Choosing fragments and assays

One of the advantages of running lots of fragment screens is that it generates lots of data that you can mine for general trends and insights. Astex and Vernalis have both done this; in a paper just published online in J. Biomol. Screen. Peter Kutchukian and (former) colleagues at Novartis provide their own meta-analysis of 35 fragment screening campaigns.

The Novartis fragment library consists of 1400 fragments with molecular weights ranging from 102 to 306 Da and logP values from -2.19 to 3.9. This library has been screened against dozens of targets using a variety of different methods. The researchers looked at the hit rates and used Bayesian methods to try to answer three broad questions.

What makes a fragment amenable for fragment-based screening?

Many people have found that some fragments hit many targets while others hit none, and the results here are no different. Over a set of 20 targets, only 37% of fragments came up as a hit, as opposed to the 54% that would be expected if the odds of hitting a target were the same for all fragments (and using the hit rates actually observed). Correspondingly, some fragments hit more targets than expected. Indeed, 1.4% hit six or more, which is orders of magnitude more than would be expected by chance. Given justifiable concerns about artifacts, one might be tempted to dismiss these hits, but the researchers found that these frequent hitters turn out to be more likely to generate crystal structures than other active fragments. In other words, these are privileged fragments (think 7-azaindole).

Do these privileged fragments have anything in common? Previous work from Astex and Vernalis has suggested that fragment hits tend to be slightly more lipophilic than non-hits, and this trend is all the more apparent here. In fact, fragments that hit more than five targets had a median logP of 2.47 versus 1.45 for fragments that hit just a single target. Promiscuous fragments also tended to be slightly larger than other fragments, in contradiction to the molecular complexity hypothesis. They also tended to have more aromatic bonds, fewer rotatable bonds, and higher solubility.

How do hits from different fragment screening technologies and target classes compare with each other?

Do different target classes find different sets of hits? An analysis of substructures identified in hits against various target classes suggests the answer is yes. Certain substructures are preferred by kinases, for example, while other substructures are preferred by serine proteases. This suggests that building fragment libraries specific to a target class may be productive, though certainly not essential.

Regarding screening technologies, the researchers examined both biophysical (for example NMR, SPR, and DSF) as well as biochemical (such as fluorescence) assays. In general, the hit rates were similar for different technologies, with two exceptions. In SPR, a number of fragments nonspecifically interacted with the surface of the chip, giving a higher number of false positives. On the other hand, DSF gave an anomalously low hit rate, and on closer inspection the researchers found that about 1% of the fragment library appeared to denature proteins.

Interestingly, there was less overlap of hits between biophysical methods and biochemical methods than among biophysical methods or among biochemical methods. In other words, hits from an NMR (biophysical) screen were less likely to be found in a fluorescence (biochemical) screen than in an SPR (biophysical) screen. This is similar to the results of a previous study, though not stated explicitly there.

What is the best way to pair FBS assay technologies?

Given this finding, the researchers suggest that, to find the greatest number of hits, it is best to pair a biochemical method with a biophysical method. Of course, this assumes that the goal is to find as many hits as possible, but these may come at the expense of false positives. Still, if you’re going after a tough target, you want to find every possible hit you can. And if you are more interested in weeding out false positives than finding every viable hit, choosing fragments that hit in both a biochemical and a biophysical assay is probably a good starting point.

This is a fascinating paper and contains far more data than can be practically summarized here. It will be fun to see whether similar analyses, from different organizations, come to similar conclusions. 

FBLD 2014

FBLD 2014, the fifth in an illustrious series of conferences, took place in Basel, Switzerland last week. Organizers Wolfgang Jahnke (Novartis), Michael Hennig (Roche), and Rod Hubbard (University of York & Vernalis) put together a fantastic event. With 35 talks, 45 posters, and more than 200 delegates, I won’t attempt to give more than a few impressions here. In addition to Teddy’s (and others’) Tweets, Derek Lowe put up several posts (see here, here, and here); please share your thoughts below.

Harren Jhoti delivered a lively and wide-ranging opening keynote summarizing the past 15 years of FBLD as viewed from Astex. Among many other innovations, researchers there are responsible for the Rule of 3, which has been the subject of some debate. Harren emphasized that the “Voldemort Rule” should not be a strait-jacket. Like the Kobayashi Maru, rules are there to be broken, though you need to be something of a James T. Kirk to do so effectively.

Astex has produced what is likely the largest collection of protein-fragment crystal structures, and Harren noted that many proteins appear to have fragment binding sites outside the active site: across 25 different proteins, the average number of total sites is slightly greater than 2. Astex is increasingly targeting these sites for allosteric lead discovery.

The theme of crystallography carried through the conference. As Armin Ruf (Roche) exhorted, “more crystals, more structures.” One challenge is that not all crystal forms are suitable for fragments, and it is not always clear from the outset which forms will work. Armin described their chymase project in which an initial crystal form gave 8 fragment structures, but additional crystal forms allowed them to obtain 6 more. Without the different crystal forms these later fragments would have been crystallographic false negatives, yet the potential of different crystal forms to reveal more hits is under-appreciated: Armin noted that the majority of recent fragment papers reported using only a single crystal form.

The importance of crystallography was emphasized again by Nick Skelton (Genentech), who discussed their NAMPT program (which we covered here). In this project, which utilized dozens of crystal structures, a single atom change to a fragment could completely and unpredictably alter the binding mode.

Obtaining a good crystal is not necessarily easy, though. Andreas Lingel (Novartis) described their efforts to produce a form of B-RAF that would diffract to higher resolution and allow fragment soaks (as opposed to co-crystallization). They tried reducing the “surface entropy” by mutating glutamate and lysine residues to alanine, but only one of a dozen or so mutants expressed well and gave superior crystals. Although this turned out to be useful, the team is still at a loss to explain why the specific mutants are effective.

Continuing the theme of crystallography, Matt Clifton (Beryllium) described what looks to be a significant advance for the protein MCL-1. (This is a collaboration with the Broad Institute, and we previously noted some of their progress here.) The researchers have developed a maltose-binding protein (MBP) fusion of this oncology target that diffracts to 1.9 Å in the absence of any ligand. (MBP fusions are used to help crystallize challenging proteins.) Since they developed this construct in May of this year, the researchers have already solved more crystal structures than had been reported publicly to date, and uncovered some interesting findings. For example, the initial fragment that Steve Fesik’s group found binds in a similar manner to one of his more potent later leads, as does one of the AbbVie fragments; however another AbbVie fragment binds in a somewhat different fashion than the elaborated lead.

The subject of how to effectively sample chemical space was another theme, and to this end Alison Woolford (Astex) proposed the concept of a “minimal pharmacophore”: the minimal interactions necessary to drive fragment binding. Researchers at Astex have systematically cataloged several dozen of these, which include such simple entities as amines, acids, aromatic chlorides, and more abstract concepts such as a 1-bond donor-acceptor (think pyrazole). Alison showed an interesting graph with targets on the y-axis and minimal pharmacophores on the x-axis which revealed some obvious patterns such as the preference of donor-acceptor minimal pharmacophores by kinases, but there were unexpected features as well. In a sense, this is an empirical realization of early docking studies, but it also has interesting implications for library design. For example, Alison suggested avoiding fragments with more than one minimal pharmacophore, as these will not be able to effectively sample as many different sites on a protein: with two pharmacophores, a fragment would be limited to binding sites having matching recognition elements to both rather than just one. This idea ties in with the concept of molecular complexity, but from a more chemocentric point of view.

On the subject of chemistry, Dalia Hammoudeh (St Jude’s Hospital) gave a lovely talk on her experiences developing allosteric inhibitors of DHPS, an antibiotic target. Among other fragment hits from the Maybridge library, one was ostensibly 4-trifluoromethylbenzylamine, but turned out to actually be the Schiff base of this with the corresponding aldehyde. Yet another reminder to always carefully check what you think you have.

Practical Fragments has previously discussed the Genentech MAP4K4 program (here and here), and Terry Crawford gave a nice summary. One of the challenges they faced was that their initial leads had excellent brain penetration, leading to animal toxicity. This forced them to increase size and polar surface area. Although it was problematic in this case, this emphasizes how small and drug-like fragment-derived leads can be. Indeed Vicki Nienaber, who was a prime mover behind the original FBLD 2008 meeting, has devoted much of her efforts at Zenobia to CNS targets.

Finally, Derek Lowe (Vertex) gave a rollicking history of the drug industry, ending with his view of where fragments fit in. He noted that chemists – Valinor not withstanding – play a central role, and in that sense the field is a departure from the general trend of the past decade or so. It still remains to be seen how many of the 30+molecules FBLD has delivered to the clinic will come out the other side, but at least for now the field is thriving. As Chris Lipinski stated last year, “if I had to single out one technology that really took me by surprise and has been very successful, it has been fragment screening.”

Fragment merging for renin

Renin, an aspartic protease involved in regulating blood pressure, is one of those drug targets that has been around forever; it took decades before the first direct inhibitor was approved. In a recent paper in J. Med. Chem., Daniel Baeschlin and colleagues at Novartis (where the approved drug was discovered) describe how they’ve taken a fragment-merging approach to look for additional inhibitors of this target.

The researchers started by assembling a small (113 compound) fragment library designed to target aspartic proteases. This was screened against renin by NMR, resulting in hits such as compound 3. Although these were too weak to yield dissociation constants or IC₅₀ values by NMR or biochemical screens, the researchers were able to obtain crystal structures of at least two of these fragments bound to renin, including compound 3. Interestingly, although the amino alcohol moiety of this compound was designed to target the catalytic aspartic acids, this turned out not to be the case. Instead the binding appears to be largely driven by hydrophobic contacts between the tricyclic moiety of the fragment and the so-called S₃-S₁ pocket of the protein.

Along with the fragment effort, the researchers also undertook an HTS screen, resulting in the discovery of compound 5, which itself had come from a 950-compound library targeted towards aspartic proteases. Crystallography revealed that this molecule binds with the diphenylmethane moiety in a similar position as the tricycle of fragment 3, and indeed when a rigid tricyclic framework was grafted onto compound 5 the resulting compound 9 showed a satisfying boost in potency. Further optimization led to a pure enantiomer of compound 12 with low nanomolar potency, good selectivity, moderate oral bioavailability and efficacy in rats, though it did also show some time-dependent CYP3A4 inhibition.

This is really a structure-based design paper, and there is obviously much more detail than would be appropriate here. What caught my eye is that it is a nice example of fragment-assisted drug discovery, in which fragment information is used as one aspect of an overall lead discovery program. In this case a cynic could argue that the only bit that came from the fragment was the tricylic motif. However, given the limitless number of analogs that could be made, such information can be both unexpected and valuable.

Allosteric FPPS inhibitors – not so negative

The protein farnesyl pyrophosphate synthase (FPPS) has long been a target of drugs for osteoporosis, and some data have suggested it could be a productive target for other diseases too. However, approved drugs that target FPPS contain two phosphonates, highly negatively charged moieties that, while nicely directing the drugs to bone, lead to low plasma and soft tissue levels. Researchers at Novartis have now identified fragments that bind to an allosteric site on FPPS and so lack the phosphonate moieties. An article in the September issue of Nature Chemical Biology describes the discovery of these fragments and how they were advanced to nanomolar inhibitors.

The researchers, led to Wolfgang Jahnke in Basel, Switzerland, started by screening a library of only 400 fragments using NMR. Several low affinity (millimolar) hits such as compound 1 were identified, but surprisingly these were not competitive with bisphoshonate drugs, and some even bound synergistically. Crystallography revealed that they were binding in a previously undiscovered allosteric site. (The same group also used fragment screening to explore an allosteric site in another protein.)


To follow up on these observations, the researchers tested 40 related compounds in the Novartis internal compound collection, again using NMR to detect binding. This led to the more potent compound 5, and two rounds of focused library assembly and screening led to low micromolar inhibitors such as compound 7. Structure-based design led to molecules such as compound 11, which has comparable potency to approved drugs that target FPPS. Compound 11 was further characterized using ITC and crystallography, and although its two carboxylic acids likely account for a relatively low cellular permeability, it does not show any affinity for bone.

Interestingly, a high-throughput screen conducted against FPPS did not yield any inhibitors with an IC50 better than 5 micromolar. So in this case not only did a fragment-based approach discover a new series of molecules against a new site on an old target, it succeeded where conventional HTS didn’t.

Fragments vs Abl: antagonists and agonists

A common concern with using biophysical techniques to identify fragments is that the functional implications of identified binders are not always clear, an issue we’ve discussed previously. In a new paper in J. Am. Chem. Soc., Wolfgang Jahnke (co-editor of the first book on FBDD) and colleagues at Novartis describe a clever NMR approach to address this problem and identify both agonists and antagonists that bind to an allosteric site on the protein tyrosine kinase Abl.

Abl is less well known than its famous cousin, Bcr-Abl, an oncogenic fusion protein in which the kinase activity is always turned on. Bcr-Abl is targeted by imatinib and a number of other kinase inhibitors; indeed, the success of imatinib against certain types of cancer has been largely responsible for the rush to develop drugs targeting kinases.

Most kinase-targeted drugs (including imatinib) bind in or near the conserved ATP-binding site. However, Abl offers another binding site, a pocket that can be filled by the fatty-acid myristic acid. This interaction causes conformational changes in the protein, stabilizing an inactive state. Indeed, previous research had identified molecules that bind in this pocket and block activity. Jahnke and colleagues used NMR screening of a 500-fragment library to try to identify new chemical scaffolds.

Several fragments were identified, some of which bound relatively tightly as judged by NMR and ITC. However, these fragments did not inhibit kinase activity. Crystallographic analysis of the fragments bound to Abl revealed that, although the fragments do bind in the myristate pocket, their binding modes are incompatible with the conformational changes needed to inhibit the kinase. Realizing that a specific valine residue is structurally disordered in the absence of myristate, the researchers established an NMR assay using Abl in which valine had been isotopically labeled to assess which molecules bind in a similar fashion to myristate (and thus block activity).

But what of the molecules that bind in the myristate pocket without causing conformational changes? Some of these can actually activate the kinase by competing with endogenous myristoyl groups. Fragment-based discovery of agonists is not unprecedented (see for example here and here), but it is rare. Assays such as the one described here to distinguish between different conformations of a protein could be practical complements to approaches that focus on binding alone. The paper is also a useful reminder that binders are not necessarily inhibitors, and can in fact be just the opposite.

ACS Spring Meeting 2010

The spring national meeting of the American Chemical Society has just concluded in (uncharacteristically sunny) San Francisco. The main fragment event was a full day session organized by Rachelle Bienstock of the NIH. The theme was “Fragment based drug discovery: success stories due to novel computational methods applications.” Rachelle is planning on getting some of the speakers to write chapters for a book, so I won’t do more than give a very brief overview here.

The session was very multinational, with speakers from France, Germany, Russia, and the UK, in addition to the US, and a good mix of companies and academics. On the computational corporate side John MacCuish from Mesa Analytics described the molecular shape fingerprints approach, Carsten Detering of BioSolveIT provided several examples of applying his company’s methods for fragment linking and scaffold hopping, and Francois Delfaud of MEDIT described mining the pdb for protein-fragment interactions and applying this to Eg5 inhibitors. On the computational academic side, Tobias Lippert of the Center for Bioinformatics in Hamburg discussed the Qsearch program, Vladimer Poroikov of the Institute of Biomedical Chemistry in Moscow discussed PASS, which relies on a large training set to predict actives and inactives, and Dima Kozakov of Boston University presented the FTMap approach for predicting fragment-binding pockets in protein-protein interactions.

Moving away from the purely computational, Yongjin Xu of Novartis described the application of virtual fragment screening to identify p38 and BRaf inhibitors, Vicki Nienaber of Zenobia described iterative fragment screening to identify potent and selective LRRK2 inhibitors, and I presented Carmot’s Chemotype Evolution approach. Finally, GPCRs appear to be increasingly amenable to FBLD; Richard Law of Evotec presented a number of applications of computational methods to various programs including histamine receptors, while Miles Congreve of Heptares presented their StaR Technology for generating stabilized GPCRs suitable for SPR, NMR, and crystallography and discussed applications to the adenosine A2A receptor and the beta-1 adrenoreceptor. In the later case, the researchers were able to obtain 9 co-crystal structures and found that agonists and antagonists bound somewhat differently.

There were also a few other relevant posters and talks throughout the conference. For example, I learned that Locus Pharmaceuticals has transformed itself into Ansaris; Fouzia Machrouhi presented a poster on developing nanomolar inhibitors of the protein kinase AMPK.

Finally, Andrew Woodhead presented an update on Astex’s CDK2 program. One of the earliest posts on Practical Fragments described Astex’s fragment-based discovery of AT7519, which is in clinical trials for cancer. However, with an oral bioavailability of less than 1%, this compound is administered intravenously. Extensive medicinal chemistry ultimately revealed that a relatively minor change – capping the secondary amine with a methyl sulfonamide – led to a molecule with dramatically improved oral bioavailabilty. This molecule, AT9311, also retains good activity in mouse xenograft models. This is a useful reminder that fragment-based methods are not a replacement for solid (and inevitably subsequent) medicinal chemistry.