Creating a Predictive Model

As a means to aid the design of future Series 4 compounds, Dr. Murray Robertson made an informal attempt to create a pharmacophore model based on the 28 compounds that were found to cause disruption against Na⁺ regulation in PfATP4. Murray used Discovery Studio from Accelrys (now BIOVIA) to screen the 28 compounds using the Common Feature Pharmacophore Generation protocol, producing 10 four-feature models. These were narrowed down based on poses and score to one model that was developed further (Figure 4).

Figure 4: The four-feature model chosen for further development.

The 28 compounds were then aligned in the model to create a shape feature that could be used to manually predict the shape of the active site (Figure 5). Exclusion features were then added in areas where high scoring, inactive ligands penetrated out of this shape.

Figure 5: Fine tuning the model. (A) Overlay of the 28 active compounds. (B) Shape feature based on the overlay. (C) High scoring, inactive ligands penetrating out of the shape feature. (D) Exclusion features indicated as grey spheres.

Unfortunately, when this model was applied to the set of compounds evaluated for ion regulation activity in 2014, the predictions were found to correlate poorly with the experimental results (Figure 6). It was thought this was due to the model not taking into account overlapping binding sites or compound chirality.

Figure 6: Poor correlation was seen between the prediction and experimental data.

This model was also used to screen the Maybridge library of compounds to identify a small and diverse selection of molecules to evaluate in the PfATP4 assay. The results were manually filtered based on fit value, pose, shape, fit with "cavity feature" and diversity to give a final selection of 18 compounds with fit values between 3.88 and 3.66. These 18 compounds were subsequently evaluated in the Kirk ion regulation assay, however none were found to exhibit activity, which confirmed that the model required further optimisation.