Lindsey Burggraaff Abstract

Predicting Polypharmacology in Kinases

L. Burggraaff1, E.B. Lenselink, B.J. Bongers, X. Liu, M. Gorostiola Gonz&aac1, J. van Engelen, H. Hoos2, J.K. Wegner3, M. Steijaert4, W. Jespers, H. Gutiérrez-de-Terán5, H.W.T. van Vlijmen1,3, A.P. IJzerman, G.J.P. van Westen1

1Division of Drug Discovery and Safety, Leiden, The Netherlands
2Leiden Institute of Advanced Computer Science, Leiden, The Netherlands
3Janssen Pharmaceutica NV, Beerse, Belgium
4Open Analytics NV, Antwerp, Belgium
5Department of Cell and Molecular Biology, Uppsala, Sweden
The term polypharmacology is used when drugs (purposely) interact with multiple protein targets. This is in contrast to selectivity, in which compounds are designed to target only one protein. Although polypharmacology is generally associated with side effects, the design of multi-target drugs is gaining attention to tackle more complex disease mechanisms, such as inflammation and cancer [1]. Many polypharmacological drugs in the context of cancer are aimed at (mutant) kinases. Previous efforts show that selective kinase inhibitors become ineffective over time as a result of drug resistance [2]. Additionally, multi-drug treatments are not ideal as drug-drug interactions may occur. Therefore, polypharmacological drugs are of interest when targeting multiple kinases simultaneously, as it is hypothesized that they reduce the onset of resistance [2].
In the light of the Multi-Targeting Drug DREAM challenge 2017 [3] we developed an elaborate strategy to predict compound activity profiles for kinases. In the cancer case study, polypharmacology modeling was limited to nine pre-defined kinases of which four were on-targets and five off-targets. These kinases were subjected to three successive modeling techniques to assess ‘activity’ of compounds: statistical modeling, structure-based modeling, and molecular dynamics simulations. Ultimately we virtually screened the ZINC database [4] for compounds that adhered to our desired pharmacological profile.
We utilized statistical models that were trained and validated on public data derived from ExCAPE [5] and ChEMBL [6]. These models were used to filter the screening compounds and promising compounds succeeded to the structure-based phase.
In structure-based modeling we first thoroughly benchmarked all available protein crystal structures of the kinases at hand. We applied docking of actives, inactives, and decoys [7] to select the most promising and best enriching protein structures. Taking structural diversity into account by comparing co-crystalized ligands, we selected the top 5 enriching crystal structures for each protein. Subsequently SPLIF [8,9] scores were generated for compounds in all protein structures. The multiple docking and SPLIF scores per compound for each target were combined using Z2 scoring [10]. The resulting ensemble models for each target outperformed single models and had high (early) enrichment scores indicating that these models are predictive.
The most promising compounds were re-scored with molecular dynamics using pose-metadynamics [11]. Binding pose metadynamics enhanced compound ranking when tested on sets of 100 compounds per target. Therefore, the most promising virtual screening compounds were additionally subjected to binding pose metadynamics.
Finally, from the resulting compounds we proposed kinase inhibitors that fitted the desired polypharmacology profile and were tested by the challenge organizers. Although for this case study we utilized our strategy for a given set of kinases, this method can easily be applied to any target, provided that sufficient data is available.References
1. Reddy, A. S. & Zhang, S. Polypharmacology: drug discovery for the future. Expert Rev. Clin. Pharmacol. 6, 41–47 (2013).
2. Daub, H., Specht, K. & Ullrich, A. Strategies to overcome resistance to targeted protein kinase inhibitors. Nat. Rev. Drug Discov. 3, 1001 (2004).
3. Schlessinger, A. et al. Multi-targeting Drug Community Challenge. Cell Chem. Biol. 24, 1434–1435 (2017).
4. Irwin, J. J. & Shoichet, B. K. ZINC − A Free Database of Commercially Available Compounds for Virtual Screening. J. Chem. Inf. Model. 45, 177–182 (2005).
5. Sun, J. et al. ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics. J. Cheminform. 9, 17 (2017).
6. Gaulton, A. et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 40, D1100–D1107 (2012).
7. Mysinger, M. M., Carchia, M., Irwin, J. J. & Shoichet, B. K. Directory of Useful Decoys, Enhanced (DUD-E): Better Ligands and Decoys for Better Benchmarking. J. Med. Chem. 55, 6582–6594 (2012).
8. Lenselink, E. B., Jespers, W., van Vlijmen, H. W. T., IJzerman, A. P. & van Westen, G. J. P. Interacting with GPCRs: Using Interaction Fingerprints for Virtual Screening. J. Chem. Inf. Model. 56, 2053–2060 (2016).
9. Da, C. & Kireev, D. Structural Protein–Ligand Interaction Fingerprints (SPLIF) for Structure-Based Virtual Screening: Method and Benchmark Study. J. Chem. Inf. Model. 54, 2555–2561 (2014).
10. Sastry, G. M., Inakollu, V. S. S. & Sherman, W. Boosting Virtual Screening Enrichments with Data Fusion: Coalescing Hits from Two-Dimensional Fingerprints, Shape, and Docking. J. Chem. Inf. Model. 53, 1531–1542 (2013).
11. Clark, A. J. et al. Prediction of Protein–Ligand Binding Poses via a Combination of Induced Fit Docking and Metadynamics Simulations. J. Chem. Theory Comput. 12, 2990–2998 (2016).