Fragments vs DAPK3, computationally and experimentally

Computational approaches for discovering hits often involve sorting through many possibilities and examining a few closely. With luck, some of the predicted molecules will bind to the protein of interest. However, these don’t always bind for the “right” reason: sometimes a fragment predicted to bind one way will turn out to bind in quite a different manner. A recent Angew. Chem. Int. Ed. paper by Gisbert Schneider and colleagues at the ETH in Zürich and SARomics in Lund reports a possible example.

The researchers were interested in death-associated protein kinase 3 (DAPK3), which is implicated in several diseases. Previous work had shown that fasudil inhibits this kinase, though it hits others as well. Fasudil was used as a starting point for de novo fragment discovery using software called DOGS (Design of Genuine Structures). This is a scaffold-hopping approach in which virtual chemistry is used to generate readily accessible alternatives to a starting molecule. In this case, 347 of the 521 suggested inhibitors were fragment-sized. These were prioritized using in-house software, and compound 2 – one of the top hits – was chosen for synthesis and characterization.

Happily, compound 2 turned out to be fairly potent for its size, with impressive ligand efficiency. It is also quite different from fasudil (Tanimoto similarity = 0.16). Indeed, while fasudil is likely to be positively charged at physiological pH, compound 2 is likely to be negatively charged. Moreover, of 27 other kinases tested, compound 2 hit only one other with similar potency.

For those who have worked on kinases, compound 2 does appear unusual. A crystal structure of this molecule bound to DAPK3 revealed that it sits in the ATP-binding pocket but without making any conventional hydrogen bond interactions to the so-called hinge region of the kinase. Although no reported crystal structures show fasudil bound to DAPK3, structures with other kinases reveal the nitrogen of the isoquinoline moiety making a hydrogen bond to a backbone amide in this part of the protein.

The software used to prioritize compound 2 is based not on docking but on machine learning using the ChEMBL database, and the researchers were interested in what else this fragment might inhibit. Not surprisingly given the aryl sulfonamide moiety, several carbonic anhydrases came up, and two were confirmed experimentally.

Interestingly, the diuretic drug azosemide, whose physiological target is unknown, contains compound 2 as a substructure, and the researchers found that this molecule inhibits DAPK3 with low micromolar affinity. It also binds human carbonic anhydrase IX with similar affinity. The researchers suggest that these targets could at least partially explain the mechanism of the drug, as well as some of its side effects. It would be interesting to see cell data against these two targets, as well as the crystal structure of azosemide bound to DAPK3.

The ability to predict biological targets of molecules with the aid of machine learning would clearly be valuable (see also here). And of course new approaches for scaffold hopping are always valuable. In this case DOGS did retrieve an active (albeit odd) molecule when fed a conventional kinase inhibitor; it is as if you threw a ball and your dog fetched a slipper. I will be curious to see this applied to more systems.