(Not)Drugging the Undruggable

Well after spending last week at a Structure-based Drug Discovery conference that ended up being horribly academic and computational (think force fields and not the cool Star Trek kind), I come back to my blogging pile to find this paper. 

In this paper, the authors are looking at HIV Integrase.  HIV integrase has long been considered an "undruggable" target, although I would think that the presence of a marketed drug would kind of kill that perception.  But as we all know, preconceived notions die hard and slowly.  Anyhow, viral evolution necessitates the discovery of next generation integrase inhibitors.  To that end, the authors decided to use GLIDE to dock quinolines to the HIV Integrase-IN−LEDGF/p75 interface because some previous work had shown these molecules to have "been previously explored" for anti-integrase activity.  They quickly found out that 8-hydroxyquinoline made far more favorable contacts (one more hydrogen bond) than quinoline.  

They immediately started testing 8-OH-quinoline fragments for potency (via an Alpha-screen) and found molecules with very low micromolar potency but they do not report LEANs.  [It absolutely should be a requirement for any paper citing itself as "fragment" to include some sort of ligand efficiency metric with its data.] QA, QB, and QC have LEANS of 0.45, 0.41, and 0.37.  Yet, these compounds were cytotoxic.  

To expedite their search for non-cytotoxic, they generated a pharmacophore model.  This was used to screen ~7000 compounds generated from a query of 8-OH-quinoline and found that 5- (typically phenyl) and 7-substituted (typically phenyl) 8-OH-quinolines were the best output.

 In a perfect case of fitting the results to your preconceived notions the authors note: 

Although the pharmacophore identified the neutral form of compounds from the database, we have used the ionized form for the purpose of pharmacophore mapping, as we consider these compounds to be ionized in the context of receptor binding.

Despite adding a tremendous amount of heavy atoms, none of the compounds had activity much better than the original 12 HAC quinoline and were still cytotoxic.  They eventually found that only the 5-substituted-8-OH-quinolines did not have cytotoxicity. However, and please note, the potency is still NOT better than the original fragment hits.  Changing the C5-phenyl for piperazine or piperidine did increase potency (to 0.4uM, 0.35 LEAN) and reduced cytotoxicity.  What about the 8-OH moiety you ask?  Well, they found out that it could be substituted and/or be a thioether.  Of course, these weren't any more potent than the lead fragment and cytotoxicity remained an issue.

Well, to continue in a vein started by Dan, these authors seem to have wasted valuable NIH/NIAID and Campbell Foundation money.  Did they discover inhibitors of HIV Integrase, sure.  Are these useful frameworks?  Proabably not.  Did they establish a robust SAR from which they can move forward?  No.  Their activities floated right around 2 uM, with one or two getting below the uM line.  As I said, your computation is only as good as your experimental follow up.  In this case, it doesn't pass muster.

 

GIGO: pollution of the literature continues

Practical Fragments has repeatedly warned about the dangers of what Jonathan Baell dubbed pan-assay interference compounds, or PAINS (see for example here, here, and here). These are compounds that hit numerous unrelated targets through mechanisms that can charitably be described as “non-druglike”. Regrettably, many people still do not recognize these nuisance compounds for the artifacts they are, and PAINS continue to show up in high-profile fragment libraries. An unintentional illustration of why this is a problem was recently published in J. Med. Chem.

The researchers were interested in STAT3, a popular oncology target. They used a computational approach to extract fragments from reported inhibitors and then recombined them into new molecules, a few of which were made and tested. Unfortunately, some of the previously reported inhibitors were PAINS, and, like HeLa cells contaminating cell cultures, the resulting pathological fragments contaminated this research. The most active molecule out of this exercise, compound 8, is a para-quinone:

Quinones are troublemakers for two main reasons. First, they can nonspecifically react with thiols (see figure), and STAT3 does indeed have several free cysteine residues. Second, quinones are well-known redox-cyclers: they can be reduced and then re-oxidize in air, generating reactive hydrogen peroxide in the process.

The researchers showed that compound 8 is active in several cell assays and a mouse xenograft tumor model, but of course any generic alkylator could also show these effects (mustard gas, anyone?) and hydrogen peroxide is itself an important second messenger. It is impossible to say whether the activity of compound 8 is due to interaction with STAT3 on the basis of the experiments reported in the paper. The only evidence that compound 8 interacts with STAT3 at all comes from a fluorescence-based assay which appears to show 70% inhibition at >100 micromolar compound 8, a concentration far higher than the cell experiments.

In other words, what this paper shows is that a quinone has modest but ill-characterized biological activity. Of course, just because a compound can be a bad actor doesn’t necessarily mean it is behaving as one, but in the case of PAINS it is best to assume guilty until proven innocent. Indeed, a figure in the Supporting Information shows that the compound also inhibits STAT5 phosphorylation, supporting the notion that it acts through multiple mechanisms.

We can do better than this.

I hesitated before writing this post – I don't want to come across as a mean-spirited vigilante – but one of the strengths of science is its self-correcting nature. Researchers should learn to recognize PAINS when they inevitably show up as screening hits. If and when they don’t, editors and reviewers evaluating manuscripts and grants have an obligation to hold them to account.

It is easy to ignore or shrug off sloppy science, perhaps with a cynical chuckle, but papers like this fill me with a mixture of sadness and outrage. This research consumed the time and efforts of four scientists, not to mention scarce funding from the NIH and Alex’s Lemonade Stand Foundation, a charity founded by a young girl who subsequently died of her cancer at the age of eight.

We owe it to society to stop wasting resources chasing artifacts.

Fragments vs DAAO

Fragment-based approaches are often applied to tough targets, such as protein-protein interactions or BACE1, that have stymied more conventional approaches. Although kinases have certainly been the focus of many fragment efforts, other more “traditional” enzymes are sometimes ignored. In a recent paper in J. Med. Chem., Takeshi Hondo, Tatsuya Niimi, and colleagues at Astellas Pharma show that fragments can play a valuable role here too.

The researchers were interested in D-amino acid oxidase (DAAO), a potential schizophrenia target. This enzyme catalyzes the deamination of amino acids such as D-serine, so it is not surprising that very small, fragment-sized molecules can bind to it quite tightly (as indeed we noted here). Recognizing this, the researchers conducted a high-concentration screen of 3500 fragments. One of the more interesting hits was compound 8, which is actually a fragment of a previously reported molecule. With low micromolar potency and just 8 atoms, the fragment has a ligand efficiency of just over 1 kcal mol^-1 atom^-1, one of the highest values I’ve ever seen.

 In addition to its impressive affinity and ligand efficiency, compound 8 induced a conformational change in the protein to open a nearby subpocket, and growing into this pocket led to dramatic improvements in potency. Additional optimization for permeability and brain penetration eventually led to compound 30, with low nanomolar activity in both biochemical and cell-based assays. This compound proved to be selective against a panel of 57 potential off-targets, and was found to be active when dosed orally in a mouse model of schizophrenia.

This is a lovely example of structure-based fragment growing. Although it’s rare to find such a small, potent fragment, examples such as this do support the inclusion of very small fragments in screening libraries.

What you are Missing.

The nice part about blogging is you can post anything you want.  Mostly, we post on interesting (to us at least) papers, news about fragments, conferences, and so on.  We have assiduously avoided being a commercial outfit; what we post is what we believe in and we don't post for cash or barter.  So, we tend to not promote people/companies.  I am so totally breaking that rule right now.  

Many of you probably know Chris Swain, the principal of Cambridge MedChem Consulting.  [Full Disclosure: Chris and I have worked and published together.  I hope that does not diminish your feelings about him.  :-)] If not, this post will introduce you to him and the many wonderful resources he curates, particularly in the FBHG world.  There is so much there, sometimes I forget what he has; in fact, Chris's site goes by the rule, "Why make them buy the milk, give it to them, and the cow too!"  

One of the great resources Chris has is a graphical snapshots of a metric pile of fragment collections.  On Monday, Chris added nPMI (Principal Moment of Inertia) to these snapshots.  As has been discussed some, I (and Justin Bower from the Beatson) think this is the best way to evaluate "3D-arity".  It is interesting to just browse through the snapshots.  Some of the collections that are VERY large do seem to have a good to excellent amount of 3D-arity.  Does this correlate with increased hit rates for certain target classes?  

Well, Chris has thought of that.  He has been collecting the compounds from the literature that are reported as fragment hits.  He has just updated that snapshot with nPMI also.  What do the reported fragment hits tell us?  I would say that the vast majority of the reported fragments are Voldemort Rule compliant.  No surprising.  What I would like to see is a breakdown of fragment property against target type.  This may the part of the cow Chris isn't giving away.  There may be other slices/dices too.  

Trust me, go and spend some time on Chris site.   I won't call it a time waster, but it will suck you in.

Multiple methods find fragments on MEK1, but fluorimetry shines

The tendency of fragments found in one assay to reproduce – or not – in another assay is a frequent topic at Practical Fragments. In a recent paper in Bioorg. Med. Chem. Lett., researchers at Sanofi describe their experience screening the oncology-associated kinase MEK1.

The researchers were interested in the ATP-binding site of MEK1, and they started with a virtual screen (using Glide-SP) of a 10,000 compound library. The top 196 hits were then tested experimentally by differential scanning fluorimetry (DSF) and surface plasmon resonance (SPR), leading to 30 and 44 hits, respectively, with 12 in common. A subsequent biochemical assay of the same 10,000 compound library yielded 106 hits, only 13 of which were in common with the virtual screen. 158 different fragments were identified by one or more of the three experimental methods.

Of 13 hits selected for follow-up experiments, crystallography ultimately yielded structures for 7 of them, of which 5 had been identified in the virtual screen. Interestingly, SPR had only confirmed 2 of these molecules, while DSF had confirmed all of them. Thus, in contrast to some reports, the Sanofi folks are quite sanguine about DSF and advocate using it widely and early in a project (as indeed many people do seem to be doing). The technique is fast and easy, and in this case the researchers were able to run the DSF screen before they had finished developing their biochemical assay.

The paper includes detailed comparisons between virtual screening, DSF, SPR, biochemical, and X-ray approaches, and is well worth examining if you are putting together a screening cascade.

The researchers conclude:

There is no gold-standard method for screening fragments. The general approach is to conduct a primary screen and then follow this up with at least another method to confirm hits, which are subsequently prioritised for structure determination. Different groups adopt different methods based on availability of materials, in-house expertise and prior experiences screening fragments.

In other words, multiple methods can find fragments. Ultimately, you’ll probably find real hits whatever methods you use, as long as you’re careful.

Inhibition in Solution Assay (ISA) – on a surface

As illustrated by our poll, surface plasmon resonance (SPR) is one of the most widely used techniques for finding fragments. However, as commonly practiced, SPR – like all techniques – has drawbacks. For one thing, despite impressive recent advances, it is still not particularly high throughput. Also, the protein is typically immobilized on a sensor chip, and the detection of binders depends on the mass ratio of the ligand to the protein. With larger proteins and smaller fragments, this can quickly push the signal below the noise.

A seemingly simple solution is to reverse the experiment: immobilize the small molecule and add the (comparatively large) protein to get a whopping signal. Indeed, this is the approach that Graffinity (now part of NovAliX) takes, and is conceptually similar to work done at RIKEN. However, both these techniques require dedicated surfaces functionalized with fragments.

In a recent paper in J. Med. Chem., Stefan Geschwindner, Jeffrey Albert, and colleagues at AstraZeneca sought to simplify matters. Their idea is to immobilize a single high-affinity molecule to a chip. Protein in solution should give a good signal when the protein binds to the surface, and adding competitor to the solution should decrease protein binding to the immobilized target compound, thereby reducing the signal. They call this the “inhibition in solution assay”, or ISA.

The researchers used the protein PDE10A as a test case and attached a previously characterized small molecule to the surface; this modified small molecule has an IC₅₀ of 991 nM for the target. They then used two different approaches for detecting interactions, SPR (GE/Biacore) and a 384-well plate-based optical waveguide grating (OWG) from SRU Biosystems. Both formats work and give comparable results for a set of molecules ranging in affinities from 40 nM to 0.5 mM.

One nice feature of this approach is that, as a competition assay, it should only identify molecules that are competitive with a known binder. On the flip side, ISA does require a reasonably potent binder for your protein, and you must be able to modify this molecule such that it can be immobilized to the surface. And of course, there are still problems at high concentrations; the researchers mention that high loading of immobilized small molecule can cause other molecules to stick to the surface. Still, this is an interesting approach that should be easily applied to many systems. I’d be curious to know whether you’ve tried it or a variant, and how it compares to more conventional SPR methods.

Poll results: how small are your fragments?

The results of our most recent poll are in - here are the smallest fragments readers would allow in their library:

 

It looks like the smallest fragment most people would include in their library has a median of 7 or 8 non-hydrogen atoms, just slightly smaller than azaindole. More than 85% of respondents set a minimum size of 5 to 10 heavy atoms, so if we take the Pfizer rule of thumb that each heavy atom averages 13.3 Da, we’re talking 67 to 133 Da for the smallest fragments.

These sound like reasonable limits; slightly smaller molecules start becoming too volatile to handle reliably. Also, as Teddy pointed out, you’ll probably need either very sensitive methods to detect the smallest fragments, or very impressive ligand efficiencies.

Our poll last year asked about the largest fragments, so together these polls set a range of 5 to 20 heavy atoms for typical fragment libraries.

Thanks to the 75 of you who voted in this most recent poll.