14 May 2018

Fragments vs Gram-negative bacterial PPAT

Of the 30+ fragment-derived drugs that have entered the clinic, only one is an antibiotic. In part this reflects a shift away from this therapeutic area by many companies. Novartis, though, has continued to invest, as demonstrated by two consecutive papers in J. Med. Chem.

The researchers were interested in the enzyme phosphopantetheine adenylyltransferase (PPAT, or CoaD), which catalyzes the penultimate step of coenzyme A biosynthesis from ATP and 4'-phosphopantetheine. Although the enzyme is present in all organisms, the bacterial form is highly conserved across prokaryotes and significantly different than the human form. It is also essential for bacterial growth, thus making it an attractive target.

In the first paper, Robert Moreau and colleagues start big: a high-concentration screen (at 500 µM) of 25,000 fragments as well as NMR-based screens of their core 1408 fragment library. Triaging both hit sets led to a cornucopia of 39 crystal structures of bound fragments; the chemical structures of a dozen are provided in the paper, with IC50 values from 31 to >2500 µM. Perhaps surprisingly, all of these bound at the pantetheine binding site of the enzyme, suggesting that this is a “hotter” hot spot than the ATP-binding site.

Three of the fragments are described in more detail. The first was optimized from 273 µM to 4.3 µM, but subsequent advancement was unsuccessful. The second fragment, with an IC50 of 230 µM against E. coli PPAT, could be optimized to mid-nanomolar inhibitors; unfortunately these were much less active against PPAT from P. aeruginosa, so this series was also abandoned. But the third fragment discussed, compound 6, proved to be more tractable.


Initial optimization based on other hits led to compound 32, and addition of a methyl to the benzylic linker provided a satisfying 30-fold improvement in potency for compound 33. This “magic” methyl appeared to help desolvate the adjacent NH as well as pre-orient the molecule in the bound conformation. Further growing from this position led to compound 53, which provided a further 7-fold improvement in potency. Crystallography revealed a hydrogen bond between the nitrile nitrogen and a protein backbone amide. Unlike the previous series, this compound was active against PPAT from both E. coli and P. aeruginosa.

The second paper, by Colin Skepper and colleagues, describes further optimization of the molecules to picomolar binders. There’s a lot of lovely medicinal chemistry in both papers, but unfortunately all the molecules displayed at best only modest antibacterial activity. One problem is that Gram-negative bacteria have two membranes: an outer one which blocks lipophilic molecules and an inner one which blocks hydrophilic molecules. Compounds that can make it past these barriers also face an array of diverse efflux pumps, and these seemed to be the downfall of this project. The core of the molecule makes multiple hydrogen bonds to PPAT; about twenty different heterocycles were tested, but most of these had significantly lower potency, and the active ones were efflux pump substrates.

These difficulties in part explain why companies have been moving away from antibiotics. This was not a minor effort: each paper listed more than twenty authors. The second ends somewhat wistfully. “Although none of the series disclosed… yielded a clinical candidate, it is our hope that these studies will help pave the way toward the discovery of new Gram-negative antibacterial agents with novel modes of action.” It is a worthy – if arduous – quest.

07 May 2018

Fragment growing via virtual synthesis and screening

Practical Fragments has covered virtual screening for nearly ten years, and the tools continue to improve. More recently, researchers are using computational approaches not just to dock libraries of molecules, but to decide what compounds to make. The latest example, called AutoCouple, is described in an ACS Cent. Sci. paper by Cristina Nevado, Amedeo Caflisch, and colleagues at University of Zurich.

The researchers have a standing interest in bromodomains, epigenetic “reader” proteins that bind acetylated lysine residues. In particular, they were interested in CBP (cyclic-AMP response element binding protein). Previously the researchers had identified compound 1 through virtual screening, but although this compound had sub-micromolar affinity, it showed no cell-based activity, presumably due to the carboxylic acid, a moiety usually associated with poor cell permeability. Indeed, a CBP series we discussed earlier this year that also contained a carboxylic acid had no cellular activity.

To come up with better molecules the researchers used a program they developed and named AutoCouple because it virtually “grows” a fragment using common coupling reactions such as amide formation, Buchwald-Hartwig amination, and the Suzuki-Miyaura reaction. An initial set of 270,000 commercial compounds was computationally filtered to remove large molecules and those containing undesirable moieties. Potentially self-reactive building blocks were also removed. Ultimately 70,000 virtual compounds based on growing compound 2 (the key fragment of compound 1) were designed and docked into multiple crystal structures of CBP, and 53 were actually synthesized and tested.


Four of the 33 amides synthesized were sub-micromolar, compound 5 being one example; another 17 were low micromolar. (Five of the 10 Suzuki-derived compounds were also sub-micromolar, as was at least one of the amines.) Compound 5 was improved by using information from one of the other tested molecules to generate compound 16, with low nanomolar affinity. Crystallography confirmed that this compound binds as the docking had predicted, in a similar manner to compound 1.

Happily, not only was compound 16 more potent than compound 1, it was also active in cells. Moreover, it showed reasonable selectivity against a dozen other bromodomains.

Overall AutoCouple looks like it could be a useful tool to design and prioritize compounds for synthesis. Moreover, like the growing via merging “PINGUI” approach we highlighted earlier this year, the Python scripts appear to be freely available. It would be fun to benchmark both methods on the same targets to see how they compare.