Fragments vs renin: growing this time

Renin is a key player in the regulation of blood pressure, and thus an important therapeutic target for hypertension. Indeed, the approved drug aliskiren is a renin inhibitor. However, this drug has very low oral bioavailability as well as other problems – surely something better could be developed? This was the goal of a team of researchers at Takeda, described in two recent papers in Bioorg. Med. Chem.

One challenge with renin is that it is an aspartic protease with a large active site – similar to the difficult target BACE1. Like BACE1, fragment-based approaches proved to be useful. In the first paper, Michiko Tawada and colleagues conducted an enzymatic screen (at 100 µM) of their fragment library. Although this library contained many positively-charged fragments – which would be expected to interact with the negatively charged catalytic aspartic acid residues – none came up as hits. Neutral compound 1, however, was identified, and crystallography revealed that it binds in the hydrophobic S1, S3, and S3^sp pockets. Novartis researchers published a similar experience several years ago.

Compound 1 was poorly soluble, lipophilic, cytotoxic, and offered suboptimal vectors for fragment growing, so the researchers sought an alternative by constructing and testing a library of analogs. Compound 4a had a similar affinity to the initial fragment, and crystallography revealed a similar binding mode. This was used as the core of a second library, leading to compound 6b, which also displayed a similar binding mode to the initial fragment. Although the affinity was similar (and indeed, the ligand efficiency was lower), the new fragment had better physicochemical and biological properties. It was also more synthetically tractable for subsequent optimization, which is the focus of the second paper.

As with the initial fragment, compound 6b did not make interactions with the catalytic aspartic acid residues, though they are nearby. By redesigning the compound and introducing a basic nitrogen in compound 7, Yasuhiro Imaeda and colleagues were able to engage these residues. Also, the crystal structure of compound 4 (top) revealed that a hydrophobic substituent would be tolerated, which led to compound 9, with nanomolar affinity. Further growing into a hydrophobic pocket led to compound 14, with high picomolar activity. This compound was active in human plasma, showed excellent selectivity against other aspartic proteases, and exhibited encouraging bioavailability and pharmacokinetic properties. The paper notes that this molecule has been optimized further.

For me, the most striking lesson from these two papers is how much effort it took to improve the potency of the initial fragment hit: lots of analogs were made without notable improvements, and it would have been easy to give up. But in the end, a combination of managed serendipity and careful structure-based design increased the affinity by 74,000-fold to a promising lead. Something to keep in mind the next time you find yourself lost in a forest dark, where all paths seem to lead to dead compounds.

Fragments vs secreted phospholipase A2: AZD2716

Many success stories were presented at the recent FBLD 2016 meeting in Boston, some of which are appearing in the literature. A case in point is published in this month’s issue of ACS Med. Chem. Lett.

Fabrizio Giordanetto, Daniel Pettersen, and colleagues at AstraZeneca were interested in finding inhibitors of secreted phospholipase A₂ (sPLA₂) enzymes, which cleave glycerophospholipids and are implicated in the lipid accumulation and inflammation associated with atherosclerosis. Of the eleven different isoforms, sPLA₂-IIa, sPLA₂-V, and sPLA₂-X are considered particularly good targets, and the researchers sought an inhibitor that would hit all three. Other companies had shown that a primary amide can form multiple hydrogen bonds at the catalytic site, so the AstraZeneca team reanalyzed previous internal screening data to look for fragment-like hits (defined as having 10-18 non-hydrogen atoms) containing a primary amide. They found many, including compound 1.

In addition to being a potent inhibitor of both sPLA₂-IIa and sPLA₂-X, compound 1 was quite active in human plasma, which is physiologically relevant. A crystal structure of  sPLA₂-X revealed that the compound bound as expected, and modeling suggested that adding a carboxylic acid moiety could make additional interactions with the catalytic calcium ion. Several molecules were made, the most potent of which turned out to be compound 4, with a satisfying 2000-fold boost in activity against sPLA₂-X. Shortening or lengthening the linker connecting the acid with the rest of the molecule reduced affinity, observations which could be rationalized by modeling.

Compound 4 was characterized in some detail, which revealed bioavailability in rats and dogs, good pharmacokinetics, and a fairly clean off-target profile. Unfortunately, it was reasonably active against OATP1B1, which recognizes carboxylic acids. Among other duties, OATP1B1 transports statins to the liver, and since many people with atherosclerosis are taking statins this activity would obviously be a problem. However, crystallography suggested that introducing substitutions very close to the carboxylic acid moiety would likely be tolerated by sPLA₂ but not by OATP1B1. Indeed, simply adding a methyl group maintained or increased activity against the three relevant sPLA₂ isoforms while completely abolishing OATP1B1 inhibition. Happily, AZD2716 had excellent pharmacokinetics and bioavailability in mice, rats, dogs, and cynomolgus monkeys.

This is a lovely example of what has been called fragment-assisted drug discovery. The researchers explicitly looked for a small, ligand-efficient starting point and relied heavily on structure-based design during optimization. The paper ends by noting that AZD2716 was selected as a clinical candidate, though it does not appear in the AstraZeneca pipeline or in clinicaltrials.gov; if this changes we’ll make a note on our running list.

At FBLD 2016 Jenny Sandmark presented this story, and she also described another compound derived from a different fragment. This turned out to be selective for sPLA₂-X over sPLA₂-IIa and sPLA₂-V, and was therefore deprioritized. The experience working with this earlier series was, however, useful in guiding the discovery of AZD2716 – a reminder of the importance of having multiple good fragment starting points.

FBLD 2016

Last week the sixth FBLD meeting was held in Cambridge, MA. Like its predecessors in 2014, 2012, 2010, 2009, and 2008, this meeting was an enormous success, mixing more than 230 scientists with excellent (and liberal) food and drink. With 33 talks, more than 30 posters, and several vendor booths and workshops I won’t be able to do more than capture a few highlights.

The most striking feature for me was the number of success stories. This began with Steve Fesik’s keynote lecture, in which he discussed the MCL-1 inhibitors he and his team at Vanderbilt have discovered. When we highlighted his work last year he had reported low nanomolar inhibitors, but these did not have cell-based activity. His group has now optimized the molecules to low picomolar biochemical potency, low nanomolar cellular activity, and good activity in mouse xenograft models. This has not been easy: more than 2210 compounds were made, guided by 60 X-ray structures and dozens of pharmacokinetic experiments. It seems to be paying off though, and the researchers are developing biomarkers with the goal of advancing a compound into clinical testing.

Two other notable success stories about clinical candidates must be mentioned, though I’ll wait until publications come out before going into detail. Kathy Lee described how she and her colleagues at Pfizer chose a fragment that was less potent and ligand-efficient than other hits due to its interesting binding mode and were able to advance it to PF-06650833, an IRAK4 inhibitor with potential for inflammatory diseases. And Wolfgang Jahnke discussed how he and his colleagues at Novartis were able to discover and advance ABL001, an allosteric inhibitor of BCR-ABL, despite having the project halted twice – a reminder that persistence is essential.

Several other success stories have been covered at least in part on Practical Fragments, including inhibitors against PDE10A (presented by Izzat Raheem of Merck), Dengue RNA-dependent RNA polymerase (presented by Fumiaki Yokokawa of Novartis), lipoprotein-associated phospholipase A2 (presented by Phil Day of Astex), and BACE1 (presented by Doug Whittington of Amgen).

Crystallography was another theme, and several of the success stories relied on crystallographic fragment screening. Frank von Delft of the Structural Genomics Consortium described developments that allow screening 1000 crystals per week at Diamond’s Xchem facility in the UK, which include acoustic dispensing of compounds into crystallization drops – while carefully avoiding hitting the crystals head-on.

Several computational talks reported results that run contrary to conventional wisdom. Vickie Tsui of Genentech discussed their CBP bromodomain program (which we recently discussed here). Several water molecules form a highly ordered network in the protein, and a WaterMap analysis suggested that these were high-energy and that displacing them would lead to an enhancement in activity. Unfortunately this turned out not to be the case, though the researchers were able to get to low nanomolar inhibitors by growing towards a different region of the protein.

Li Xing mined the Pfizer database of 4000 kinase-ligand structures to extract 595 unique hinge binders. Not surprisingly, some of these – such as adenine and 7-azaindole – bound to multiple kinases, but 427 were complexed to just a single kinase. Hinge binders typically form 1 to 3 hydrogen bonds to the protein, and while there didn’t seem to be a correlation between the number of hydrogen bonds and potency, more hydrogen bonds did correlate – perhaps counterintuitively – with lower selectivity. To the extent that hydrogen bonds are thought of as enthalpic interactions, this further muddies the argument that enthalpy and entropy can be useful in drug design.

On a more positive note, Sandor Vajda (Boston University) suggested that, according to analyses done in FTMap, perhaps 60-70% of protein-protein interactions may be druggable – as long as we accept that this may require building larger molecules than commonly accepted. And Chris Radoux (Cambridge Crystallographic Data Centre) discussed the computational tool for characterizing hotspots that we previously covered here; a web server for easy search should be available soon.

Library design was also a key topic. Richard Taylor of UCB described his analysis of all FDA-approved drugs, which revealed >350 ring systems. Interestingly though, 72% of drugs discovered since 1983 rely exclusively on ring systems used prior to that date. Clearly there is plenty of untapped chemical real estate.

But getting there won’t necessarily be easy. David Rees stated that 33 fragments recently added to the Astex library required 13 different reaction types. Importantly, many of the fragment to lead successes at Astex have required growing the fragment from the carbon skeleton rather than from more synthetically tractable heteroatoms. Knowing in advance how to do this with every new member of a fragment library should make life much easier in the long run, though it is a serious challenge for chemists.

There is far more to write about, including a great discussion led by Rod Hubbard on how FBLD is integrated effectively into organizations and how it enables difficult targets, but in the interest of space I’ll stop here. If you were at FBLD 2016 (or even if you weren’t) please share your thoughts!

Tips for high-throughput crystallography

X-ray crystallography is tied for second place among methods used in fragment-based lead discovery, according to our most recent poll. This makes sense, since structures are usually essential for advancing fragments to leads. Faster fragment-finding methods are usually used to triage fragments down to a manageable number of hits to feed into crystallography, but the high incidence of false negatives means that promising fragments might be inadvertently discarded. If structures are key goals at the end of a fragment screening campaign, why not start directly with crystallography?

In fact, this is exactly what more and more groups seem to be doing. The problem, historically, has been throughput. Increasing automation has been solving some of the mechanical issues (such as mounting crystals and collecting data at a synchrotron), but what about the actual processing? A recent paper in Structure by Andreas Heine and collaborators at Philipps-University Marburg and Helmholtz-Zentrum Berlin für Materialien und Energie provides some useful advice.

The protein in question is endothiapepsin, a model aspartic protease that is easy to crystallize and diffracts to high resolution. Earlier this year, we discussed the researchers’ work soaking 360+ fragments against this protein, and a companion paper gives detailed information on how several dozen fragment hits bind. The Structure paper describes an automated refinement pipeline, and highlights some of its most important features.

Determining a crystal structure involves iterative cycles of modeling the protein backbone and side chains into regions of “electron density.” One risk is “model bias,” illustrated memorably in this brief video. This is especially important for small molecules: since they represent such a tiny fraction of the overall structure, it is especially easy to see what you want to see. To avoid this, people often look for regions of electron density – which in addition to a bound small molecule could represent co-solvents, buffer, or an amino acid side chain that has unexpectedly moved – before doing much refinement.

The problem is that the electron density might be very spotty and easy to overlook. This is especially true for fragments that bind weakly and which are small by definition. Some initial refinement can thus improve the quality of the electron density maps. The researchers find that adding water molecules and including these in the refinement is the single most important step. Adding bound hydrogen atoms to the protein model is also helpful: even though each hydrogen only contributes one electron to the overall density, there are more than enough to make a meaningful difference. Finally, for very high resolution structures (better than 1.5 Å), it can help to treat each atom of the protein individually (anisotropic refinement of B factors, or atomic displacement parameters). However, at lower resolution, doing this can lead to overfitting. Incorporating these steps into the automated process revealed that 25% of fragments would have been missed had conventional methods been used.

The paper includes lots more detail that will be of interest primarily to crystallographers. Moreover, the data for all 364 fragment soaks has been uploaded to the protein data bank. This is a very high-quality data set: all the crystals diffracted to better than 2.0 Å resolution, with the mean being 1.35 Å, and should be a useful resource for those of you establishing your own automated processing system.

Poll results: affiliation, metrics, and fragment-finding methods

The latest poll has just closed, and the results are quite interesting – I’ll get to these in the next paragraph. First, a quick note on methodology. The poll ran from August 27 through September 30. Due to issues with polling in Blogger, we began running polls in Polldaddy in 2013; its interface gives the total number of votes for a question but not the number of individual respondents. Thus, for the questions on metrics and methods, I assumed that the number of respondents was equal to the number of people who identified themselves as practicing FBLD in the first question, or 123 out of a total of 154. The true percentages for the metrics and methods that people use could be higher or lower if not everyone answered all the questions.

Readership demographics have been remarkably stable since 2010 and 2013, with just over half of respondents from industry, and around 80% of all respondents actively practicing FBLD.

The next question asked about screening methods, and here things get more interesting.

The first thing to notice is that, as we also saw in 2013, nearly all fragment-finding techniques are being used more, with the average user employing 4.1 distinct methods today compared with 3.6 in 2013 and 2.4 in 2011. Ligand-detected NMR has jumped to first place in terms of popularity, with SPR and X-ray crystallography tied for second, followed closely by thermal shift. MST, while still in the minority, has had the largest percentage increase. The use of crystallography has certainly jumped since 2011, which fits with recent publications.

Finally, with regards to metrics, ligand efficiency (LE) continues to dominate, followed by LLE (or LipE), though overall usage of both is down compared with 2014. Only one of the other metrics broke the 10% mark. 

Again, if some practitioners answered the first question of the poll, but not the next two, the use of all methods and metrics could be underestimated. Still, these results seem to fit with what I’ve heard talking with folks – any surprises?