256th American Chemical Society National Meeting

This week thousands of chemists converged on the venerable city of Boston for the Fall National ACS meeting. One brief but rich symposium was entitled “Best practices in fragment-based drug design,” organized by Amy Hart, David Marcoux, and Heidi Perez (BMS).

After an introductory presentation by me, Anil Padyana described FBDD at Agios. The company is focused on metabolic enzymes, many of which have dynamic, shallow, and polar active sites – challenging even for fragments! Indeed, a summary of 11 targets screened using a variety of methods revealed generally low hit rates, usually < 3%. The company’s first approved drug came out of an HTS screen against IDH2 of 80,000 compounds that yielded just 24 hits. The molecule that ultimately led to enasidenib was essentially a (large) fragment, with 21 atoms. Agios’ current fragment library is just over 5000 about 10,000 molecules, though they are in the process of expanding this to 20,000 – perhaps part of a general trend. Given the history of enasidenib, they are including molecules beyond the rule of 3, with an upper molecular weight limit of 350 Da, far higher than most respondents in our recent poll.

Cullen Cavallaro presented an early, though still unpublished, FBDD story from BMS: KAT II, a brain enzyme implicated in schizophrenia. Screening 3700 fragments using NMR, SPR, and TSA yielded 236 hits, only 6 of which were common to all methods. All 236 hits were soaked into crystals of KAT II, resulting in 43 structures, 13 of which bound in the active site. Strikingly, 12 of these contained carboxylic acids, which generally don’t cross the blood brain barrier. The lucky thirteenth fragment showed no activity in an enzymatic assay, no thermal stabilization, and only a marginal STD NMR signal. However, through a combination of library synthesis and structure-based design the researchers were able to obtain nanomolar inhibitors. Unfortunately the project was stopped when BMS exited neuroscience.

Anna Vulpetti provided an overview of work done by her and her Novartis colleagues to discover inhibitors of Factor D. A high-throughput screen of the serine protease didn’t yield anything useful, but a combination of fragment screening and structure-based design led to multiple series of inhibitors. Anna is a proponent of fluorine NMR, and Novartis has recently expanded its fluorinated fragment library to 4000 members. Like Agios, they have chosen to include some larger fragments, up to 350 Da.

Finally, David Norton described the initial work done at Astex to discover an orally available ERK1/2 inhibitor, which entered phase 1/2 clinical trials in May of this year. We highlighted some of this work a couple months ago so I won’t cover it in detail, but among other lessons David emphasized the importance of initial fragment optimization before starting to grow.

There were plenty of fragment talks outside the symposium too. Last year we highlighted Ben Cravatt’s strategy for performing fragment screening in cells. Chris Parker, the first author on that paper, has just launched his independent academic career at Scripps Florida, and provided an update. Chris noted that, for phenotypic screening, the approach is essentially a target-finding method, and indeed more than 4000 proteins have been identified, varying over five orders of magnitude in abundance. Having proper controls is critical, and recent efforts include screening pairs of enantiomeric fragments and looking for differences.

Taekyu Lee provided an update of the Vanderbilt MCL-1 program, most recently described in this paper. Some of the molecules shown have low picomolar affinity, mid nanomolar cell activity, and are more than 10,000-fold selective for MCL-1 over BCL-2 and BCL-x_L. The program was partnered with Boehringer Ingelheim earlier this year, and they’ve got competition: four other molecules have entered the clinic.

Covalent fragments were also a theme. Peter Wipf (University of Pittsburgh) described the construction of a 300 compound “mercaptophilic” library. In contrast to other academic reactive fragment libraries we’ve covered (see here, here, and here), this one contains a wide variety of different warheads with varying reactivities.

Finally, Jeff Neitz (UCSF) described efforts against Taspase-1, which is involved in cancer cell proliferation. A high-throughput screen of 242,000 molecules yielded seven chemical series – all of which ultimately proved to be artifacts. The enzyme is a threonine protease but contains a cysteine residue near the active site, so the researchers conducted a Tethering screen with 1280 disulfide-containing molecules, which led to 64 hits in five classes. Converting the disulfide to more drug-like warheads ultimately led to nanomolar molecules with cell-based activity, and the researchers even had some success removing the warhead entirely.

If you missed the meeting, you still have time to catch what should be an epic conference: FBLD 2018 returns to San Diego where it originated ten years ago. People still talk fondly about that meeting, so don’t miss this one!

Poll results: the modern fragment library

Our most recent poll has just closed, and the results provide a snapshot of fragment libraries in 2018.

The first two questions asked about the smallest and largest fragments (in terms of non-hydrogen or heavy atoms) you would include in your library. As shown below, the median lower bound is 7-8 heavy atoms, while the median upper bound is 17-18 heavy atoms. This is comparable to what we saw when we last asked these questions five or six years ago.

(Methods note: of the 98 respondents to the question “What is the largest number of heavy atoms you would allow in a fragment,” 10 answered ≤ 10 atoms, but no one answered 11-12. I suspect that the 10 answers were erroneous and that they actually meant to answer the second question, “What is the smallest number of heavy atoms you would allow in a fragment,” which was answered by 89 respondents. Therefore I have excluded these answers from the figure. Please leave a comment if I’ve mischaracterized your vote!)

The third question asked how many fragments people have in their libraries, and more than 40% of the 99 respondents answered 1001-2000.

The distribution is similar to the results from four years ago when we last asked this question. Notably though, the number of respondents with very large libraries has more than doubled, admittedly from a small base.

Overall then this poll could probably be summarized as, “plus ça change.” Or, as the growing number of clinical success stories attests, if it ain’t broke, don’t fix it!

Fragment growing and merging: an inverse agonist of RORγt2

The nuclear receptor transcription factor RORγt2 is involved in the differentiation of Th17 cells, and is thus a target for inflammatory diseases. The protein contains a large, hydrophobic ligand binding site, and as a result most known inverse agonists have less than ideal physicochemical properties. In a paper recently published in J. Med. Chem., Samuel Hintermann and colleagues at Novartis have taken a fragment-based approach.

The researchers screened a library of 1408 fragments using a “differential static light scattering (DSLS)” assay, which is a type of thermal shift assay that measures denaturation and aggregation of the protein. A few dozen molecules that stabilized RORγt2 were tested in dose-response curves prior to crystallization trials, ultimately yielding 13 structures. Compound 1 was particularly interesting because it binds in the center of the cavity, providing growth vectors in two directions. It also makes a couple hydrogen bonds with the protein, as opposed to purely hydrophobic interactions.

Growing from the ethoxy position quickly led to improvements in affinity. To avoid the possibility of toxic iminoquinone metabolites, the researchers replaced the central phenyl ring with a pyridine, resulting in the low micromolar inverse agonist compound 8.

To further improve affinity, the researchers merged an element from a previously reported GlaxoSmithKline molecule (compound 2) onto compound 8, resulting in the potent compound 9, which was characterized in a battery of assays.

The crystal structure of compound 9 bound to the protein revealed that the core fragment moiety binds in the same manner as the original compound 1, though the added benzyl ether binds in a subpocket that had not previously been observed to bind ligands.

Kinetic studies using a reporter displacement assay revealed that compound 9 has both a slow on-rate as well as a slow off-rate, consistent with the fact that it is fully enclosed by the protein. The researchers performed molecular dynamics simulations to try to determine how the ligand could enter or leave, which suggested large conformational changes in a flexible region of the protein. Isothermal titration calorimetry (ITC) showed that the binding of compound 9 is enthalpically driven, with an unfavorable entropy. Although interpreting thermodynamics is fraught, this result makes intuitive sense given the hydrogen bonds formed and the fact that the molecule seems to rigidify the protein.

Biophysics is interesting, but of course biology is what was driving the program. Compound 9 is potent in a variety of cell assays and is also selective for RORγt2 over other nuclear hormone receptors. However, it is also mostly insoluble, and although it did show efficacy in a rodent inflammation model, plasma concentrations of compound 9 were highly variable between individual rats, which the authors attribute to poor physicochemical properties.

This is a nice application of fragment growing and merging that demonstrates how difficult it is to find useful leads for lipophilic sites: even with favorable biochemistry and biophysics, the pharmacokinetics are a slog. That said, others have made progress against similarly hydrophobic targets, so it will be fun to watch this story progress.

Conservation of fragment binding modes revisited

A common assumption when growing or linking fragments is that the binding mode will remain the same. This is often the case, but exceptions occur frequently enough to keep life interesting. Last year we highlighted a study that tried to answer the question of when ligands changed their binding mode by analyzing the protein data bank (PDB). In a new J. Med. Chem. paper, Esther Kellenberger and collaborators at Université de Strasbourg and Eli Lilly have conducted an even more exhaustive study.

The researchers considered all protein structures deposited in the PDB between 2000 and mid-2016 solved to at least 3 Å resolution. This yielded 1079 different fragments (MW < 300 Da) and 1832 larger (“drug-like”) ligands, as well as 126 crystallization additives such as buffers and detergents. In comparing the same protein with different ligands, care was taken to remove mutant proteins that could cause a change in binding mode.

This dataset was used to address several questions.

First, how often does the same fragment bind to the same pocket in the same manner? Often a crystal structure will have several different copies of the same protein in the asymmetric unit. In nearly three-quarters of cases, the fragments bound in a similar manner to the different copies. The exceptions often involved protein conformational changes, in some cases due to different crystal contacts.

Second, how often does a fragment maintain its binding mode when incorporated into a larger molecule? The data set included 359 pairs of ligands on 51 proteins. Again, about three-quarters of fragments had similar binding modes as their larger counterparts. When binding modes changed, protein flexibility often played a role. Polar contacts such as hydrogen bonds were much more highly conserved than hydrophobic contacts. As the earlier study also found, binding modes of very small fragments (MW < 110) were most likely to change, while fragments with MW > 150 almost always retained their binding modes.

Third, do fragments and larger ligands make similar interactions? The data included 235 proteins in which at least one structure contained a fragment and another structure contained a larger ligand.  (The larger ligand didn’t necessarily contain the fragment.) Obviously larger ligands are able to make more interactions than smaller ligands, but, as Stephen Roughley and Rod Hubbard observed back in 2011, enough fragments should allow you to map out the important interactions. After systematically exploring the data, the current researchers suggest that fully mapping a pocket requires nine or more different fragments, a high bar satisfied by just 11 proteins.

Finally, do crystallization additives behave as fragments? The researchers looked at all additives with MW < 300, and separately considered those bound to otherwise free (apo) proteins and those bound to proteins containing other ligands. In general additives showed more variation in their binding modes, though those binding to apo proteins often made similar contacts as made by fragments and larger molecules. Intriguingly, small polar molecules such as DMSO and glycerol often made hydrophobic interactions with proteins.

There is plenty more in the paper than can be summarized here. Laudably, the researchers have provided all of their data in a convenient web portal that even supports chemical substructure searches. Overall the results reassuringly suggest that the binding mode of a fragment usually remains the same as it is optimized. But of course these types of analyses are subject to survivor bias: fragments that change binding mode unexpectedly may be more difficult to optimize, and thus less likely to lead to larger ligands.

The odds may be ever in your favor, but look out for the exceptions.