Hidden gem of a finding or not?

Todays paper is from a group in Korea.  It's a typical "we did some in silico screening, limited biochemical testing, made a compound or two, and voila!" paper.  In this case, the target is Tyk2 (the target of Xeljanz).

Figure 1.  Xeljanz (tofacitinib)

2000 diverse fragments were selected from the Otava library and docked against Tyk2.  64 top ranked fragments were selected and 9 were selected that had inhibition over 50% at 100 microM, with the best compound (1) having 60% inhibition at 3 microM.  

Figure 2.  Cpd 1 docked to Tyk2. 

What I don't like here is that they didn't do full dose-response curves.  That seems lazy.  Also, the only structures they show are the docked structures.  Maybe its just me, but show me some line drawings.  They then did some limited SAR (3 cpds) based on 1 as the scaffold.  Cpd 12 was the best compound 

Figure 3.  Cpd 12

(10nM IC50).  In the end, 12 was equipotent (or superior) with tofacitinib in terms of shutting down Tyk2/Stat3 signalling.  However, they could not rule out that this is due to non-specific inhibition of other JAK proteins. So, is this a great result?  If so, why BOMCL (not to be snobby)?

Fragments in the clinic: PLX3397

Practical Fragments covers a wide variety of journals. J. Med. Chem., Bioorg. Med. Chem. Lett., Drug Disc. Today, and ACS Med. Chem. Lett. are all well-represented, but we also range further afield, from biggies such as Nature and Science to more niche titles such as ChemMedChem, Acta. Cryst. D., and Anal. Chim. Acta. The increasingly clinical relevance of fragment-based approaches is highlighted by a recent paper by William Tap and a large group of collaborators appearing in the New England Journal of Medicine. This reports on the results of the Daiichi Sankyo (née Plexxikon) drug PLX3397 in a phase I trial for tenosynovial giant-cell tumor, a rare but aggressive cancer of the tendon sheath.

The story actually starts with a 2013 paper by Chao Zhang and his Plexxikon colleagues in Proc. Nat. Acad. Sci. USA. The researchers were interested in inhibiting the enzymes CSF1R (or FMS) and KIT; both kinases are implicated in cancer as well as inflammatory diseases. The team started with 7-azaindole, the same fragment they used to discover vemurafenib. Structural studies of an early derivative, PLX070, revealed a hydrogen bond between the ligand oxygen and a conserved backbone amide. Further building led to PLX647, with good activity against both CSF1R and KIT. Selectivity profiling against a panel of 400 kinases revealed only two others with IC₅₀ values < 0.3 µM. The molecule was active in cell-based assays, had good pharmacokinetics in mice and rats, and was active in rodent models of inflammatory disease.

The new paper focuses on the results of a clinical trial with PLX3397, a derivative of PLX647. Despite its close structural similarity to PLX647, it binds to CSF1R in a slightly different manner. Both inhibitors bind to the inactive form of the kinase, but PLX3397 also recruits the so-called juxtamembrane domain of the kinase to stabilize this autoinhibited conformation. Pharmacokinetic and pharmacodynamics studies in animals were also positive.

Tenosynovial giant-cell tumor seems to be dependent on CSF1R, so the researchers performed a phase 1 dose-escalation study with an extension in which patients treated with the chosen phase 2 dose were treated longer. Of the 23 patients in this extension, 12 had a partial response and 7 had stable disease. A quick search of clinicaltrials.gov reveals that PLX3397 is currently in multiple trials for several indications, including a phase 3 trial for giant cell tumor of the tendon sheath.

Several lessons can be drawn from these studies. First, as the authors note, one fragment can give rise to multiple different clinical candidates. Indeed, in addition to vemurafenib, 7-azaindole was also the starting point for AZD5363. This is a good counterargument to those who believe that novelty is essential in fragments.

A second, related point is that selectivity is also not necessary for a fragment. The fact that 7-azaindole comes up so frequently as a kinase-binding fragment has not prevented researchers from growing it into remarkably selective inhibitors. An obvious corollary is that even subtle changes to a molecule can have dramatic effects: the added pyridyl nitrogen in PLX3397 is essential for stabilizing a unique conformation of the enzyme.

Finally, careful patient selection is critical to answering biological questions. I confess that I had never heard of tenosynovial giant-cell tumor, nor the role of CSF1R, but I’m glad others had. I look forward to seeing an increasing stream of fragment papers in clinical journals.

LO-MS...Coming of Age.

There are many ways to screen for fragments.  One of the really emerging areas uses mass spec detection: WAC, native mass spec, HDX, and ligand-observed MS.  The group that we highlighted back in March has a new paper where they look at the method in terms of accuracy of Kd determination and compare the results to other biophysical methods.  

Their previous work they looked at relative affinity ranking of bound fragments.  In this study, they compared the accuracy of Kd determination in this method to ITC.  First they used a pool of 3 or 4 known CAI inhibitors and a pool of 50 fragments (from their collection).  The conditions were defined:

The hCAI protein was incubated with each inhibitor mixture in the binding buffer with a total volume of 50 mLat room temperature for 40 min. The protein concentration was maintained at 25 mM and the inhibitor concentration increased from 1 mM to 50 mM. The control was prepared by using the binding buffer substitute for hCAI during incubation. The incubation solution was then filtered through a 10 kDa MW cutoff ultrafiltration membrane by centrifugation at 13,000 g for 10 min at 4 C followed by a quickwash with 10mM ammonium acetate (pH 8.0) to remove the unbound compounds.

 Two different methods were used to calculate Kd: 1. saturation curves and 2. measuring the unbound fraction of ligand .  Method 1 was deemed unsuitable for determining Kds for ligands with largely different Kds.  Method 2 did not depend on saturation curve fitting, instead using a calibration curve and did not observe any fragment competition at higher P:L ratios (6:1 or 8:1).  This approach also improved the sensitivity of the assay allowing better detection of lower affinity ligands.  Kds determined using this method matched those determined by ITC.  

To further test the method, they ran a pool of 50 fragments against HCV RNA polymerase NS5B.  This gave, as expected, a complicated chromatographic baseline.  To exclude promiscuous binders, they ran against BSA in parallel.  Eight fragments in the mixed pool showed selective binding to NS5B using unbound fraction analysis vs. 2 from the bound fraction analysis.  7 of 8 fragments were confirmed by SPR (ITC could not be used).  The 1 BFA fragment to be analyzed by SPR showed that it was a very weak binder.  

This sort of work makes me happy.  Methods always need to be pushed and evaluated.  When evaluating methods, if this sort of cross validation hasn't been done, question as to why not? 

Fragments vs Trypanosoma cruzi spermidine synthase, allosterically

Chagas disease, caused by Trypanosoma cruzi, is spread throughout Central and South America by a nasty blood-sucking insect. A couple drugs are approved to treat it, but they can cause severe nausea and peripheral neuropathy, so there is room for improvement. In a recent paper in Acta Cryst D., Yasushi Amano and colleagues at Astellas Pharma describe their efforts against the T. cruzi spermidine synthase (TcSpdSyn).

TcSpdSyn transfers an aminopropyl moiety from the cofactor decarboxylated S-adenosylmethionine (dcSAM) to the evocatively-named putrescine (1,4-diaminobutane) as one step in the synthesis of an essential antioxidant. Small amines can bind in the putrescine-binding pocket and inhibit the enzyme with low micomolar activity, so the researchers decided to find other fragments that could bind in this pocket. They screened in the presence of dcSAM, using surface-plasmon resonance (SPR), with each fragment present at 0.25 mM, as well as in thermal shift assays, with each fragment present at 2 mM. Although nothing is reported about library size or hit rate, hits from either assay were taken into crystallography, resulting in six structures described in detail and deposited in the Protein Data Bank (pdb).

Two fragments were found that bind in the putrescine-binding pocket, and in both cases the enzyme shows some conformational changes to accommodate the fragments. Although these two fragments have only modest potency (IC₅₀ = 0.18-0.48 mM), they do have satisfying ligand efficiencies, and are good starting points for structure-based design.

Unexpectedly, the other four fragments bound not in the putrescine-binding pocket but at an interface between two proteins of TcSpdSyn, which forms a homodimer. One of these fragments, an isothiazolinone, showed mid-nanomolar activity in a functional assay. Readers may recall a paper we pilloried earlier this year which also reported an isothiazolinone as a screening hit. In that case, the researchers failed to recognize that this PAINS compound has ample precedent for reacting with thiols. Happily, in the current paper the researchers are not only aware of this, they actually see covalent bond formation between the fragment and a cysteine residue in the crystal structure. Interestingly though, the fragment reacts with only a single cysteine residue at the dimer interface, despite the presence of six other cysteine residues in the protein.

The researchers carefully analyzed this structure and found that binding of the fragment disrupts the putrescine-binding pocket; in other words, the fragment is an allosteric inhibitor. Moreover, the other three fragments that bind at the dimer interface also appear to act allosterically, and one of them is a single digit micromolar inhibitor.

This is a nice example of how even PAINS compounds can be useful if they are well-characterized and not hyped. Moreover, the structures suggest new approaches for tackling a target for a neglected tropical disease, either covalently or more conventionally.

Magic Methyl and SBDD

As sites have closed down, we have seen a fair number of paper come out describing work at various sites.  Today's paper is another of these, from Roche, Nutley.  The target is Tankyrase, blogged previously here.  

This group started with a biochemical screen of their in house fragment library (here for analysis of their library) against both TNKS1 and 2.  This screen resulted in two compounds: 1. a pyranopyridone and 2. a benzopyrimidone (that looks like the first published TANK inhibitor).  

Figure 1.  Fragments identified from biochemical screen.

They were able to co-crystallize 1 with the TANKS2 and then were able to model 2's binding based upon the known inhibitor.  This pleasingly revealed both similarities and differences in the binding.  the main recognition elements are largely the same.  A big difference is that the phenyl maintains hydrophobic interactions that the cyano group does not. 

Figure 2.  TANKS2 X-ray structure with 1 (orange carbons) and 2 (yellow carbons). 

This lead to the obvious decision to merge the two fragments leading to fragment 3. This compound was reasonably potent (320 nM), however it should moderate to high clearance in an in vivo PK study. 

Compound 3. 

Then of course, the medchem kicks in aiming to decrease cLogP and increase solubility by focusing on two areas: the fused phenyl ring and the isopropyl side chain. This work was able to achieve their goals, but they also stumbled on a "magic methyl" (9) which improved potency 100 fold.  

9.  Magic methyl on fused phenyl ring.

This magic methyl works by sliding into a little pocket defined by three tyrosine residues. 

Figure 3.  Magic methyl binding mode. 

The t-butyl alcohol derivative of 9 had excellent properties: reasonable solubility, excellent permeability, stopped axin degradation in cells in a dose dependent manner, prevented mRNA production of beta-catenin dependent genes, and in mouse had satisfactory in vitro activity and PK profile.  

This is an excellent example of FBLG.  The magic methyl does exist and X-ray is a highly enabling technology. 

What works for crystallography?

As a recent post emphasized, crystallography is a key technique for fragment-based lead discovery. We’ve occasionally touched on things that can go wrong in crystallography, but in a recent paper in Drug Discovery Today, Helena Käck and colleagues at AstraZeneca (Mölndal) put things in a more positive light by asking what factors lead to success.

The paper starts with a literature review of successful fragment structures published between 2012 and 2014 and summarizes some of the key findings. First is the need to easily generate robust crystals that diffract well and are stable for long periods of time. If the ligand-binding site is known, it is important that this is accessible and not occluded by protein or ligands. Finally, the crystals should be stable when soaked in high concentrations (> 10 mM) of ligand, ideally in the presence of 10% DMSO.

None of these factors will come as a surprise to experienced crystallographers, but the authors do a nice job of concisely summarizing them as well as providing solutions to common problems. For example, the use of surrogate proteins can help in cases where the target itself is hard to crystallize. Proteins can be grown in the presence of known ligands, which can then be soaked out. And various additives can also help.

All of this is nice, but what really makes this paper noteworthy is the second part, in which the authors discuss their own experience with soluble epoxide hydrolase (sEH), a potential cardiovascular and immune target that we’ve previously discussed here, here, and here. This protein seems to have all the hallmarks of technical success. Indeed, both HTS and fragment screens at AstraZeneca produced high hit rates, and 65% of hits taken into crystallography produced structures. In all, 55 structures were determined, with ligands ranging in size from 130 to 540 Da and affinities ranging between 0.003 and 600 µM. Of these, 38 could be considered fragments. As seen before, the protein is relatively rigid, and the ligands bind in a variety of subsites within the large lipophilic active site.

With so much data, the researches asked whether ligand properties could predict crystallographic success. The most robust correlation was seen with affinity: 94% of compounds with affinities below 0.1 µM produced structures, while only 36% of compounds with affinities above 100 µM did.

Ligand efficiency (LE) was also correlated with crystallographic success, though three small fragments (MW < 160 Da) with very high LE values did not produce structures – a phenomenon which has been noted by others.

In contrast to another recent study that compared many fragment-screening approaches, solubility did not predict success. The researchers suggest that this is because crystal conditions are so different from the conditions under which standard solubility measurements are run.

Admirably, the structures of 52 of the ligands are reported in the supplementary material – along with their measured affinities – and the resulting crystal structures have been deposited in the protein data bank. Some of the ligands bind at multiple sites and some have dual conformations; these ambiguities are noted. Moreover, a set of inactive analogs has also been included. Together with smaller sets of previously released structures, this provides a bonanza of structural and affinity data with which to benchmark computational docking programs. Hopefully we’ll see more of this public sharing of data.

Fragment finding smackdown: 2015 edition

Our current poll (right-hand side of page) asks about NMR. But of course, there are lots of other ways to find fragments, and the question often arises as to which ones are best. This is the subject of a recent paper in ChemMedChem by Gerhard Klebe and collaborators at Philipps University Marburg, Proteros, NovAliX, Boehringer Ingelheim, and NanoTemper.

Long-time readers will recall that the Klebe group assembled a library of 361 fragments, some of which violated strict “rule of 3” guidelines. These were screened in a high-concentration functional assay against the model aspartic protease endothiapepsin, resulting in 55 hits, of which 11 provided crystal structures. The authors wondered how other techniques would fare. In the new paper, they retested their entire library against the same protein using a reporter displacement assay (RDA), STD-NMR, a thermal shift assay (TSA), native electrospray mass spectrometry (ESI-MS), and microscale electrophoresis (MST). To the extent possible they tried to use similar conditions (such as pH) for the different assays, though the fragment concentrations ranged from a low of 0.1 mM (for ESI-MS) to a high of 2.5 mM (for TSA), while protein concentrations ranged between 4 nM (for the biochemical assay) to 20 µM (for ESI-MS).

All told, 239 fragments hit in at least one assay – a whopping hit rate of 66%. Actually, the number is even higher since, for various reasons, not all fragments could be tested in all assays. And yet, not a single fragment came up in all of the assays! Overall agreement was in fact quite disappointing, with most methods having overlaps of less than 50%, and often below 30%. This is in contrast to a study from a different group highlighted a couple years ago.

What’s going on? One clue might be the solubilities, which were experimentally measured for all library members. In general, hits tended to be more soluble than the library as a whole, emphasizing the importance of this parameter not just for follow-up studies but for identification of fragments in the first place.

Another possibility is that some fragments bind outside the enzyme active site, and thus would not be picked up in a biochemical assay or the RDA. Some evidence for this is provided by follow-up NMR studies in which hits were competed with ritonavir, which binds in the active site. Ritonavir-competitive binders shared greater overlap with biochemical and RDA hits, while there was more overlap between ritnovair-uncompetitive binders and hits from methods such as ESI-MS, TSA, and MST that rely solely on binding. (This could also explain similar observations made earlier this year.)

If a picture is worth a thousand words, how many of the 11 hits that had previously yielded crystal structures would have been identified had they been tested in other methods? Here the numbers vary significantly, from 27% for ESI-MS and MST to 100% for NMR, though these statistics should be taken with a grain of salt since – for example – only 7 of the 11 crystallographically-confirmed hits could actually be tested in the NMR assay. Also, it is possible that some hits from these methods might have generated new crystal structures for fragments not identified in the initial biochemical screen.

One admirable feature of this paper is that the authors provide all their data, including structures and measured solubility numbers for each component of their library. This should provide an excellent dataset for a modeler to use in benchmarking computational methods.

All in all this is a thorough and important analysis and a sobering reminder that, even if a fragment doesn’t hit in orthogonal assays, that doesn’t necessarily mean it’s not a useful starting point. On the other hand, artifacts are everywhere, and paranoia is often justified. The art is deciding which hits are worth pursuing – and how.

Uninteresting GPCR Fragment Work...meant as a Compliment!!

There are certain movies that when they are on TV, I can't not watch.  I call these Broken Leg Movies (as in if I were laid up with a broken leg what would I watch).  As I have said, Road House is one, Apollo 13 another.  Its about America's blase attitude towards the amazing feat of putting men on the moon.  It takes a potential horrific tragedy (For those off you who haven't seen it, let me say (**Spoiler Alert**) don't worry it has a happy ending.) in order for America to care about men in space.  Which of course is in direct contrast to Pigs in Space (with Swedish Subtitles)! 

One of the field changing technologies is Heptares' STAR technology (for creating stabilized, soluble GPCRs).   We have discussed it often on this blog.  Well, they are back with another paper, this time working the voodoo they do so well on a Class C GPCR.  Negative allosteric modulation of the mGluR has the potential for significant medical impact in a variety of diseases.  In a relatively well trod drug space (there have been several molecules in late stage trials), an issue appears to be the acetylenic moiety in these drugs (which appears to be manageable).  So, non-acetylenic molecules would be desireable.  

To attempt to ligand this molecule, they screened 3600 non-acetylenic fragments using a radio-labeled assay.  This is in contrast to previous work where they used SPR. From this screen, 178 fragments were tested in concentration-response curves leading to "a number of promising" hits, including the compound shown below. 

Cpd 5.  pKi=5.6, LE=0.36

This compound was advanced using the tools you would expect (especially from Heptares): modeling, X-ray crystallography, medchem, and so on.  The final molecule is an advanced lead with excellent mGluR selectivity and in vivo activity, clean tox, and so on. 

This is excellent work, but "yawn".  I think it might be interesting to hear why they went with the radioligand approach, as opposed to SPR.  You could quibble that 5 is too big to be a fragment, but really?  Papers like this are uninteresting, we know its going to work.  The science is excellent, but I want to see the triumph out of tragedy.  Not here.  I want to congratulate Heptares for making an achievement like this paper perfectly uninteresting.  And I mean uninteresting as the very best of compliments. 

Aggregation alert

Practical Fragments has quite a few posts about PAINS, or pan-assay interference compounds. In part this reflects their sad prevalence in the literature, but it’s also fair to say that they are easy targets because many are readily recognizable.

But not all artifacts are so easily spotted, as discussed in a new paper just published in J. Med. Chem. by John Irwin, Brian Shoichet, and colleagues at the University of California San Francisco (see also here for Derek Lowe's excellent summary).

The researchers took on one of the most insidious problems, compound aggregation, in which small molecules form colloids that bind to and partially denature proteins, causing false positives in all sorts of assays. This can happen even at nanomolar concentrations of compound, and is all the more problematic at higher concentrations used in fragment screening and early hit to lead optimization. In many cases aggregates can be disrupted or passivated by including nonionic detergents such as Triton X-100 or Tween-80, but not all assays tolerate detergent, and some aggregates form even in the presence of detergent.

Worse, all sorts of molecules can form aggregates, including many approved drugs. Previous attempts to try to predict which molecules will aggregate have not been very successful. Colloid formation is essentially a phase transition, and like other such transitions (crystallization, for example) it is fiendishly difficult to predict what molecules will do this under what conditions. But if we can’t predict from first principles which molecules will form aggregates, can we at least draw empirical lessons?

The researchers assembled a set of >12,600 known aggregators and put together a very simple model that assesses how similar a molecule of interest is to one of these aggregators (using Tanimoto coefficients, or Tcs). Aggregators have a wide range of physicochemical properties, with ClogP values from -5.3 to 9.8, but 80% have ClogP> 3.0. The team hypothesized that a molecule sufficiently similar to a known aggregator – and also somewhat lipophilic – would have a higher probability of being an aggregator than a molecule chosen at random.

To test this idea, the team took  a batch of 40 molecules and tested them for aggregation. Among those most similar to known aggregators (Tc ≥95%), 5 of 7 molecules were confirmed as aggregators. This fell to 10 of 19 for the next set (Tc 90-94%), 3 of 7 after that (Tcs 85-89%) and only 1 of 7 for the least similar (Tcs 80-84%). Thus,Tc ≥85% was chosen as the cutoff.

Next, the researchers examined molecules that had been reported as active in some sort of biological assay, and found that 7% were ≥85% similar to a known aggregator and had ClogP> 3. Ominously, this rate is an order of magnitude greater than the number of commercially available compounds that also fit these criteria. More damning, most of this enrichment has occurred since 1995, when high-throughput and virtual screening really went mainstream. In other words, the past couple decades have seen a sizable enrichment of potential aggregators in the literature.

All of this is fascinating, but what really makes this paper significant is that the researchers have made all their primary data available, and also built a simple to use website called “Aggregator Advisor”. Just draw your molecule or paste a SMILES string to generate a report. For example, entering gossypol tells you that this molecule has previously been reported as an aggregator. (With two catechol moieties, it’s also a PAINS.) Perhaps not coincidentally, it shows up in more than 1800 publications.

Of course, as the researchers note, “just because a molecule aggregates, under some conditions, in the same concentration range as it is active, does not establish that its activity is artefactual.” Indeed, 3.6% of FDA-approved drugs are known aggregators. Still, particularly if your hit has only modest activity (0.1 µM or worse), similarity to a known aggregator should at least make you cautious.

The researchers are at pains to emphasize that their model is “primitive and subject to false negatives and false positives.” Thus, any hits need to be tested to see if they behave pathologically in any given assay. More importantly, a molecule that comes up as a negative should not be presumed to be innocent.

All these caveats aside, Aggregator Advisor is very easy to use. It’s certainly worth running the next time you find an interesting molecule – whether in your lab or in the literature – particularly if there was no detergent in the assay.