New tools for covalent fragment-based lead discovery

Covalent fragments provide an opportunity to both drug difficult targets and to more completely shut down targets. Success has spurred interest, and the literature is exploding. It has been just over a month since our last post on the topic, and already three new papers are worth highlighting.

 

The first, in Eur. J. Med. Chem. by György Keserű and collaborators at the Hungarian Research Centre for Natural Sciences and University of Szeged, describes a library of 24 covalent fragments. All of these contain the same relatively simple core but vary in their covalent warheads or how the warhead is attached.

 

The idea is to explore warhead reactivity in the context of a “vanilla” fragment that could provide modest but nonspecific hydrophobic interactions with proteins. The 14-atom 3,5-bis(trifluoromethyl)phenyl core was chosen because it is commonly used in medicinal chemistry and lacks polar atoms likely to make specific interactions to proteins. Also, the electron withdrawing trifluoromethyl groups make the warheads more reactive. The UV absorbance and lipophilicity also make derivatives synthetically easy to work with, and the fluorine atoms are useful for ¹⁹F NMR.

 

The warheads themselves span a vast range of reactivity as assessed both computationally and experimentally (by reactivity with glutathione). Some, such as maleimides and isothiocyanates, are so highly reactive that they are often used for nonspecific protein labeling, while others, such as styrene and acetylene, are quite unreactive. In the middle are moieties like acrylamides, chloroacetamides, and epoxides.

 

The researchers screened the library (at 100 µM) against four unrelated kinases: BTK, ERK2, RSK2, and MAP2K6. Unsurprisingly, four of the hottest fragments inhibited all the kinases, while the seven weakest warheads were inactive. Things got interesting in the middle though, with different inhibition profiles seen for different kinases.

 

Next, the researchers tested their fragment sets against two new kinases, JAK3 and MELK. Both kinases yielded several hits. Replacing the vanilla fragment with small hinge-binding elements for the relevant warheads rapidly yielded nanomolar inhibitors. Covalent inhibitors had already been reported for JAK3 but not for MELK. The researchers suggest using their library as a rapid tool for assessing cysteine accessibility. If you are interested in trying this at home, the authors have offered to send the library upon request.

 

The second paper, in ChemBioChem by György Keserű, Stanislav Gobec, and a large multinational group of collaborators, describes a slightly expanded covalent library consisting of 28 compounds representing 20 different warhead chemotypes, all with the same 3,5-bis(trifluoromethyl)phenyl core. Usefully, glutathione reactivity kinetics are provided for all the fragments. The fragments were screened against six different (non-kinase) targets, providing hits against all of them. ¹⁹F NMR as well as mass spectrometry was used to confirm binding.

 

It is always nice to see new types of covalent warhead chemistries, but medicinal chemistry tends to be somewhat conservative: if something works clinically and isn’t (too) toxic, we’ll stick with it. Thus the continuing interesting in acrylamides, which are found in five of the six approved covalent kinase inhibitors. Enter the third paper, in J. Med. Chem., by Adam Birkholz and colleagues at Amgen, which systematically explores the glutathione reactivity of substituted N-phenyl acrylamides.

 

The researchers first examine 11 α-substituted N-phenyl acrylamides. For the most part electron-withdrawing substituents increase the reactivity of the warhead, though fluorine has the opposite effect, attributed to its mesomeric electron-donating ability.

 

Next, the researchers turn to 21 β-substituted N-phenyl acrylamides. Again, electron withdrawing substituents increase the reactivity of the acrylamides. For aminomethyl substituents, the reactivity is lower than the parent unsubstituted acrylamide for amines with pK_a < 6, while the more basic amines show increased reactivity. All experiments were conducted at pH 7.4, and computational modeling suggests that the protonated amine inductively withdraws electron density from the acrylamide, thereby increasing its reactivity.

 

While the general trends reported in the paper are expected, the actual numbers provide a valuable resource. One of the challenges of covalent drugs is ensuring the warhead is reactive enough to bind to the target but not so reactive that it binds to other targets or is cleared too rapidly. By knowing how much a given substituent is likely to increase – or decrease – reactivity, chemists can more precisely tune their molecules.

 

Our medicinal chemistry toolkit is expanding, and covalent molecules are playing a growing role.

Fragment mixtures vs protein mixtures

In FBLD – as in most areas of research – speed and efficiency are prized. The faster you can find quality fragments, the faster you can advance them. NMR-based screening remains one of the most popular fragment-finding methods, and in a recent Molecules paper William Pomerantz and collaborators at the University of Minnesota and Gustavus Adolphus College provide an accelerated workflow.

 

The Pomerantz lab is well known for protein-observed ¹⁹F (PrOF) NMR, in which fluorine-labeled residues are incorporated into proteins. This is easily accomplished by supplementing the media with fluorine-containing amino acids during protein expression. To date more than 15 fluorinated amino acids have been tested in more than 70 proteins, ranging from 7 to 180 kDa in size. Because the chemical shift of fluorine is so sensitive to its environment, a fragment binding nearby can be readily detected by PrOF NMR.

 

When a single type of amino acid is fluorinated, the resulting protein spectrum is considerably simpler than in traditional protein-observed NMR methods. Taking advantage of this, the researchers mixed two different bromodomain proteins: the human oncology target BPTF and PfGCN5 from the malarial parasite Plasmodium falciparum. Both of these bromodomains contain a tryptophan in their N-acetyl-lysine binding sites, so each protein was labeled with 5-flurotryptophan. The proteins were then screened (at 50 µM each) against 467 fragments from Life Chemicals in pools of 4-5 (at 400 µM each). Chemical shift perturbations of the binding-site tryptophan were seen for half of the 98 pools. To determine which fragments were responsible for these shifts, the researchers tested their fragment mixtures against the relevant proteins using (ligand-detected) CPMG NMR. Since they had previously determined the ¹H NMR spectra of all their fragments, it was easy to pick out the binders.

 

Hit rates were similar for both BPTF (9.8%) and PfGCN5 (9.2%), and 4.1% of fragments hit both bromodomains. The researchers had previously screened this library, which is enriched for shapely fragments, against the bromodomain BRD4 D1 (see here) and obtained a similar hit rate. Statistical analyses revealed that the 3D-character for PfGCN5 hits is similar to the library as a whole, as had also been seen for BRD4 D1, while the BPTF hits tended to be flatter.

 

The researchers also followed up on several  fragments individually. One in particular had low micromolar affinity for PfGCN5 as assessed with both PrOF NMR and ¹H-¹⁵N HSQC NMR titrations. Interestingly, this fragment also caused a chemical shift in a different 5-fluorotryptophan residue some 22 Å away from the canonical binding site. Binding at this site could not be competed by a known high-affinity ligand, and a computational screen using FTMap suggested that this does appear to be a secondary binding site.

 

Overall this approach appears to be an appealing workflow as judged by comparing required time, protein, and ligand amounts to other NMR-based screening cascades. As the researchers note, it is advantageous to assess both protein and ligand behavior, as done here. Have you tried using PrOF, and if so how has it performed for you?

Stabilizing protein-protein interactions: part 2

Stabilizing – rather than disrupting – protein-protein interactions is increasingly popular, particularly in the context of PROTACs and molecular glues. Last year we highlighted research in which covalent Tethering was used to identify fragments that could stabilize a protein-protein interaction, but only when the fragments were disulfide-bonded to one of the proteins. In a new (open access) J. Med. Chem. paper some of the same researchers, including Christian Ottmann and a large group of collaborators at Eindhoven University of Technology, University of Lille, AstraZeneca, and UCSF, have extended the approach to non-covalent fragments.

 

As before the researchers were interested in the adapter protein 14-3-3; the various isoforms of this “hub” protein each interact with hundreds of other proteins, many of which are implicated in disease. Natural products such as fusicoccin A stabilize interactions between 14-3-3 and some protein partners, suggesting ligandability, but the 13 chiral centers make analoging somewhat daunting. Two 14-3-3 partners implicated in oncology include the transcriptional coactivator TAZ and the tumor suppressor p53. Stabilizing the interactions of either of these proteins with 14-3-3 could have anti-cancer effects.

 

To find fragments that could stabilize these interactions, the researchers first grew crystals of 14-3-3σ in complex with peptides derived from either TAZ or p53. These crystals were then soaked in 100 pools of five fragments, each at 10 mM. Electron density was seen for several fragments; all fell into one of two scaffolds. Interestingly, both scaffolds contained an amidine moiety that forms a salt bridge with a glutamic acid side chain in 14-3-3σ.

 

One fragment, AZ-003, made interactions both with 14-3-3σ as well as with the TAZ-derived peptide. The peptide itself actually changed conformation so that a carbonyl oxygen could form a hydrogen-bond to the fragment, and three additional peptide residues could be resolved in the electron density that were not seen in the absence of fragment.

 

In the case of the other complex, none of the fragments interacted directly with the p53-derived peptide, though they bound nearby. Fragment growing and crystallography revealed that some of these larger molecules made water-mediated interactions to the peptide and were specific for p53 over TAZ, demonstrating that selectivity can be achieved. In the case of one molecule, AZ-008, crystallography was unsuccessful but modeling and protein-observed NMR experiments suggested direct interactions with the p53-derived peptide. Fluorescence polarization and SPR experiments revealed that 1 mM AZ-008 improved the affinity of 14-3-3σ for the peptide by an unambiguous but modest two-fold.

 

As with the previous paper, there is still much to do, and it will likely not be easy. The Supporting Information contains more than 30 crystal structures of small molecules bound to 14-3-3, both covalently and non, hinting at the effort required. However, the fact that both AstraZeneca and Novartis are working on 14-3-3 proteins – along with at least one startup – bodes well. Expect a part 3!

Fragments vs (lots of) RNA

RNA is hot. Hundreds of millions of dollars have gone to startups focused on finding small molecules that bind RNA, and plenty of large pharmaceutical companies are pursuing this class. As with all difficult targets, fragments have a role to play, as demonstrated by a paper just published (open access!) in ChemBioChem by Harald Schwalbe and collaborators at Johann Wolfgang Goethe-University Frankfurt and Saverna Therapeutics.

 

The researchers assembled a collection of 101 fluorine-containing fragments and pooled these into five sets of 20-21 compounds each. These were then screened (at 50 µM) against 14 different RNA targets, ranging from 14-nucleotide hairpins to a 127-nucleotide riboswitch. The primary screen was ligand-detected ¹⁹F NMR using CPMG, in which ligand binding causes a change in relaxation which is detected as a decreased signal. All hits from the mixtures were confirmed as single molecules. To assess selectivity, all the compounds were also screened against 5 DNA targets and 5 proteins.

 

The results are not entirely unexpected. Some of the fragments did not hit any targets, while others were rather promiscuous: a couple showed strong binding to 8 of the 14 RNA targets. That said, strong is a relative term; the highest affinity of any fragment measured was 375 µM.

 

The RNA targets spanned a variety of structures, but the most hit-rich were aptamers and riboswitches, which bind specific small molecules or ions. These had 7 to 26 hits each. In contrast, the other RNA targets tested all had six or fewer fragment hits. For the aptamers, competition experiments with the natural ligands suggested binding at the orthosteric sites in some cases but not others.

 

Hit rates against four of the five proteins were also high, with 16 to 55 hits each. The fifth protein, a challenging phosphatase, had only four hits. This was still better than the 24-nucleotide DNA duplex, with a single hit. The four G-quadruplex DNA targets had between 12 and 20 hits, consistent with prior research. Taken together, the results suggest that RNA riboswitches and aptamers may be reasonably ligandable, while RNA targets that do not normally bind small molecules may be more challenging.

 

The researchers also conducted cheminformatic analyses. Not surprisingly given the relatively small library, there were no strong correlations between molecular features and targets bound, though fragments that hit had a slight tendency towards more aromatic atoms and fewer sp³-hybridized carbons. This is consistent with a vigorously debated paper we highlighted in July.

 

Finding weak hits is one thing, but advancing them has been challenging for RNA targets. The researchers provide an example in which they link a fragment to an intercalator to generate a low micromolar binder, but intercalators are often nonspecific, and affinity would still need to be further improved. Practical Fragments first highlighted a fragment screen against RNA more than a decade ago, and earlier this year we noted a high nanomolar binder, but I have yet to see an attractive low nanomolar lead emerge.

 

Nonetheless, this paper provides a solid launching point for such an effort to succeed. In particular, the researchers laudably include structures of all the library members as well as the raw screening data of all the hits on all the targets in the Supporting Information. If you are feeling adventurous, you now have plenty of starting points to choose from.