Dr Henriette Willems Abstract

A case study of academic drug discovery enabled by virtual screening

Henriette Willems¹, Steve Andrews¹, Helen Boffey¹, Jon Clarke¹, Simon Edwards¹, Christopher Green¹, John Skidmore¹, David Winpenny¹

¹ALBORADA Drug Discovery Institute, University of Cambridge

Autophagy is a mechanism by which cells remove unnecessary and dysfunctional components. Increased autophagic clearance of aggregated proteins is postulated to be beneficial in treating neurodegenerative diseases. The PI5P4K lipid kinases play a key role in regulating autophagy (Mol Cell. 2015, 57, 219-234.) Three isoforms of PI5P4K are known (alpha, beta and gamma), with alpha the most active isoform. They have very low homology to eukaryotic protein kinases, and no structural similarity in the active site. No potent and selective inhibitors of any PI5P4K isoform have been published to date. This abstract describes the identification of potent and selective inhibitors of PI5P4Kalpha and gamma through virtual screening.

At the ALBORADA Drug Discovery Institute, we started our PI5P4K program with a virtual screening (VS) approach. As our capabilities expanded we added other lead identification strategies including both an HTS and fragment screening. For the VS approach, our initial aim was to purchase around 1000 compounds for biological screening. We built several MOE 3D pharmacophores on a public PI5P4Kalpha apo crystal structure to use as the first step in a virtual screening protocol. All pharmacophores included donor and acceptor features thought to be involved in ATP binding. Excluded volumes were added to the active site to define the shape. The pharmacophores were used to screen a 31,000-compound commercially available compound library. At the same time, this database was also docked to the protein using a GOLD library screen protocol.

The MOE pharmacophores we generated were either too selective (<1000 hits), or not selective enough (16,000 hits) and GOLD docking returned almost the entire database. However, using the MOE pharmacophore to filter the docked poses gave around 6,000 hits, mostly in reasonable binding modes. A diverse set of 960 compounds was picked from this subset based on docking scores and clustering. From these, our kinase function bioassays identified 11 compounds with IC50 potencies below 10 μM for PI5P4Kalpha and 5 for PI5P4Kgamma. This equates to hit rates of 1.2% and 0.5%.

As the virtual screening approach was successful with the first library, it was applied to a second commercial compound library containing around 1.5 million compounds. 848 compounds were purchased using a protocol very similar to the one above, and another 10 compounds for PI5P4Kalpha and 12 for PI5P4Kgamma were identified with IC50 potencies below 10 μM (hit rate 1.2% and 1.4% respectively).

A number of hits were followed up with analogue purchases, with compounds selected by substructure, Tanimoto similarity or shape similarity. More hits were identified for 4 of the PI5P4Kalpha hit clusters and 3 of the PI5P4Kgamma clusters. In total, 2 PI5P4Kalpha compounds, and 3 PI5P4Kgamma analogues under 100 nM were identified. Thus, potent novel inhibitors with developable ADMET properties (moderate permeability, high stability in human microsomes, low or moderate efflux) were identified by computational methods and compound purchasing.

Further work was carried out to identify alternative active series through an HTS screen of a 175,000 compound library. The HTS had a much lower hit rate than the virtual screens, 0.015%. The cost of purchase and screening of the VS selected compounds was £2000-4000, with the cost for the HTS library over 10-fold more. This demonstrates that virtual screening of commercial compounds can be a very accessible and cost-effective method for identifying new leads for structure-enabled targets, even those for which only an apo structure is available.