Substrate activity screening for phosphatase inhibitors

Regular readers will be aware that there are lots of ways to find fragments, but one approach we haven’t covered yet is substrate activity screening, or SAS. A new paper in J. Med. Chem. by Jon Ellman and coworkers at Yale uses this technique to find inhibitors of striatal-enriched protein tyrosine phosphatase (STEP), which is implicated in cognitive decline in a variety of diseases.

Many enzymes can accept a wide range of substrates, and these are often fragment-sized. The basic idea behind SAS is that, since substrates (by definition) bind to a target, finding new substrates gets you new binders, and for some target classes it is straightforward to transform substrates into inhibitors. Of course, you could screen for inhibitors from the start, but the nice thing about looking for substrates is that you are far less likely to encounter artifacts. This is because artifacts normally muck up assays; it’s harder to envision a spurious substrate.

Phosphatases clip phosphates from their substrates. Protein tyrosine phosphatases (PTPs), for example, dephosphorylate tyrosine residues in proteins; they essentially perform the opposite reaction of protein tyrosine kinases. Like kinases, though, finding selective inhibitors can be challenging. The researchers started by building a small library of 140 phosphorylated fragments (previously described here) and looking for those that were particularly good substrates. One of the best for STEP was substrate 8, which looks quite different from phosphotyrosine.

Replacing the substrate phosphate group with a bioisostere (difluoromethylphosphonic acid) that could not be hydrolyzed by the enzyme gave compound 12, which had an inhibition constant (K_i) similar to the Michaelis constant (K_m) of substrate 8. Subsequent optimization led to compound 12s, with a low micromolar K_i and at least 18-fold selectivity against four other PTPs.

Unfortunately, the highly acidic phosphate bioisosteres in these molecules limit membrane permeability: although compound 12s inhibits STEP activity in rat neuronal cell cultures, it is not permeable in a model of the blood-brain barrier. Perhaps some of the less polar phosphate bioisosteres discovered in a previous virtual screen could help.

SAS is an interesting method, and I’m curious as to why more people aren’t using it. Of course, it does require generating bespoke libraries of fragment substrates, but once you have these they are useful for many members of a target class. What do you think?

Biophysics with White Wine (pt 2)

Tarte flambe or flammekuchen.  Doesn't matter what you call it, DELICIOUS!  Another fantastic find from the Novalix Biophysics in Drug Discovery Conference.  Yesterday I wrote up the Biophysical Characterization section, today's yummy-ness: Mechanistic Analysis. 

Ann Boriack-Sjodin -Epizyme:  She emphasized that X-ray is the key to Epizyme's work, but they also use STD, ITC, SPR, Fortebio, thermal shift, and enzymology.  This was a theme, especially for the non-fragment specific talks: we use any and all biophysical techniques.  This talk focused on the methyl-transferase DOT1L. They struck out with a diversity library and in silico screening.  They did find, with SBDD, a selective inhibitor.  They key to this compound was its VERY long residence time: 24 hours.  The concept of koff driven inhibitors was brought up in several other talks.

Glyn Williams -Astex: First off, let me say the best thing Glyn said during his talk was that the Ro3 was meant as a guideline.  His talk spoke about the variety of methods in use at Astex: MS, NMR, Thermal shift, ITC, and X-ray.  MS is used for protein validation and QC, thermal shift was used for affinity ranking, but they have moved away from it (his comment, "when it doesn't work, you don't know"), ITC is a good way to discriminate good compounds from bad, and NMR is used in competition mode.  In terms of their library, they have had 2600 fragments EVER and their current fragment library iteration has 1500 members.  Their core library has a avg MW of 176 (~13 HA) and clogP of 0.9 and the X-ray subset of 350 fragments 146 Da (~10 HA) clogP of 0.5.  40% of their fragments are NOT commercially available.  These are small fragments and he noted that on average each fragment has hit two targets.  Greater than 50% of their hits have never generated a X-ray structure, but have hit in the biophysics assays.  He presented how they do 3D-arity.  The draw a "best plane" through the molecule and then calculate the average deviation of each atom from that plane.  I found this approach unwieldy and I still think PMI is a better way to go.
They take several approaches to fragment screening: with their core fragment library they WATER-LOGSY and thermal shift which then goes into X-ray follow up ( with MS, ITC, and 2D-NMR).  If the X-ray works, they have a X-ray validated hit and it moves forward.  They also sometimes go straight into a X-ray screen (with a 350 fragment subset).  One of the advantages of the NMR-based screening is that NMR can detect hits < Kd, while X-ray can only detect hits > Kd. 
In terms of properties, he showed a fascinating graph (that I bet lots of people have) that shows that improvements in enthalpy occur during H2L and improvements in entropy during LO.  LogP occurs in H2L and stays the same in LO. 

Marku Hamalainen -HealthCare:  This was an interesting talk, especially when contrasted with Goran Dahl's.  He showed a very interesting graph (I am not showing slides without specific permission; I have asked for this one) that shows binding site occupancy as a function of on/off rates.  It is fascinating as it buckets your compounds in various regimes: "Ancient Medchem knowledge", "Without on you are off", "High affinity does not help if clearance is rapid", and "With slow off, you might still be on when the drug is gone". 

Goran Dahl - AZ: This talk was definitely in the "Yeah, of course" category.  Not to diminish his talk, which was excellent, but it makes sense in a only after someone points it out to you kind of way.  Kudos to him for saying it first (chronologically at least): koff does NOT correlate with PK.  Prolongation kicks in when koff< elimination rate.  Pure and simple, yet how many people had actually thought about it that.  Plasma t1/2/ residence time > 1, duration is driven by PK, < 1 and it is driven by binding kinetics. 

Geoff Holdgate -AZ: This was an excellent talk giving a high-level overview and then diving into some very interesting topics.  He spoke on combining thermodynamics and kinetics to drive chemistry.  Key Questions: "How do you improve medchem decisions with kinetic data?" One Kd can arise from many different kinetic profiles this would allow you to pick and chose one that could be beneficial, but how do you know what that would be?
 "Is biophysics simply useful for retrospection?"  There are no examples of the use of biophysical data to drive medchem prospectively. 
"Should you drive affinity/LO by Delta H only?"  From Glyn's talk, it seems like LO is driven by entropy, NOT enthalpy. 

His take home lesson, which I wholeheartedly agree with: the Drug Discovery paradigm of focusing on affinity needs to change. 

Biophysics in the Alsace

Two weeks ago, the first Novalix conference on Biophysics in Drug Discovery was held in Strasbourg.  I was lucky enough to be one of 160 people in attendance (this was largely a european affair, with ~10% of attendees from outside the EU).  The split of attendees was 60/40 industry/academia.  The conference was split into four themed sessions: Biophysical characterization, Mechanistic Analysis, Emerging Technologies, and Biophysical Methods for Identifying Hits and Leads.  This was not a fragment conference, but many of the talks were specifically about fragments, and the rest could be impactful in fragments.  I want to share my impressions/thoughts on the speakers relevant to the readers here.  You can also go to my website to see my thoughts on the speakers not relevant to FBHG. 

Michael Hennig- Roche: His talk discussed the various methods and showed examples for each.  This was a great talk giving a great overview of the various methods available for active follow up.  Specifically fragments: The Roche fragment library is ~5000 compounds.  In terms of QC, 80% of the samples show >85% purity (by LC-UV-MS).  Purity of fragment libraries has been discussed here previously.  For Roche's uses, every fragment hit is followed up by MC and NMR, so a lower threshold of purity will not have a negative impact.  He also presented results from a 2D-HTS.  This was a new concept for me and I found it intriguing.  The basic concept is to graph the results from two screens (or related proteins) to identify compounds that activate one, but not the other, or activate one and stimulate the other, etc.  He also presented direct and in-direct methods using Mass Spec methods.  To me, this area was one of the more fascinating areas discussed at the conference.  Theoretically, this could be applicable to fragments, but I would really like to see specific applications.  Lastly, he spoke on biophysical methods and membrane proteins. 

Rob Cooke- Heptares:  He presented the STaR approach that has been widely published and presented here, here, here, and here.  The talk was very similar to other talks that Heptares and Rob have presented in various fora over the past year.  The main thing that I was taken by was that there was no mention of NMR.

Matthias Frech - Merck: I really enjoyed this talk.  One of the main things I noted was his use of the phrase "hit affirmation".  Confirmation (according to the dictionary) is a piece of corroboration, while affirmation means it is true.  Is this parsing meaning where none exists?  Maybe, but I think it may also inform on mindset.  He said that SPR is the workhorse for FBHG, but they also use NMR, MST, ITC, stop-flow, and X-Ray.  95-98% of their projects are accomplished using SPR and ITC.  However, he stated that SPR is used to rule out compounds, not rule them in.  This is key to the proper use of SPR.  I would be interested to see if anyone else takes this approach.  They use SPR and ITC to obtain the enthalpic and entropic terms for compound binding.  ITC yields the enthalpy, SPR yields the DeltaG et voila, simple math (my favorite kind) yields the entropic term.  One other very interesting item that he noted was that there was no correlation between affirmation rate and target class for 31 projects they undertook (2009-2012).  

 

Tomorrow I will update the Mechanistic Analysis session.  

Fragment merging revisited: CYP121

Last year we highlighted a paper from Chris Abell and colleagues at the University of Cambridge in which they applied FBLD to CYP121, a potential anti-tuberculosis target. Several fragments with different binding modes were identified, and while some could be successfully merged to produce higher affinity binders, others couldn’t. In a new paper in ChemMedChem, the researchers take a closer look at why some of their initial attempts at fragment merging failed, and figure out how to succeed.

In the original paper, fragment 1 was particularly interesting for two reasons. First, crystallography revealed that it did not make direct interactions with the enzyme’s heme iron, as do most inhibitors of CYPs, suggesting that higher specificity might be achievable. Moreover, the co-crystal structure revealed that fragment 1 could bind in two nearly overlapping orientations, practically begging to be merged. Unfortunately, the resulting merged compound 4 actually bound worse than the initial fragment.

Computational modeling suggested that a primary reason for this disappointing result is the steric clash between two hydrogen atoms on the two phenyl rings of compound 4. These are forced into an unfavorable configuration when the molecule binds the protein. To fix this, the researchers sought to introduce a new interaction with the protein that would allow the molecule to relax into a lower energy conformation, alleviating the steric clash. This led to compound 5, with a satisfying increase in affinity. But lest folks become too cocky, an attempt to pick up an additional hydrogen bond to the protein (compound 8) actually led to a decrease in affinity despite the presence of the designed hydrogen bond, as assessed by crystallography. More successfully, building into a cavity led to the most potent compound 9. (Geeky aside: the aminopyrazole versions of fragment 1 had similar affinities as fragment 1, suggesting that the aminopyrazole moiety per se only gives a boost in potency in the context of the merged molecule.)

High-resolution co-crystal structures were solved for several of the molecules; the figure below uses color-coded carbons to show the overlay of the two different binding modes of fragment 1 (green), compound 4 (cyan) and compound 5 (magenta). What’s striking is how closely all the molecules superimpose, despite their very different affinities.

This is a nice case study in fragment merging that emphasizes just how difficult the strategy can be, even when it looks like it should be easy. And while Practical Fragments has not always looked kindly on computational methods, this is a beautiful example of how modeling can be used to understand why things that look good on paper don’t work, as well as how to fix them.

Tessera, Tessera Everywhere

A lot of papers come across the editorial desk here at Practical Fragments.  Most of them appear because of a keyword in a search, sometimes somebody says, "Hey did you see this?", and sometimes we miss them (so if you see one that you think is interesting, doesn't hurt to ping us).  I recently came across this paper.  Well, the first thing that struck me was my branding was working, my eminence (LOL) in the field is working its way into people's thought process; I mean seriously the first line of the abstract is the entire reason my company is called what it is.  So, with great interest I dove into the paper.  So, what is it about?  The authors describe a fragment library and its use in a chemogenomics approach against three diverse target classes: GPCRs, Ligand-gated ion channels, and a kinase.  

The authors propose that promiscuous hits are driven by desolvation.  Thus, they propose a new sub-field: Fragment-based Chemogenomics (which obviously is a subset of Fragonomics) which is:

"an approach to accurately characterize protein–ligand binding sites by interrogating protein families with libraries of small fragment-like molecules."

They constructed a library of 1010 fragments "inspired" by the Voldemort Rule: number of heavy atoms 22, log P < 3, number of H-bond donors 3, number of H-bond acceptors 3, number of rotatable bonds 5, number of rings 1.  They then applied some medchem filters followed by a scaffold diversity analysis. 81 novel scaffolds were purchased to supplement underrepresented scaffolds.  Very nicely, they also identified scaffolds that were over-represented and selected those with high "cyclicity".  Lastly, the removed any compounds that showed any aggregation or preciptitation in any of the (published or unpublished) biochemical/biophysical assays.  Honestly, I wish they were more explicit here. The chemical space seems to be well represented (not shown) with underrepresentation of small aliphatic ring systems.

Physicochemical Properties of Fragment Library

They then screened the library against their various targets and, as expected, were able to identify actives with different hit rates.

This figure shows selected (how selected and are they necessarily important ones?) properties of the actives for selective and non-selective hits.  [N.B.  I am pretty sure that the second graph in (b) should be MW, not LogP again.]  Why these properties and not all of them?  To me this smacks of hiding data that do not support the central thesis.  More than a half of their actives bind to one specific target; however, 44 bind to two targets, 12 to three targets, and 11 to four targets.  Interestingly, none of the actives bind to all targets, so while they tried for some promiscuity, they did not get anything truly promiscuous.  There is no correlation between hydrophobicity and non-selectivity which they conclude means for fragments, unlike lead-like molecules, non-selectivity is NOT driven by desolvation.  

They then discuss two different types of cliffs: affinity (where two similar molecules differ in their ability to have activity) and selectivity (where two similar molecules are active against different targets).  I must admit my naivete here, but these two cliffs appear to be what I know as SAR. 

I ended up wanting more out of this paper, but for a first attempt it lays the groundwork for future refinement.  Is it a new idea? No.  The whole reason I came up with the Fragonomics name in the first place was as a joke/rebuttal to the huge array of -omics that were underway at Lilly at the time: chemogenomics, genomochemics, and so on. I would hope that this paper is a prelude to a much larger analysis of all properties and their correlation to specficity and non-specificity, down to the level of side chain and scaffolds. 

**EDIT** Dan just pointed out that he already blogged this paper back in MAY!  That's a lesson for me to blog while jetlagged.  

Biophysics bonanza of fragments for pantothenate synthetase

As Teddy just mentioned, biophysics provides multiple methods to find fragments, and it's best to use several. They may not always agree, but fragments that hit in several assays are more likely to be real. Although this requires using different skills, using diverse methods does not take a village, as illustrated by a recent paper in Proc. Nat. Acad. Sci. USA by Hernani Silvestre, Tom Blundell, Chris Abell and Alessio Ciulli at the University of Cambridge.

The researchers were interested in the enzyme pantothenate synthetase (Pts) from the bacterium that causes tuberculosis (a target we’ve covered previously here and here). They used a thermal shift assay to screen 1250 fragments from Maybridge, each at a whopping 10 mM concentration, resulting in 39 hits (3.1%) that stabilized the enzyme by at least 0.5 ˚C. Another 17% of the fragments slightly stabilized Pts, while 73% of the fragments destabilized the protein (a poorly understood phenomenon that does not seem to reflect specific binding).

Despite being relatively small, the fragment library had some scaffolds that were over-represented, and some of these were enriched among the hits, providing early SAR. Perhaps not surprisingly given the anionic character of the enzyme’s ATP cofactor and pantoate substrate, about half the hits were carboxylic acids.

All 39 hits were analyzed by WaterLOGSY and STD NMR, and only 17 showed evidence of binding, although the NMR experiments were done at a much lower concentration of fragment (0.5 mM). Competition experiments revealed that all except one of the 17 fragments bound at either the substrate or cofactor sites.

Next, isothermal titration calorimetry (ITC) was used to measure the dissociation constants of the 17 validated fragments. Measurements could not be obtained for three of the fragments; values for the rest ranged from 0.5 to 17 mM. There seemed to be a rough correlation between affinity and the extent of stabilization in the thermal shift assay, and binding was enthalpically-driven.

The 17 fragments were then soaked into crystals of Pts, resulting in 8 co-crystal structures. Most of the fragments that did not produce structures also had low affinities as assessed by ITC. Four fragments bound in a pocket normally occupied by the adenine ring of the cofactor ATP, while the other four bound in the substrate pantoate pocket. One of these bound quite deeply in this pocket and caused a conformational change in the protein. The researchers were able to obtain a structure of Pts bound to both this fragment and ATP.

There’s lots of nice data in this paper, and all the new structures have been deposited in the protein data bank. More generally, the “integrated biophysical approach” provides a practical template for applying FBLD. The paper has just four authors, and only two of them actually performed experiments according to the author contributions section. That short list provides more evidence that a very small but dedicated team can successfully find and validate fragments.

Biophysics in Drug Discovery

Folks,

Posting has been light since Dan has been traveling and I have been super busy.  The fun news is that I will be attending the Novalix conference on Biophysics in Drug Discovery this week.  The agenda is shaping up to be very exciting and I think of keen interest to the readership of this blog.  I will (WiFi-willing) be blogging daily about the conference and tweeting random thoughts about the talks. 

For those of you who will be there, I am always happy to meet up and chat.  There are a number of very interesting Fragment-focused talks, so I will be focusing on those. 

To have shape or not?

I consider the debate/discussion/civil discourse on what/where/when/why/how of 3D fragments to be one of the most interesting topics in the fragment field right now.  A paper has come out from the 3Dfrag consortium.  The first take away from this paper is we have a NEW acronym, FBHI (Fragment-based hit Identification).  Their central thesis is that fragments are too flat for certain target classes of targets, thus resulting in too low hit rates. This paper describes their efforts in pre-competitive space to prove/disprove this hypothesis.  It's fascinating.  

Using the nPMI as their metric for "3D-arity" they compare 1000 fragments representation "commercial" space to Zinc InMan Compounds fragmented by RECAP.  It is obvious that InMan has more 3D-arity, but the question is always, is this biased by target type of compound type (like natural products)?  But, to the greater point, will 3D fragments give better hit rates than 2D for certain target classes?  Many people point to the Hann complexity model a (MIP) and say no, because there is more potential "bad" interactions.  However, the authors point out if the NUMBER of sites of interaction is the same, this should not be the case.  Recent data from Evotec seems to support this they state. 

The 3Dfrag consortium's goal is to explore the role of "3D-arity" in fragment screening success.  As shown in this figure, they apply pretty standard selection rules for their collection, including the Pfizer Rule (medchemists inform on the final selection process).  

Interestingly, even after all of this, trying to select for "3D-arity", 70% of their final 200 molecules were largely flat (nPMI1+nPMI2< 1.1).  Making matters worse, 10% of the compounds could not be delivered by suppliers and 5% failed QC (by NMR). The nPMI and maximum similarity is shown below for 170 compounds.  For the maximum similarity, a right-shifted plot shows a high degree of internal similarity, while left-shifted would be more diverse.  So, it appears that their library is diverse (most compounds < 0.8).  Importantly, of the 170 compounds, they appear to fill a much wider region of 3D-space.   

One of the tools generated by the consortium is 3Dfit, a webtool that is free and allows fragments to be evaluated.  They generate up to 9 conformers (blue to red, low to highest energy) that are then plotted on the PMI triangle.  They also calculate a radar plot of molecular properties:

Molecules outside the consortium's GUIDELINES are flagged for consideration (see (b)).  The consortium has now undertaken chemistry to generate additional 3D fragments.

As the authors state:

We believe that this strategy will build a library with broader coverage of biologically relevant chemical space compared with current fragment collections and the identified hits will offer higher-quality start points for medicinal chemistry projects.

They will now experiment and try to validate (or invalidate) this hypothesis.  But, as they point, these sorts of Pre-Competitive Initiatives have difficulty getting buy in from partners willing to devote time, resources, assets to this.  They are looking for more people to participate.  I have no skin in this game, but I would love to see people step up and help this initiative prove/disprove a crucial hypothesis in Fragments.

Fluorinated fragments vs FAAH – functionally

Fluorine NMR is a topic that has come up several times on Practical Fragments (see here, here, and here). As readers will recall, the ¹⁹F nucleus can be readily detected with a properly equipped NMR spectrometer. The isotope has a wide range of chemical shifts, and sensitivity to the local environment makes it easy to detect whether fluorine-containing fragments bind to a protein. But you don’t need a dedicated fluorine-containing library: in a new paper in ChemBioChem, Claudio Dalvit and coworkers at Fondazione Istituto Italioano di Tecnologia describe using fluorinated substrates to screen a membrane enzyme.

The researchers use an approach they call n-fluorine atoms for biochemical screening (n-FABS). A substrate or cofactor is labeled with fluorine, and when this is processed by an enzyme, the resulting change in chemical structure affects the ¹⁹F chemical shift, which is easily detected by NMR. Either substrate or product (or both) can be observed, and a decrease in product can be attributed to inhibition of the enzyme.

The researchers were interested in the protein fatty acid amide hydrolase (FAAH), a membrane-bound enzyme that hydrolyzes lipids such as endocannabinoids. Of course, membrane-proteins are tough to screen using fragment-based approaches, and the fact that this enzyme processes lipophilic substrates makes things even more challenging. The researchers synthesized several fluorine-containing substrates, but most of these turned out to be insoluble or formed aggregates, even at low micromolar concentration and even in the presence of detergent. Ultimately they were able to make one substrate that was soluble at 30 micromolar, sufficient for screening.

Next, the researchers assembled a library of fragments. Although these did not need to contain fluorine for the n-FABS assay, the researchers chose to focus on fluorine-containing fragments anyway, perhaps so they could use other NMR methods to confirm binding. Of 160 commercial fluorine-containing fragments purchased, 113 showed solubility ≥ 0.1 mM, purity ≥ 75%, and no aggregation. These were combined into 23 pools of 5 and screened for inhibition in the n-FABS assay at 200 micromolar of each fragment. Pools that showed >15% inhibition were deconvoluted to find the active fragments; some contained more than one hit. This process led to a remarkably high hit rate of 16.5%. The IC₅₀ values of all 19 of these hits were then determined using n-FABS and they showed quite a range, from quite potent (3 micromolar) to low millimolar.

These are nice results and there are clear opportunities for advancing some of the fragments, but I must admit I was left wanting more. The n-FABS assay is essentially an inhibition assay, and of course there are all kinds of things that can show inhibition without proper binding. However, since all the fragments do contain fluorine, it would be straightforward to actually measure direct binding using NMR; it would be very interesting to see how many fragments show up in both assays. Perhaps we will see this in a follow-up study.