Applicability Domain

The foundational metric for defining an AD, measuring the Tanimoto distance or Euclidean distance between a query molecule's fingerprint and the training set's centroid or nearest neighbors.

• Key methods: Morgan fingerprints (ECFP4), MACCS keys, or graph kernels.
• Thresholds: A common heuristic is a Tanimoto similarity > 0.7 to the nearest training compound.
• Limitation: Structural similarity alone fails to capture activity cliffs, where nearly identical molecules exhibit drastically different biological potency.

The AD is defined by the convex hull or bounding box of the training data's physicochemical descriptor space. A prediction is reliable only if the query molecule's descriptors fall within the observed ranges.

• Critical descriptors: Molecular weight, LogP, topological polar surface area (TPSA), and number of rotatable bonds.
• Leverage approach: The Williams plot visualizes standardized residuals versus hat values (leverage) to identify high-leverage extrapolations and potential outliers.
• Extrapolation risk: Querying beyond the descriptor bounds constitutes an extrapolation, where model behavior is undefined and often dangerously overconfident.

A probabilistic measure of how far a query instance is from the model's training distribution in the latent or feature space, often using Mahalanobis distance or standard deviation of ensemble predictions.

• Ensemble variance: High variance among predictions from an ensemble of models indicates the query is in a region of high epistemic uncertainty, signaling it is outside the AD.
• One-class classification: Models like One-Class SVMs or Isolation Forests can be trained solely on the training data to learn a decision boundary that envelops the AD, rejecting foreign samples.
• Gaussian Process confidence: The predictive variance of a Gaussian Process naturally increases in regions far from training data, providing a mathematically rigorous AD boundary.

The AD can be modeled by estimating the probability density function (PDF) of the training data in chemical space. A query molecule is considered out-of-domain if its likelihood falls below a calibrated threshold.

• Kernel Density Estimation (KDE): A non-parametric method that places a Gaussian kernel on each training point to create a smooth, continuous density landscape.
• Gaussian Mixture Models (GMM): Captures multi-modal distributions, representing distinct chemical series or clusters within the training set.
• Deep generative models: Variational autoencoders (VAEs) learn a continuous latent representation of the training molecules; the reconstruction error or latent space likelihood serves as a powerful AD metric.

A rigorous, distribution-free framework that provides a finite-sample, model-agnostic validity guarantee for the AD. Instead of a binary in/out decision, it produces prediction sets with a user-specified error rate (α).

• Inductive conformal prediction: Uses a held-out calibration set to compute nonconformity scores, defining a threshold that guarantees marginal coverage (e.g., 95%).
• Mondrian conformal prediction: Extends the framework to provide conditional validity across different classes or molecular scaffolds, preventing biased coverage.
• Credibility and confidence: The framework outputs a p-value for each possible label, where a low p-value for the predicted class indicates the molecule is likely outside the model's reliable domain.

A domain-specific filter that validates whether a query molecule's predicted property is physically or biologically plausible, acting as a sanity check beyond statistical metrics.

• Lipinski's Rule of Five: A heuristic filter for oral bioavailability; a model predicting high oral absorption for a molecule violating multiple rules should be flagged.
• PAINS and ALARM NMR filters: Identifies pan-assay interference compounds or reactive entities; a potent bioactivity prediction for a known PAINS compound is likely a false positive.
• Metabolic logic: A predicted site of metabolism (SOM) that contradicts established cytochrome P450 reactivity rules indicates the query is outside the model's reliable mechanistic domain.

A computational modeling method that establishes a mathematical relationship between a molecule's structural features and its biological activity or chemical property. The applicability domain is a critical component of QSAR validation, as defined by OECD Principle 3, which mandates that a model's predictive space be explicitly characterized.

• Maps molecular descriptors to endpoints
• Relies on the assumption that similar structures have similar properties
• Domain definition prevents extrapolation to novel chemotypes

A technique for encoding the structural features of a molecule into a fixed-length binary or integer vector for use in machine learning and similarity searching. Fingerprints serve as the coordinate system for defining the boundaries of an applicability domain.

• ECFP4: Circular fingerprints encoding atom neighborhoods up to diameter 4
• MACCS Keys: 166-bit structural key-based fingerprint
• Morgan Fingerprints: Generalization of ECFP with configurable radius
• Distance metrics like Tanimoto similarity quantify proximity within the domain

The process of assigning a confidence interval or probability distribution to a model's prediction, distinguishing between aleatoric uncertainty (inherent data noise) and epistemic uncertainty (model ignorance). This is the statistical backbone of applicability domain determination.

• Gaussian Processes provide native predictive variances
• Ensemble methods capture epistemic uncertainty through model disagreement
• Conformal prediction offers distribution-free, finite-sample coverage guarantees
• High uncertainty signals that a query lies outside the reliable domain

A model-agnostic, distribution-free framework that produces prediction sets with a rigorous, finite-sample guarantee of coverage. Unlike heuristic distance-based domain definitions, conformal prediction provides a statistically valid measure of confidence for molecular property predictions.

• Produces prediction intervals with user-specified confidence levels (e.g., 90%)
• Inductive conformal prediction uses a held-out calibration set
• Mondrian conformal prediction provides conditional validity across subpopulations
• Larger prediction sets indicate greater uncertainty and potential domain violation

A pair of structurally similar molecules that exhibit a drastic difference in biological potency. Activity cliffs represent the most challenging regions of chemical space for applicability domain assessment, as small structural changes produce large property shifts.

• Defined by a high similarity (e.g., Tanimoto > 0.9) and large potency difference (e.g., > 100-fold)
• Often caused by critical pharmacophoric features or binding pocket constraints
• Matched Molecular Pair Analysis (MMPA) systematically identifies these cliffs
• Models trained without cliff-aware strategies may fail catastrophically near these boundaries

The computational estimation of a drug candidate's Absorption, Distribution, Metabolism, Excretion, and Toxicity properties. Each ADMET endpoint has its own distinct applicability domain, and a compound may fall within the domain for one property while being outside it for another.

• LogP and aqueous solubility define lipophilicity domains
• CYP450 inhibition models require specific isoform domain definitions
• hERG cardiotoxicity and AMES mutagenicity are critical safety domains
• Multi-parameter optimization requires simultaneous domain assessment across all endpoints