GeinQSAR Studio
Home › Glossary

Glossary

The QSAR terms used in the program and in this manual, each in a few plain sentences.

Terms are in alphabetical order. Where the program applies a threshold, the threshold and its literature source are given on the page that shows the statistic, for instance The model dashboard.

Activity cliff

A pair of compounds that are structurally very similar but differ strongly in potency. Cliffs are where the assumption behind QSAR, that similar structures have similar activity, fails, and they limit the accuracy any model can reach on a dataset.

Applicability domain (AD)

The region of chemical space in which a model's predictions can be trusted, defined by the training compounds (OECD principle 3). GeinQSAR marks each predicted compound as inside or outside the domain and states the reasons, using leverage, distance to the nearest training compounds (Euclidean, Mahalanobis or Tanimoto), descriptor ranges or the latent space of a PLS model.

Atom-based cubes

A 3D-QSAR model on a lattice: for every cube of the lattice, the volume filled by the atoms of each of six atom classes (donor hydrogen, hydrophobic, negative, positive, electron withdrawing, other) is a column, and PLS relates the columns to the activity (Dixon et al., 2006).

CCC (concordance correlation coefficient)

Measures how closely predicted values agree with observed ones, penalising both scatter and any systematic offset or slope away from the line y = x. A CCC of at least 0.85 on the test set is a common acceptance threshold (Chirico and Gramatica, 2012).

Censored value

A measurement reported only as a bound, such as IC50 > 10 μM, because the compound did not reach the effect within the tested range. It is not an exact number and is left out of a regression by default; for classification it can still decide the class.

Conformal prediction interval

A prediction interval built from the errors a model made on compounds it was not fitted on, rather than from the model's own standard error, which is the error of its fit. With n such calibration compounds the finest coverage that can be certified is 1 minus 1/(n + 1), so a small calibration set limits how fine an interval can honestly be claimed.

Contour map

The display of a 3D-QSAR model: surfaces around the aligned molecules that enclose the regions where a field contributes most to activity, coloured by whether more of that field raises or lowers it. For a lattice model the usual quantity contoured is the standard deviation times the coefficient (StDev x Coeff); a KEFRA map contours the activity field projected back from the kernel model.

Cross-validation

Estimating how well a model predicts compounds it was not fitted on, by refitting it repeatedly with part of the training set left out and predicting the left-out part. Leave-one-out leaves out one compound at a time; k-fold divides the set into k groups.

Desirability

A way of combining several objectives into one number: each objective is mapped onto 0 to 1 by its own limits and direction, and the mapped values are combined with their weights. It gives a single order, at the price of the weights, which are a judgement.

ECFP (extended-connectivity fingerprint)

A circular fingerprint: each atom's environment up to a given radius is encoded as an identifier, and the set of identifiers describes the molecule. ECFP4 uses a radius of 2 bonds (diameter 4); the identifiers are folded into a bit vector of fixed length, such as 1024 bits.

Eudismic ratio

The ratio of the potencies of the more active enantiomer (the eutomer) and the less active one (the distomer). Its logarithm, the eudismic index, measures how much the activity depends on stereochemistry.

KEFRA

Kernel Ensemble Field Regression Analysis, the lattice-free 3D-QSAR method of GeinQSAR. Each aligned compound is a sum of atom-centred Gaussian densities of its properties, optionally averaged over an ensemble of conformers; two compounds are compared by the overlap integral of their densities, and a Gaussian process fitted on that similarity predicts the activity with an uncertainty and an applicability domain. It builds on continuous molecular fields (Baskin and Zhokhova, 2013).

Latent variable

A combination of the original descriptors, constructed by PLS or principal component regression, on which the model is built. A PLS model with three latent variables (components) uses three such combinations; too many of them fit noise.

Leverage

A measure of how far a compound lies from the centre of the training set in the space of the model. Compounds with leverage above the warning value h* = 3p'/n (p' the number of model parameters, n the number of training compounds) are extrapolations; the Williams plot shows leverage against the standardised residual.

LIBQORA

LIBrary design by QSAR-Oriented Ranking Analysis, the library design of GeinQSAR. It builds the molecules a scaffold and a set of substituents imply, or single-point analogues of a hit, computes their properties, predicts each one with the models you supply, and ranks them by objectives you name, either as one desirability score or as Pareto fronts. See Designing a library.

OECD principles

Five conditions a QSAR model should meet to be used for regulatory purposes: a defined endpoint, an unambiguous algorithm, a defined applicability domain, appropriate measures of goodness of fit, robustness and predictivity, and a mechanistic interpretation where possible. The validation reports of the program are organised around them.

p-scale

Potencies converted to the negative decadic logarithm of the molar concentration: 1 μM becomes 6, 10 nM becomes 8. Models are built on the p-scale because free energies of binding are proportional to it and errors are then roughly constant across the range.

Pareto front

The set of candidates that no other candidate beats on every objective at once. A molecule on the first front can only be improved in one objective by giving up another, so the fronts show the trade-offs that a single combined score averages away.

pChEMBL

The value ChEMBL publishes for a potency measurement on the negative logarithmic molar scale (for example pIC50 = -log10 IC50 in mol/L), given only for exact, comparable records. The Dataset Builder checks its own p-scale conversion against it.

PLS (partial least squares)

A regression method that compresses many correlated descriptors into a few latent variables chosen to explain the activity. It is the classical method of lattice 3D-QSAR, where there are thousands of field values and only tens of compounds, and the default linear method of the 2D-QSAR recipes.

Predictive r2 (r2pred)

The squared correlation of the predictions of an external test set, computed with the training mean as reference. It is the usual measure of external predictivity of a 3D-QSAR model.

Q2 (q2)

The cross-validated coefficient of determination: 1 minus the sum of squared prediction errors of the left-out compounds divided by the total sum of squares of the training activities. Values above 0.5 are usually required; q2 describes robustness, not predictivity for new compounds.

Q2F1, Q2F2, Q2F3

Three forms of the external predictive squared correlation of a test set. They differ in the mean used in the denominator: the training mean (F1, identical to the predictive r2), the test mean (F2) or the training variance scaled to the test set size (F3). A threshold of 0.70 is common.

r2m (r2m)

A test-set metric of Roy and co-workers that combines the squared correlation with the squared correlation through the origin, so that a good correlation with a systematic offset is penalised. The average r2m should exceed 0.5 and the difference between its two forms be below 0.2.

SALI (structure-activity landscape index)

The potency difference of a pair of compounds divided by one minus their similarity. A large SALI marks an activity cliff; for two compounds the descriptors cannot tell apart, such as enantiomers under a 2D fingerprint, the index becomes infinite and the program reports the pair as indistinguishable.

SAMPLS

Sample-distance partial least squares (Bush and Nachbar, 1993): an algorithm that computes PLS from the matrix of inner products between compounds instead of from the field columns. It gives the same model as ordinary PLS and makes leave-one-out cross-validation of 3D-QSAR field matrices fast.

Scaffold

The ring systems of a molecule with the linkers between them, side chains removed (the Bemis-Murcko framework). A scaffold split puts all compounds of a scaffold on the same side of the training/test division, which is a harder test than a random split and a stricter test of a model meant for a new chemical series. It does not give a more honest internal estimate than a stratified random split: on 45 independent data sets it was the closer of the two to the external result on 21 of them, which is 47 %.

Tanimoto coefficient

The similarity of two fingerprints: the number of features they share divided by the number of features present in either. It runs from 0 (nothing in common) to 1 (identical fingerprints).

Training, test and external sets

The training set is used to fit the model; the test set is set aside before fitting and used only to measure predictivity; an external set is a further set of compounds, often from another source, predicted with the finished model.

Williams plot

Standardised residuals plotted against leverage. Compounds beyond three standard deviations are response outliers; compounds beyond the warning leverage are structural outliers, outside the applicability domain.

Y-randomisation

A test for chance correlation: the activity values are shuffled among the compounds and the model is refitted many times. The scrambled models must be clearly worse than the real one; the corrected value cR2p should exceed 0.5.