A Primer on Molecular Similarity in QSAR and Virtual Screening
Part II - How reliable are experimental measurements (endpoints) in QSAR studies?
b. Bioactivity
One of the endpoints routinely modeled are bioactivity data, also giving rise to the area of QSAR (quantitative structure-activity relationships). There are at least three different classes of bioactivity endpoints:
1. Affinity assays measuring the drug-receptor dissociation constant (commonly abbreviated as Ki values),
2. Functional assays which measure, for example, inhibition constants (IC50 values) and
3. Cell-based assays that give some type of phenotypic readout.
While cell-based assays are generally more complex and have potentially more sources of error, they offer the huge advantage of taking intra- (and inter-) cellular signaling into account3. Affinity assays are a good choice for the quick measurement of binding energies, but they do not necessarily give much information about how effective a compound actually is. Two compounds with identical Ki values can (but don’t have to) have vastly different IC50 values4. Functional assays are in some sense located between these two assay types - they are less complex than cell-based assays (at the expense of omitting the signaling network in the cell), but they provide some information about the efficacy of the compound against the target. (Under the assumption that the compound will reach the target in later stages such as animal models - an assumption that needs to be thoroughly validated separately.)
The error sources in biological assays are manifold. Solubility in the solvent may be limited, as shown in the previous paragraph, leading to different than expected concentrations in solution (and thus less likelihood of observing activity). In addition, "frequent hitters" may form aggregates in solution5, leading to unspecific inhibition (with potentially misleading structure-activity relationships that can be inferred from those "activities"). DMSO may influence the target protein at too high a concentration, and evaporation of solvent leads to unwanted compound concentrations. Also, from the technical side, high-throughput readouts such as those based on optical detection, often do not show consistent sensitivity over the wells of a screening plate, leading to "edge effects", which need to be normalized in a second step.
3. Summary and Conclusions
While experimental measurements are the "gold standard" against which every hypothesis and model needs to be validated, we have shown that not every number obtained from an experiment can be taken at face value. For two examples, bioactivity and solubility data, we have given reasons why values may not be comparable between different laboratories (or even among different conditions used in the same laboratory). The conclusions are two-fold: Firstly, merging data points from different sources is a tricky process. And, secondly and most importantly, before any modeling attempt is started on a particular data set, the user is advised to have a thorough look at the underlying data quality - since every model can only be as good as the input data provided.
To be followed by part III Connecting descriptors and experimental measurements - model generation.
References
1. Bhattachar, S. N.; Deschenes, L. A.; Wesley, J. A., Solubility: it's not just for physical chemists. Drug Discov Today 2006, 11, (21-22), 1012-8.
2. Llinas, A.; Burley, J. C.; Box, K. J.; Glen, R. C.; Goodman, J. M., Diclofenac solubility: independent determination of the intrinsic solubility of three crystal forms. J Med Chem 2007, 50, (5), 979-83.
3. Clemons, P. A., Complex phenotypic assays in high-throughput screening. Curr Opin Chem Biol 2004, 8, (3), 334-8.
4. Knight, Z. A.; Shokat, K. M., Features of selective kinase inhibitors. Chem Biol 2005, 12, (6), 621-37.
5. McGovern, S. L.; Caselli, E.; Grigorieff, N.; Shoichet, B. K., A common mechanism underlying promiscuous inhibitors from virtual and high-throughput screening. J. Med. Chem. 2002, 45, (8), 1712-1722.
|
