2.1.2. Pharmacophore model validation

The validation of a pharmacophore model is considered an important step before its use in virtual screening. The Predictive ability, specificity, and sensitivity of a pharmacophore model are earnest metrics for the reliability of performance. We assessed the predictive ability of the model on a decoy set to determine the accuracy of recognition of active and inactive compounds.

Sensitivity states how good the model correctly classifies compounds and specificity shows how well the model is able to exclude inactive compounds [28]. The Sensitivity (TPR) and specificity (TNR) can be measured using Eqs. (1) and (2), respectively.

Other metrics to validate the pharmacophore model are the receiver operating characteristic (ROC) curve which indicates how well a model can differentiate between active and inactive compounds [31] and the area under the ROC-curve (AUC) [33]. The ROC curve provides the true positive rate plotted against the false positive one of the hits. If the curve was sharp and then flattened, this means that the model ranked the active compounds higher than the inactive ones. The AUC value lies between 0 (bad classifier) when the model ranks all the inactive compounds first and 1 (excellent classifier) on the contrary situation. Hence, to evaluate the performance of the model, a set of 703 inactive compounds was obtained from DUD-E as a decoy set for 5 active and selective COX-2 inhibitors. The decoy and active compounds have similar physicochemical properties but different two dimensional topological ones [34, 35, 36]. Using the idbgen routine included in LigandScout, the compounds were converted into a LigandScout format [52]. In addition to these metrics, the statistical parameters of goodness of hit score (GH) [37], enrichment factor (EF), and accuracy (ACC) were determined to investigate the performance of the model. The formulas are written in Eqs. (3), (4), and (5) below:

Ht is the total number of hits and Ha is the number of active hits. The range of the GH score is (0–1) with a threshold value equal to 0.6 [38]. Ideally, when the model picks all the active compounds with no inactive ones, it will have a steep slope for the ROC curve, a high value of AUC, a high EF value, and the highest value of sensitivity and specificity which is 1 [39].