The art of developing valid QSARs
The art of developing valid QSARs
By John C. Dearden School of Pharmacy and Biomolecular Sciences, Liverpool John Moores University, Byrom Street, Liverpool L3 3AF, U.K.
A QSAR (quantitative structure-activity relationship) is a mathematical correlation between a property (biological, chemical or physical) of interest and one or more indicators of molecular structure.
There are four broad stages in the development of a QSAR model. Firstly, values for the property of interest have to be obtained, either by experiment or from the literature, for a number of representative compounds.
Secondly, indicators of molecular structure (commonly termed descriptors) have to be obtained, by experiment, from the literature, or by using appropriate software.
Thirdly, a statistical technique has to be used to select those descriptors that best model the property of interest, and to derive the mathematical expression that is the QSAR.
Fourthly, the QSAR has to be validated; that is, the model has to be tested to check that it has predictive ability, so that it can be used with confidence to predict the value of the property of interest for compounds not used in the development of the model.
Most descriptors fall into three broad classes – hydrophobic, electronic, and steric. Hydrophobicity generally models transport of a xenobiotic within an organism, although it can also reflect hydrophobic binding to a receptor. It is usually represented by the octanol-water partition coefficient (P).
Electronic descriptors are the broadest class, and include polarity, polarisability, hydrogen bonding, atomic charge, molecular energy levels and many others. Steric descriptors model molecular size and shape; the former is quite simple to model, and includes such descriptors as molecular weight and molar volume, but molecular shape is more difficult, and there is as yet no good universal way of doing so.
Some years ago, the OECD published guidelines for QSAR development, as follows: a valid QSAR should have:
1. a defined endpoint;
2. an unambiguous algorithm;
3. a defined domain of applicability;
4. appropriate measures of goodness of fit, robustness and predictivity;
5. a mechanistic interpretation, if possible.
Within each of those requirements, there are a number of more detailed requirements. However, sadly it is still the case that many published QSAR studies do not meet all those requirements.
A recent publication from our laboratory (Dearden et al, SAR & QSAR in Environmental Research 20 (2009) 241-266) reported 21 different types of error that occur in published QSAR studies.
These include failure to take account of data heterogeneity, use of inappropriate endpoint data, use of collinear descriptors, use of incomprehensible descriptors, error in descriptor values, poor transferability of QSARs, inadequate and/or undefined applicability domain, unacknowledged omission of data points, use of inadequate data, replication of compounds in datasets, too narrow a range of endpoint values, over-fitting of data, use of excessive numbers of descriptors, lack of and/or inadequate statistics, incorrect calculation, lack of descriptor auto-scaling, misuse and/or misinterpretation of statistics, lack of consideration of residuals, inadequate training set and/or test set selection, failure to validate a QSAR correctly, and lack of mechanistic interpretation.
This presentation will consider each of these errors, and will offer guidance on the correct development of QSARs.