretation
This section explains how QSAR models can inform compound design, why their interpretations need scrutiny, and how Hansch and Free-Wilson models approached structure–activity relationships.
- Descriptor-Based Activity Hypotheses: Explain how a QSAR equation can suggest which compound descriptors influence activity and how changing a compound might alter a predicted outcome. Distinguish an easily inspected descriptor from one that must be recalculated and is harder to explain.
Why it matters: Interpreting descriptors can turn a prediction into a testable idea for modifying a compound. - Regulatory Interpretation Boundaries: Describe the qualities the section identifies for regulatory use of a QSAR model and explain what limits the validity of its interpretations.
Why it matters: A regulator needs to understand both the basis of a prediction and where its interpretation is justified. - Statistical and Expert Validation: Explain why assessing QSAR performance statistically is not enough and how expert review can challenge an apparently important feature.
Why it matters: A model can fit its data while giving a misleading explanation for its predictions. - Hansch and Free-Wilson Approaches: Distinguish the historical Hansch approach from Free-Wilson analysis in terms of what each relates to biological activity.
Why it matters: These examples show two ways to make a quantitative structure–activity relationship interpretable.
Key terms: QSAR, descriptor, lnKoff, lnKon, a_base, a_nN, mechanistic interpretation, defined endpoint, unambiguous algorithm, applicability domain
6.4 Gradient Boosting Machines
This section previews a comparison of QSAR models with gradient boosting machines, followed by GBM theory and model-agnostic interpretation.
6.4.1 Informing the RNA-Binding Chemical Space
This section uses a gradient boosting classifier to distinguish RNA-binding from protein-binding small molecules and considers what its predictive features might reveal about RNA-binding chemical space.
- RNA Binding Descriptor Differences: Explain why a model that distinguishes RNA-binding from protein-binding compounds could help identify molecular descriptors associated with RNA-binding activity.
Why it matters: The classifier is useful not only for prediction but also for asking what distinguishes the two compound sets. - Gradient Boosting Comparison Results: Describe how the XGBClassifier was evaluated and how its reported performance compared with the published LASSO and neural network results.
Why it matters: The comparison shows that this gradient boosting model was a useful baseline for the RNA-versus-protein binding task without tuning.
Key terms: QSAR models, molecular descriptors, RNA-binding activity, protein-binding activity, GBM, XGBClassifier, Mordred, stratified 10-fold cross validation, AUPRC, LASSO
6.4.2 The XGBoost Magic Trick
This section explains how gradient boosting builds an ensemble, why it needs tuning, and what XGBoost feature-importance measures can and cannot reveal about RNA-binding predictions.
- Parallel and Sequential Ensembles: Distinguish random forests from boosting in how they train and combine models, and explain the different errors their ensembles primarily reduce.
Why it matters: This contrast explains why boosting learns from earlier mistakes rather than averaging independent predictions. - Residual Based Gradient Boosting: Trace how the decision-tree gradient-boosting example calculates residuals, fits a new tree, and updates the ensemble.
Why it matters: Residuals make the instruction to correct earlier errors concrete: they supply the next tree's training targets. - Boosting Tuning Tradeoffs: Explain why boosting can overfit or train slowly and describe the controls the section gives for those issues.
Why it matters: Model settings affect both how well the ensemble generalizes and how long it takes to train. - XGBoost Importance and Scope: Describe XGBoost's three feature-importance measures and distinguish their global interpretation from the desired interpretation of an individual compound's prediction.
Why it matters: A descriptor that matters across the model is not necessarily the descriptor driving one compound's prediction.
Key terms: random forests, boosting, parallel ensembles, sequential ensembles, weak learners, bias, gradient boosting, residuals, weak decision tree, loss function
6.4.3 Model-agnostic Interpretation
This section explains how sensitivity analysis and SHAP can interpret descriptor contributions to RNA-binding model predictions without restricting the choice of predictive model.
- Model-agnostic interpretation benefits: Explain why model-agnostic interpretation is useful when comparing models and choosing a predictive method.
Why it matters: You can compare what different models rely on without limiting yourself to models that are easy to interpret intrinsically. - Sensitivity analysis and derivative limits: Describe how descriptor perturbations reveal sensitivity, and distinguish the illustrated local finite-difference analysis from global transformations. Explain why finite-difference derivatives are fragile for tree ensembles.
Why it matters: A near-zero numerical derivative for a tree model may reflect its stepped predictions rather than an unimportant descriptor. - Shapley descriptor contributions: Explain why a descriptor’s Shapley value averages its marginal contributions across coalitions rather than measuring only the effect of removing it from the full descriptor set.
Why it matters: Considering different descriptor combinations avoids mistaking one context-dependent change for a descriptor’s overall contribution. - SHAP explanations and computation: Describe how SHAP uses descriptor Shapley values for an additive explanation, and why computing such explanations calls for practical choices as descriptor counts grow.
Why it matters: SHAP explanations are most useful when they are both interpretable and faithful to the model being explained. - Global and local SHAP plots: Distinguish what the section’s SHAP plots show about importance across compounds, an individual compound, and a descriptor’s relationship to predictions.
Why it matters: The plots answer different questions: what matters across compounds is not necessarily what drove one compound’s prediction.
Key terms: model-agnostic interpretability, feature importance, sensitivity analysis, global sensitivity analysis, local sensitivity analysis, partial derivative analysis, finite difference, tree ensembles, Shapley values, coalition
6.5 Summary
Where this is used later
- Challenges of Drug-Target Affinity Prediction (chapter 11) · Introduces experimental measurements of molecular activity.
- Dissimilarity-based Compound Selection (chapter 7) · The earlier Kennard-Stone example selected dissimilar compounds for representative data splitting.
- Dimensionality Reduction: Drug Repurposing (chapter 7) · Motivates dimensionality reduction as a method in its own right.
- (chapter intro) (chapter 7) · Introduces dimensionality reduction in the preceding chapter.
- Self-Organizing Maps (chapter 7) · Introduces reducing molecular data to fewer dimensions.
- Universal Manifold Approximation & Projection (UMAP) (chapter 7) · The earlier introduction to PCA as dimensionality reduction.
Concepts explained in this chapter
- model-agnostic methodModel-agnostic interpretability comprises methods that explain or analyze a model's predictions without relying on its internal implementation or architecture.
- Kennard-Stone algorithmThe Kennard-Stone algorithm selects a representative subset of samples by iteratively choosing points that maximize distance from already selected points.
- feature importance scoreFeature importance is a quantitative estimate of how much an input feature contributes to a predictive model's decisions.
- QSAR modelQSAR modeling predicts molecular properties or biological activities from quantitative relationships between chemical structure and measured outcomes.
- curse of dimensionalityThe curse of dimensionality is the phenomenon in which statistical estimation, geometric coverage, and related computational tasks become increasingly difficult as the number of dimensions grows.
- conformational spaceConformational space is the collection of accessible three-dimensional molecular arrangements arising from structural flexibility.
- protomer spaceProtomer space is the set of chemically relevant protonation-state variants considered for a ligand.
- principal component analysisPCA is principal component analysis, a linear dimensionality-reduction method that projects data onto orthogonal directions of greatest variance.
- SHAPSHAP is a model-interpretation method that attributes a prediction to input features using Shapley-value-based contributions.
- Shapley valueA Shapley value is a cooperative-game-theoretic attribution that assigns an average marginal contribution to an individual feature across coalitions.
- numerical differentiationThe finite difference method numerically approximates derivatives from function values at nearby input points.
- partial derivative analysisPartial derivative analysis estimates how infinitesimal changes in individual input descriptors affect a model's prediction.
- sensitivity analysisSensitivity analysis evaluates how changes in inputs or assumptions affect a model's outputs.
- surrogate modelA surrogate model is an inexpensive predictive model that approximates the performance or outputs of a more computationally expensive process.
- gradient boostingGradient boosting is an ensemble learning technique that sequentially adds weak models trained to reduce the current ensemble's loss or errors.
- dimensionality reductionDimensionality reduction is the transformation of data into fewer dimensions while retaining important information.
- 3D molecular descriptorA 3D molecular descriptor is a numerical representation of properties derived from a molecule's three-dimensional geometry.
- Boltzmann-weighted averagingBoltzmann weighting is the averaging of state-dependent values with weights determined by their relative energies and temperature.
- Dimorphite-DLDimorphite-DL is a software package that predicts chemically plausible protonation states of molecules over a specified pH range.
- ETKDGETKDG is a distance-geometry conformer-generation method that incorporates experimental torsion preferences and chemical knowledge constraints.
- Molecular Operating EnvironmentMOE is commercial molecular-modeling software for visualizing, analyzing, and manipulating molecular structures.
- Universal Force FieldThe Universal Force Field is a general-purpose molecular mechanics force field parameterized to model structures and energies across broad classes of elements and compounds.
- distance bounds matrixA distance bounds matrix stores lower and upper allowable distance bounds for pairs of atoms in a molecular structure.
- distance geometryDistance geometry is a method for generating molecular three-dimensional coordinates from constraints on distances between atoms.
- intermolecular force fieldA molecular force field is a parameterized mathematical model of atomic interactions and their contribution to molecular potential energy.
- protomerA protomer is a molecular form distinguished by its protonation state or by protonation-linked tautomeric structure.
- protonationProtonation is the addition of a proton to a molecule or ionizable chemical group.
- tautomerismTautomerization is the process by which tautomers interconvert through proton transfer and bond rearrangement.
- molecular descriptorA molecular descriptor is a numerical or symbolic representation of a molecule's physicochemical, structural, topological, or quantum-mechanical properties.
- boostingBoosting is an ensemble technique that sequentially trains models so later models correct errors made by earlier models.
- weak learnerA weak learner is a predictive model with limited individual accuracy that can contribute to a stronger ensemble when combined with other learners.
- covariance matrixA covariance matrix is a square matrix whose entries quantify pairwise covariance between features.
- eigenvalueAn eigenvalue is the scalar associated with an eigenvector that describes the factor by which a linear transformation scales that vector.
- eigenvectorAn eigenvector is a nonzero vector whose direction is preserved by a linear transformation up to multiplication by a scalar eigenvalue.
- explained varianceExplained variance is the proportion of variability in data accounted for by a representation or statistical model.
- principal componentA principal component is a transformed feature formed as a linear combination of original features that captures a successive amount of data variance.
- L1 regularizationL1 regularization penalizes the sum of absolute model weights, encouraging some weights to become exactly zero.
- XGBoostXGBoost is a gradient-boosting implementation designed for efficient and high-performing supervised learning.
- code coverageCoverage is a tree-ensemble feature-importance measure based on the average number of samples affected by splits using a feature.
- frequencyFrequency is a tree-ensemble feature-importance measure equal to the number of split operations using a feature across the ensemble.
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.
