Chapter 7: Unsupervised Learning: Repurposing Drugs, Curating Compounds, & Screening Fragments
The chapter introduces unsupervised learning as a way to find structure in unlabeled drug-discovery data and explains the basic purposes of dimensionality reduction, clustering, and density estimation.
After this chapter you will be able to
- Learning from Unlabeled Data. Distinguish unsupervised learning from supervised learning and explain why it is useful when life-science experiments leave much data unlabeled.
- Reduction and Clustering Purposes. Describe what dimensionality reduction and clustering each aim to do with data.
- Density Estimation Purpose. Explain what density estimation seeks to represent and how low-probability regions can help flag anomalies.
- COVID Moonshot screening dataset. Describe the COVID Moonshot campaign’s aim and the subset of compounds used here, including what IC50 measures.
- Fingerprints and optional activity labels. Explain how the section prepares compound representations and activity labels, and distinguish those labels from a requirement of unsupervised learning.
- Single-Shot Screening and Follow-Up. Describe how automated high-throughput screening tests a large compound library at one relatively high concentration and flags compounds with significant inhibition as potential hits.
- Dose-Response Curves and IC50. Explain what a dose-response curve plots and how IC50 describes a compound’s response across concentrations.
- Restrained Hill Curve Fitting. Describe, at a high level, how a restrained Hill equation is fitted to dose-response observations and distinguish the set inhibitor concentrations from the recovered parameters.
- HTS Data and Dimensionality Reduction. Explain why noisy, chemically diverse, high-dimensional HTS results challenge traditional QSAR pattern-finding and why dimensionality reduction is useful.
- Self-organizing map training. Trace how a self-organizing map places high-dimensional fingerprints on a two-dimensional grid: find each input’s best matching unit, then adjust that node and its neighbors toward the input over repeated iterations.
- Activity patterns and inference. Explain how the trained SOM’s nodes and optional activity labels can guide compound investigation and inference for an unseen compound.
- SOM Hit Designation. Trace how the trained SOM is used to designate Drug Repurposing Hub compounds as hits.
- Retrospective Recall Versus Precision. Explain what the reported SOM result establishes and why it does not establish good precision.
- SOM Strategy Limitations. Describe the proposed reasons the SOM flags so many drugs and the methodological assumptions that weaken its interpretation.
- Manifolds and nonlinear structure. Explain why a nonlinear manifold-learning method can reveal relationships that PCA or a grid-based SOM may miss.
- UMAP neighborhood construction. Trace, at the level described here, how UMAP uses high-dimensional similarities to arrange points in a low-dimensional projection.
- Neighborhood and spacing settings. Distinguish how n_neighbors and min_dist affect a UMAP projection of compounds.
- Chemical projection interpretation limits. Describe what the COVID Moonshot UMAP plots suggest and what cannot be concluded from their distances or proximities.
- Diversity and focus tradeoff. Distinguish diverse from focused compound selection, including when each is useful and what information it can provide.
- Ways to measure diversity. Explain why chemical diversity depends on the chosen representation and measure, and how that choice affects compound sampling.
- Exhaustive subset selection limit. Describe the proposed exhaustive way to find a maximally diverse library and why it is impractical for the example size.
- Combinatorial Library Construction. Describe how a combinatorial library is built from predefined fragments, attachment sites, and reaction rules, then filtered or prioritized.
- Retrosynthetic Library Tradeoff. Explain why retrosynthetic library design is offered as an alternative to combinatorial enumeration and what BRICS does not guarantee.
- Cluster-based selection workflow. Explain how clustering supports selection of a diverse compound library and how representation, distance, clustering method, and within-cluster sampling affect that selection.
- Agglomerative clusters and dendrogram cuts. Describe how agglomerative clustering builds a hierarchy, how linkage determines which clusters merge, and how a dendrogram cut defines the final clusters.
- Within-cluster sampling tradeoffs. Distinguish equal from size-proportional sampling across clusters, and explain why the example selects compounds near each cluster center.
- Taylor-Butina clustering and cutoff. Trace how Taylor-Butina forms clusters from Tanimoto-distance neighbors, and explain how its distance cutoff affects the result.
- Sample diversity and redundancy. Describe how the two sampled libraries are compared for diversity and redundancy, and state the reported outcome.
- Dissimilarity-Based Selection Process. Trace how dissimilarity-based compound selection builds a subset directly from pairwise compound comparisons, without first clustering the compounds.
- MaxSum and MaxMin Selection. Distinguish how MaxSum and MaxMin choose the next compound, and relate MaxMin to the earlier Kennard-Stone implementation.
- Fragment-Based Discovery Workflow. Explain how fragment-based drug discovery differs from traditional compound-library screening: it finds multiple target-binding fragments, then expands, merges, and optimizes them into leads with stronger binding profiles.
- Small Fragment Screening Advantage. Explain why a fragment library can be much smaller than a compound screening library yet still probe a protein and chemical space effectively.
- Fragment Hit Development Strategies. Distinguish fragment growing, linking, and merging by the starting fragments and the structural change each makes.
- Limits of Fragment-Derived Leads. Explain why growing, linking, or merging fragment hits does not by itself produce a suitable lead, including the limitation of binding observations from static structures.
- Pharmacophore Features and Geometry. Explain why a pharmacophore specifies both interaction-capable functional groups and their spatial arrangement for activity at a target.
- Pharmacophores in Fragment Discovery. Explain the connection between pharmacophores and fragment expansion or merging, and describe what decomposing screened fragment hits yields.
- Density Estimation and KDE. Explain what density estimation models and how kernel density estimation differs from a histogram.
- Kernel and Bandwidth Choices. Explain how the kernel function and bandwidth affect a KDE estimate, including the consequences of bandwidths that are too small or too large.
- Fragment Pharmacophore Distance Models. Trace how the bound-fragment ensemble becomes fitted pharmacophore-pair distance distributions.
- Unseen Molecule Scoring. Describe how fitted pharmacophore-distance densities are used to score a previously unseen molecule and what a higher score means.
What it builds on
- QSAR Modeling: Mapping Descriptors to Measurements (chapter 6) · Provides a contrast with QSAR modeling that maps descriptors to measurements.
- Dimensionality Reduction: Principal Component Analysis (chapter 6) · Introduces dimensionality reduction in the preceding chapter.
- Hashed Fingerprints (chapter 2) · Introduces hashed molecular fingerprints.
- Molecular Descriptors (chapter 2) · Introduces molecular descriptors as ways to characterize compounds.
- Loading a Virtual Screening Library (chapter 2) · Introduces working with an existing compound library.
- Fingerprints: Representing Molecules as Numbers (chapter 2) · Molecular fingerprints as compound representations.
- Defining “Similarity” (chapter 2) · Assessing similarity between compounds.
- Representative Data Splitting: Kennard-Stone Algorithm (chapter 6) · The earlier Kennard-Stone example selected dissimilar compounds for representative data splitting.
Section by section
Introduction
The chapter introduces unsupervised learning as a way to find structure in unlabeled drug-discovery data and explains the basic purposes of dimensionality reduction, clustering, and density estimation.
- Learning from Unlabeled Data: Distinguish unsupervised learning from supervised learning and explain why it is useful when life-science experiments leave much data unlabeled.
Why it matters: This distinction explains why the chapter’s methods can work with data that lacks measured outcomes. - Reduction and Clustering Purposes: Describe what dimensionality reduction and clustering each aim to do with data.
Why it matters: The two methods serve different purposes: simplifying a representation and grouping similar instances. - Density Estimation Purpose: Explain what density estimation seeks to represent and how low-probability regions can help flag anomalies.
Why it matters: It gives a way to reason about which observations fit the data and which may be unusual.
Key terms: unsupervised learning, supervised learning, labels, inputs, dimensionality reduction, clustering, cluster, density estimation, probability density function, anomaly detection
7.1 Dimensionality Reduction: Drug Repurposing
Introduces the COVID Moonshot screening data, its optional activity labels, and the high-dimensionality problem that motivates visualization.
- COVID Moonshot screening dataset: Describe the COVID Moonshot campaign’s aim and the subset of compounds used here, including what IC50 measures.
Why it matters: These measurements establish what the compounds and activity values in the chapter represent. - Fingerprints and optional activity labels: Explain how the section prepares compound representations and activity labels, and distinguish those labels from a requirement of unsupervised learning.
Why it matters: It separates the compound features used for analysis from activity labels that are available but not necessary for unsupervised learning. - High-dimensional data challenge: Explain why high-dimensional screening data can be difficult to analyze.
Why it matters: This explains the problem motivating the forthcoming dimensionality-reduction visualizations without assuming how those methods work.
Key terms: COVID Moonshot, Mpro, HTS, IC50, Morgan fingerprints, active and inactive compounds, unsupervised learning, high dimensionality, curse of dimensionality
7.1.1 High Throughput Screening Data
This section explains how high-throughput screening measures compound activity, derives IC50 from dose-response data, and motivates dimensionality reduction for noisy, diverse screening results.
- Single-Shot Screening and Follow-Up: Describe how automated high-throughput screening tests a large compound library at one relatively high concentration and flags compounds with significant inhibition as potential hits.
Why it matters: This explains what the first-pass screening result can—and cannot—tell you about a compound. - Dose-Response Curves and IC50: Explain what a dose-response curve plots and how IC50 describes a compound’s response across concentrations.
Why it matters: A concentration series provides more information about activity than a single-shot measurement. - Restrained Hill Curve Fitting: Describe, at a high level, how a restrained Hill equation is fitted to dose-response observations and distinguish the set inhibitor concentrations from the recovered parameters.
Why it matters: Fitting turns measured responses into an IC50 while keeping the estimated parameters plausible. - HTS Data and Dimensionality Reduction: Explain why noisy, chemically diverse, high-dimensional HTS results challenge traditional QSAR pattern-finding and why dimensionality reduction is useful.
Why it matters: It explains why organizing screening data is valuable before looking for structure–activity patterns.
Key terms: HTS, single-shot inhibition data, potential hits, dose response curves, IC50, E𝑚𝘢𝑥, Hill equation, Levenberg-Marquardt algorithm, restrained fit, Hill coefficient
7.1.2 Self-Organizing Maps
The section explains how a self-organizing map projects compound fingerprints onto a two-dimensional grid, how to evaluate the map, and how to interpret its activity patterns and clusters.
- Self-organizing map training: Trace how a self-organizing map places high-dimensional fingerprints on a two-dimensional grid: find each input’s best matching unit, then adjust that node and its neighbors toward the input over repeated iterations.
Why it matters: This explains why a two-dimensional map can reveal relationships among compounds represented by 2048-dimensional fingerprints. - Map evaluation and selection: Distinguish topographic error from quantization error, and describe how the section uses held-out compounds and parameter comparisons to select a SOM.
Why it matters: A visually appealing map is more useful when its organization is checked against data held out from training. - Activity patterns and inference: Explain how the trained SOM’s nodes and optional activity labels can guide compound investigation and inference for an unseen compound.
Why it matters: The map can help prioritize possible hits and highlight places where similar compounds have different measured activities. - Unified Distance Matrix interpretation: Describe what the Unified Distance Matrix displays and how its colors and overlaid connected components help identify clusters across nodes.
Why it matters: The UMAT helps identify groups that extend beyond compounds assigned to a single node.
Key terms: self-organizing map, SOM, best matching unit, BMU, neighborhood function, topographic error, quantization error, held-out, optional labels, active
7.1.3 Drug Repurposing
The section uses a self-organizing map to flag existing drugs for possible repurposing, then explains why recovering known hits does not establish reliable prediction.
- Drug Repurposing Rationale: Explain what drug repurposing seeks to do and why existing drugs can offer an advantage over developing new ones.
Why it matters: It explains why the section looks for new uses among drugs that have already been studied. - SOM Hit Designation: Trace how the trained SOM is used to designate Drug Repurposing Hub compounds as hits.
Why it matters: This is the specific rule behind the section’s proposed repurposing hits. - Retrospective Recall Versus Precision: Explain what the reported SOM result establishes and why it does not establish good precision.
Why it matters: Finding every known hit is much less persuasive when nearly every candidate is flagged. - Unsupervised Evaluation Boundaries: Distinguish the section’s four ways labels can enter an unsupervised workflow’s exploration or evaluation, and identify the SOM result as retrospective enrichment rather than supervised predictive performance.
Why it matters: Naming how labels were used prevents an exploratory result from being overstated as predictive validation. - SOM Strategy Limitations: Describe the proposed reasons the SOM flags so many drugs and the methodological assumptions that weaken its interpretation.
Why it matters: These limitations explain why the many flagged hits should remain candidates for study rather than strong activity predictions.
Key terms: drug repurposing, medical indication, safety profiles, SOM, node, plurality label, active, Drug Repurposing Hub, retrospective hits, recall
7.1.4 Universal Manifold Approximation & Projection (UMAP)
This section explains how UMAP produces a low-dimensional view of chemical data, how its settings affect that view, and how to interpret the resulting plot cautiously.
- Manifolds and nonlinear structure: Explain why a nonlinear manifold-learning method can reveal relationships that PCA or a grid-based SOM may miss.
Why it matters: It explains why a useful map of compounds need not come from flattening the data onto a plane or arranging it on a fixed grid. - UMAP neighborhood construction: Trace, at the level described here, how UMAP uses high-dimensional similarities to arrange points in a low-dimensional projection.
Why it matters: The construction process explains why nearby points, rather than exact distances across the whole plot, are the most useful feature of a UMAP view. - Neighborhood and spacing settings: Distinguish how n_neighbors and min_dist affect a UMAP projection of compounds.
Why it matters: These settings change which patterns are easy to see, so the appearance of clusters depends partly on the projection choices. - Projection selection and variability: Describe how the section selects UMAP settings for the COVID Moonshot example and why inspecting multiple runs is useful.
Why it matters: A single attractive plot is not, by itself, a sound way to choose settings for a particular task. - Chemical projection interpretation limits: Describe what the COVID Moonshot UMAP plots suggest and what cannot be concluded from their distances or proximities.
Why it matters: It keeps a useful visualization from being mistaken for experimental evidence about what a compound does.
Key terms: manifold, manifold assumption, PCA, SOM, nonlinear dimensionality reduction, UMAP, high-dimensional similarity scores, low-dimensional graph, local structure, global structure
7.2 Clustering: Curating Diverse Compound Libraries
The section connects earlier similarity-based screening to the upcoming discussion of similarity and diversity in compound-library design.
7.2.1 Diversity and Focus
This section explains how compound-library design balances broad chemical diversity against focused exploration, and why selecting the most diverse subset is difficult.
- Diversity and focus tradeoff: Distinguish diverse from focused compound selection, including when each is useful and what information it can provide.
Why it matters: The choice determines whether limited testing resources yield broad coverage or deeper insight into a promising region. - Ways to measure diversity: Explain why chemical diversity depends on the chosen representation and measure, and how that choice affects compound sampling.
Why it matters: A library can look more or less diverse depending on what molecular features the selection method compares. - Exhaustive subset selection limit: Describe the proposed exhaustive way to find a maximally diverse library and why it is impractical for the example size.
Why it matters: The vast number of possible subsets motivates practical alternatives to checking every library.
Key terms: diversity, focus, chemical space, chemical series, structure-activity relationship, hit rate, internal diversity, external data sets, molecular descriptors, scaffolds
7.2.2 Combinatorial Libraries
This section explains how to construct synthetic compound libraries by recombining molecular fragments, and why retrosynthetically guided recombination can be preferable to unrestricted enumeration.
- Combinatorial Library Construction: Describe how a combinatorial library is built from predefined fragments, attachment sites, and reaction rules, then filtered or prioritized.
Why it matters: This explains where a compound library can come from before selecting a diverse subset. - Scaffold Decoration and Hopping: Distinguish varying R-groups on a retained scaffold from replacing the scaffold, including why a replacement may be sought.
Why it matters: The distinction clarifies what is being varied in a library and why predictions may be less reliable after a scaffold change. - Bemis–Murcko Scaffold Abstractions: Explain how the section obtains an atomic framework and how a graph framework further generalizes it.
Why it matters: The chosen scaffold abstraction affects how many distinct building blocks the library appears to contain. - SMIRKS Fragment Joining: Trace what the section’s SMIRKS zip reaction joins and what happens to its dummy atoms.
Why it matters: It shows how the illustrated combinatorial library recombines fragments into new molecules. - Retrosynthetic Library Tradeoff: Explain why retrosynthetic library design is offered as an alternative to combinatorial enumeration and what BRICS does not guarantee.
Why it matters: A library is more useful when its proposed compounds have a better chance of being made.
Key terms: combinatorial library design, molecular fragments, attachment sites, combinatorial enumeration, scaffold, R-groups, scaffold decoration, scaffold hopping, chemical series, Bemis-Murcko (BM) scaffolds
7.2.3 Cluster-based Compound Selection
This section shows how clustering can select a diverse subset of compounds, then compares agglomerative and Taylor-Butina selection.
- Cluster-based selection workflow: Explain how clustering supports selection of a diverse compound library and how representation, distance, clustering method, and within-cluster sampling affect that selection.
Why it matters: The compounds selected depend on how similarity is represented and how the resulting clusters are sampled. - Agglomerative clusters and dendrogram cuts: Describe how agglomerative clustering builds a hierarchy, how linkage determines which clusters merge, and how a dendrogram cut defines the final clusters.
Why it matters: The linkage rule and where the hierarchy is cut determine which compounds can be sampled together. - Within-cluster sampling tradeoffs: Distinguish equal from size-proportional sampling across clusters, and explain why the example selects compounds near each cluster center.
Why it matters: A sampling rule must match whether the library is meant to emphasize diversity or reflect the starting population. - Clustering scale and representation: Explain why large fingerprint libraries motivate an alternative to agglomerative clustering and why ordinary k-means is not the chosen alternative.
Why it matters: A clustering method must fit both the scale of the library and the molecular representation. - Taylor-Butina clustering and cutoff: Trace how Taylor-Butina forms clusters from Tanimoto-distance neighbors, and explain how its distance cutoff affects the result.
Why it matters: The cutoff turns a desired degree of molecular similarity into the groups from which compounds are selected. - Sample diversity and redundancy: Describe how the two sampled libraries are compared for diversity and redundancy, and state the reported outcome.
Why it matters: These measures check whether the sampled library actually achieves the intended diversity rather than judging clusters by appearance alone.
Key terms: cluster-based compound selection, Morgan fingerprints, Tanimoto similarity coefficient, agglomerative clustering, intercluster distance, single linkage, complete linkage, average linkage, Ward’s method, equal sampling
7.2.4 Dissimilarity-based Compound Selection
This section explains how dissimilarity-based methods select diverse compound subsets without clustering, and compares two ways to choose each new compound.
- Dissimilarity-Based Selection Process: Trace how dissimilarity-based compound selection builds a subset directly from pairwise compound comparisons, without first clustering the compounds.
Why it matters: This gives you a way to select a diverse library subset without making clusters first. - Dissimilarity Measure Choices: Describe the dissimilarity measures illustrated in the section and why the choice of fingerprint and comparison metric depends on the use case.
Why it matters: What counts as dissimilar depends on how compounds are represented and compared. - MaxSum and MaxMin Selection: Distinguish how MaxSum and MaxMin choose the next compound, and relate MaxMin to the earlier Kennard-Stone implementation.
Why it matters: The two rules favor different notions of how a new compound differs from the subset already chosen.
Key terms: dissimilarity-based compound selection, pairwise dissimilarity, seed compound, stopping criterion, molecular descriptors, Euclidean distance, Mahalanobis distance, Tanimoto coefficients, molecular fingerprints, extended-connectivity fingerprints
7.3 Density Estimation: Fragment-based Drug Discovery
Introduces fragment-based drug discovery as an alternative to screening whole compounds, from finding binding fragments to developing lead compounds.
- Fragment-Based Discovery Workflow: Explain how fragment-based drug discovery differs from traditional compound-library screening: it finds multiple target-binding fragments, then expands, merges, and optimizes them into leads with stronger binding profiles.
Why it matters: This distinction explains what the fragment-based approach contributes to lead discovery.
Key terms: fragment-based drug discovery (FBDD), fragment library, binding affinity, lead compounds
7.3.1 Fragment-based Drug Discovery
This section explains how fragment libraries are defined, why small fragments are useful for screening, how fragment hits can be developed into compounds, and why those compounds still need validation.
- Fragment Library Rule of 3: State the Rule of 3 criteria used to qualify molecules for a fragment library, and distinguish these guidelines from rules for drug-like compounds.
Why it matters: These criteria explain what kinds of molecules enter a fragment screen. - Small Fragment Screening Advantage: Explain why a fragment library can be much smaller than a compound screening library yet still probe a protein and chemical space effectively.
Why it matters: It explains why screening fewer, simpler molecules can be a useful discovery strategy. - Fragment Hit Development Strategies: Distinguish fragment growing, linking, and merging by the starting fragments and the structural change each makes.
Why it matters: The three approaches turn weak fragment hits into candidates for stronger-binding leads. - Limits of Fragment-Derived Leads: Explain why growing, linking, or merging fragment hits does not by itself produce a suitable lead, including the limitation of binding observations from static structures.
Why it matters: A constructed compound is only a candidate until its binding profile proves better than the fragments’.
Key terms: Rule of 3, H bond donors, H bond acceptors, CLogP, fragment libraries, binding sites, hit rate, chemical space, fragment growing, R-groups
7.3.2 Pharmacophore Modeling
This section explains how pharmacophores capture the interaction features and geometry needed for binding, and how they can guide screening and fragment-based drug discovery.
- Pharmacophore Features and Geometry: Explain why a pharmacophore specifies both interaction-capable functional groups and their spatial arrangement for activity at a target.
Why it matters: The right chemical groups alone are not enough if they cannot make the needed contacts in the right positions. - Ligand Based Pharmacophore Modeling: Describe how the section constructs and checks a ligand-based pharmacophore from a set of ligands.
Why it matters: Shared features are more useful for screening when they are identified across ligands and checked against separate data. - Three Dimensional Conformer Effects: Explain how a compound’s conformations affect its chance of matching a desired pharmacophore.
Why it matters: A molecule that contains every desired feature may still rarely arrange them for binding. - Pharmacophores in Fragment Discovery: Explain the connection between pharmacophores and fragment expansion or merging, and describe what decomposing screened fragment hits yields.
Why it matters: Looking across fragment hits provides a target-related feature distribution rather than only isolated candidates for merging. - Crystal Structure Contact Caution: Explain why a single drug-protein crystal structure can mislead pharmacophore feature selection and how the section says simulations can help.
Why it matters: A contact seen in one structure is not necessarily a reliable binding feature.
Key terms: pharmacophore, functional groups, geometry, drug-target complementarity, ligand-based pharmacophore modeling, conformers, pharmacophore features, external test set, three-dimensional space, conformational flexibility
7.3.3 Density Estimation
The section explains kernel density estimation and applies it to distances between pharmacophore features in bound fragments to score unseen molecules.
- Density Estimation and KDE: Explain what density estimation models and how kernel density estimation differs from a histogram.
Why it matters: This is the basis for turning observed fragment features into a distribution against which new molecules can be compared. - Kernel and Bandwidth Choices: Explain how the kernel function and bandwidth affect a KDE estimate, including the consequences of bandwidths that are too small or too large.
Why it matters: An unsuitable bandwidth can make a pharmacophore-distance distribution misleadingly noisy or smooth. - Fragment Pharmacophore Distance Models: Trace how the bound-fragment ensemble becomes fitted pharmacophore-pair distance distributions.
Why it matters: These measured distances supply the observations that the fitted models use to score other molecules. - Unseen Molecule Scoring: Describe how fitted pharmacophore-distance densities are used to score a previously unseen molecule and what a higher score means.
Why it matters: This turns distributions learned from fragment hits into a way to prioritize molecules for continued investigation.
Key terms: density estimation, kernel density estimation (KDE), histogram, kernel function, bandwidth, overfitting, underfitting, cross-validation, log-likelihood, aligned conformations
7.4 Summary
Concepts explained in this chapter
- clusteringClustering is an unsupervised learning technique that partitions data instances into groups of similar instances.
- pharmacophore modelingLigand-based pharmacophore modeling derives common interaction features and geometry by comparing conformations of compounds with related activity.
- performanceBig O notation describes an algorithm's asymptotic growth in time or memory as input size or dimensionality increases.
- K-means clusteringK-means clustering partitions data into a specified number of groups by iteratively assigning points to nearby centroids and recomputing the centroids.
- Taylor-Butina clusteringTaylor-Butina clustering is a fingerprint-based clustering method that selects high-neighbor compounds as centroids and assigns nearby compounds using an exclusion threshold.
- Ward's methodWard's method is an agglomerative linkage criterion that merges the pair of clusters producing the smallest increase in within-cluster variance.
- agglomerative clusteringAgglomerative clustering is a bottom-up hierarchical clustering method that starts with singleton clusters and successively merges the closest clusters.
- centroidA centroid is the arithmetic mean position of the points assigned to a cluster.
- dendrogramA dendrogram is a tree diagram that represents the hierarchy and merge or split levels produced by hierarchical clustering.
- exclusion sphereAn exclusion sphere is the neighborhood around a selected cluster centroid whose flagged members cannot serve as members of another cluster or as another centroid.
- hierarchical clusteringHierarchical clustering is a clustering technique that builds a hierarchy of nested groups by iteratively merging or splitting clusters.
- inter-cluster distanceIntercluster distance is a rule for measuring separation between two clusters from the distances among their members.
- prototype-based clusteringPrototype-based clustering represents each cluster by a prototype and assigns instances according to their proximity to those prototypes.
- redundancy ratioA redundancy ratio is the fraction of unique compound pairs whose similarity exceeds a specified threshold.
- space complexityMemory complexity is an asymptotic measure of the memory required by an algorithm as input size grows.
- density estimatorDensity estimation is an unsupervised technique that estimates the probability distribution underlying sampled data.
- pharmacophoric featureA pharmacophoric or pharmacophore feature is a molecular interaction capability, such as hydrogen-bond donation, hydrogen-bond acceptance, hydrophobicity, or aromaticity, represented in a pharmacophore.
- self-organizing mapA self-organizing map is an unsupervised neural-network technique that maps high-dimensional data onto a lower-dimensional grid while preserving neighborhood relationships.
- probability densityA probability density function describes the relative likelihood density of a continuous random variable across possible values.
- drug repurposingDrug repurposing is the identification of new therapeutic uses for an existing drug outside its original medical indication.
- bandwidth parameterBandwidth is the scale parameter controlling the spatial range and smoothing of a kernel density estimate.
- kernel density estimateKernel density estimation is a nonparametric method that estimates a continuous probability density by summing kernel functions centered on observed data points.
- kernel functionA kernel function is a smooth, symmetric, non-negative weighting function that determines how an observation contributes to an estimated density at a location.
- UMAPUniform manifold approximation and projection is a nonlinear dimensionality-reduction technique for embedding high-dimensional data in a lower-dimensional space.
- global structureGlobal structure is the overall arrangement and interrelationships among points or clusters across an entire data representation.
- jaccard distanceJaccard distance is a dissimilarity measure equal to one minus the ratio of set intersection size to set union size.
- local structureLocal structure is the pattern of relationships among data points that are near one another in a representation.
- manifoldA manifold is a lower-dimensional mathematical space embedded in a higher-dimensional space that locally resembles ordinary Euclidean space.
- manifold learningManifold learning is a family of dimensionality-reduction methods that models high-dimensional data as lying on or near a lower-dimensional manifold.
- molecular fragmentA molecular fragment is a small chemical structure that can bind a biological target and serve as a starting point or building block for a larger compound.
- Gaussian neighborhood functionA Gaussian neighborhood function assigns larger update weights to nodes nearer the best matching unit and smaller weights to nodes farther away according to a Gaussian profile.
- best matching unitA best matching unit is the node in a self-organizing map whose weight vector is closest to a given input vector.
- neighborhood functionA neighborhood function assigns the magnitude of a node update according to the node's distance from the best matching unit.
- quantization errorQuantization error is the average distance between input data points and their best matching units in a self-organizing map.
- topographic errorTopographic error is the proportion of data points whose first and second best matching units are not adjacent in a self-organizing map.
- unified distance matrixA unified distance matrix is a self-organizing-map visualization that represents the average distances between each node and its neighboring nodes.
- combinatorial library designCombinatorial library design is the systematic construction of many compounds by combining predefined molecular fragments under specified chemical constraints.
- retrosynthetic library designRetrosynthetic library design constructs compound libraries by fragmenting and recombining molecules according to rules motivated by feasible chemical synthesis.
- fragment-based drug designFragment-based drug discovery, or fragment-based design, identifies small molecular fragments that bind a target and elaborates or links them into lead compounds.
- BRICSBRICS is an RDKit-supported method and module that fragments molecules and recombines the fragments using retrosynthetically motivated chemical rules.
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.
