A linear combination of pharmacophore hypotheses as a new tool in search of new active compounds--an application for 5-HT1A receptor ligands

Abstract

This study explores a new approach to pharmacophore screening involving the use of an optimized linear combination of models instead of a single hypothesis. The implementation and evaluation of the developed methodology are performed for a complete known chemical space of 5-HT1AR ligands (3616 active compounds with K i < 100 nM) acquired from the ChEMBL database. Clusters generated from three different methods were the basis for the individual pharmacophore hypotheses, which were assembled into optimal combinations to maximize the different coefficients, namely, MCC, accuracy and recall, to measure the screening performance. Various factors that influence filtering efficiency, including clustering methods, the composition of test sets (random, the most diverse and cluster population-dependent) and hit mode (the compound must fit at least one or two models from a final combination) were investigated. This method outmatched both single hypothesis and random linear combination approaches.

Figure 1. The development of an optimal combination of pharmacophore models.

The development of an optimal combination of pharmacophore models. Transparent boxes show the logical steps of the workflow; cylinders represent data sources; colored boxes reflect the compound character: gray – inactives, orange – actives or the active’s selection method (blue, red or green), which is consequently used in subsequent figures. The population of the compound set is given in brackets. Thick arrows indicate the use of data sets.

Figure 2. Affinity distribution of 3616 5-HT_1AR ligands retrieved from the ChEMBL database version 5.

Affinity distribution of 3616 5-HT_1AR ligands retrieved from the ChEMBL database version 5.

Figure 3. A dendrogram obtained using the manual clustering procedure.

A dendrogram obtained using the manual clustering procedure. The number of compounds comprising each cluster is given in brackets. The last column presents a feature composition of the pharmacophore model created for a given cluster. The feature abbreviations used are: hydrogen bond acceptor – A, hydrogen bond donor – D, hydrophobic group – H, positively charged group – P, aromatic ring – R.

Figure 4. The optimized values of MCC for each possible scheme.

The optimized values of MCC for each possible scheme. The length of combination is shown at the top of the bars. The composition of combinations based on the manual clustering approach is shown in Figure 5.

Figure 5. A composition of each top-ranked linear combination obtained using the manual clustering procedure.

A composition of each top-ranked linear combination obtained using the manual clustering procedure. Each filled square denotes presence of a hypothesis developed on a particular cluster in the optimal combination for appropriate conditions. Colors code the type of the test set: blue – diverse, red – populated, and green – random. The last row contains the total number of hypotheses forming a respective top-ranked combination. The values of the optimized statistical parameters for manual clustering are shown in Figure 4, and those for accuracy and recall are shown in Figures S1 and S2, respectively. The exemplary linear combination (manual/random/hit-once; 7 hypotheses long) is shown in Figure 9.

Figure 6. The MCC results for the validation of the top linear combinations.

The MCC results for the validation of the top linear combinations. The length of combination is shown on top of the bars.

Figure 7. Exemplary pharmacophore hypothesis selected for arylpiperazines with classical amide fragment.

Exemplary pharmacophore hypothesis selected for arylpiperazines with classical amide fragment mapping 6 out of 10 cluster representatives. The model fit 462 of the 533 compounds (87%) in the cluster. The feature abbreviations are: hydrogen-bond donor – D, positively charged group – P, aromatic ring – R.

Figure 8. An optimization curve for the investigated parameters of a top-ranked linear combination of MCC.

An optimization curve for the investigated parameters of a top-ranked linear combination of MCC (M2D/random/hit-once); arrows indicate the maximum value: MCC reached a rate of 0.686 for 10 hypotheses (also see Figure 6); the optimization of accuracy and recall had the highest values for a combination of 8 and 9 hypotheses, respectively (also see Figures S1 and S2).

Figure 9. The best linear combination of pharmacophore models obtained for manual clustering and MCC optimization.

The best linear combination of pharmacophore models obtained for manual clustering and MCC optimization (manual/random/hit-once; see also Figures 4 and 6). For each hypothesis the best fitting compound is presented, along with a matrix of distances (in angstroms) between features and a name of cluster it was developed on. The feature abbreviations used are: hydrogen bond acceptor – A, hydrogen bond donor – D, hydrophobic group – H, positively charged group – P, aromatic ring – R.