19F Target-based Screening Reduced to Practice

If you read this blog you know I love 19F NMR.  I am a big fan of it for ligand based screening or as a secondary screen in target-based mode.  Well, this paper is the first to use target-based NMR to screen small molecules.  Using 3-fluorotyrosine-labeled protein and using what they call Protein observed fluorine NMR (PrOF-NMR), the interrogate a PPI (CBP/p300-KIX) and determine this interaction's ligandability. 

Starting from the Maybridge Ro3 Library, they found 508 19F containing fragments.  These were put into 85 mixtures (5 or 6 compounds each) and screened at 833uM and 2.5 % DMSO (Figure 1, KG-501 is a control compound).  Each experiment was performed in 5 minutes (with a quick reference spectrum) which is faster than SOFAST HSQC (for 15N labeled proteins).  15 mixtures gave hits which upon deconvolution gave 4 actives (>2 SD in chemical shift), a 0.8% active rate.

Figure 1.  Typical 19F Data

 They then titrated the four confirmed fragments to determine the Kd, which for 3 of them was in the mM range.  They followed this up with Analog by Catalog and developed some SAR.  Lastly, they used H-N HSQC to verify that the compounds do bind and where they data shows.  They do. 

Some thoughts on this.  The use of 3FY is only one amino acid that can be used.  Fluorotryptophan can also be used, so this method can easily be applied to other systems.  Secondly, 19F can accomodate much larger pool sizes (within reason).  And there is no reason why both could not be used.  Of course, one of the things not noted here is that they produce single mutants in order to ID each individual residue.  I think you could live without residue specific assignments and still get tremendous value out of this method.  I would be curious what others think.  We bring up impractical tools all the time, so I really want to applaud this paper.  Here is a very practical new method for screening. 

Still Impractical, but Getting Better...

Dan and I don't see every fragment paper and so it's nice when people point out papers to us.  Its typically their recently published paper and they are looking for some sort of recognition/validation from us.  Sometimes its a paper we have on our radar, sometimes it isn't.  Recently, I received an email pointing out this paper from someone I met at FBLD2014.  Well, this is a follow up paper to a paper I discussed a few years back.  The title of the post sums up my thoughts: "Another (Impractical) NMR Method."  One comment from that post (by a co-author from Astex on the current paper) was

This is a relaxation filtered ligand based method (like T1rho or selective T1). You therefore may consider using it in competition mode if you find one suitable "spy" molecule: this would allow with a single point experiment screening mixtures and/or ranking for low affinity hits(especially if solubility is limiting). I would actually give it a try.

So, it looks like she did.  The advantages of Long Lived States (LLS) NMR is that the the dynamic range is wider and it works at low protein concentrations (3 uM or roughly the same as STD or WaterLOGSY) or with very weak affinities.  Against the workhorse HSP90 system, they screened mixtures using LLS and a spy molecule.

  

I still don't think this is a very practical method because it still requires a lot of tailoring to individual systems; in particular the compounds need pairs of protons which are suitable for excitation to the LLS.  Using it in "spy" mode gets around this. I would still hold that this is an impractical NMR method.  I like to see people developing new methods and trying to improve them.  We also need people doing good comparisons of "standard" experiments to new ones.  (Here is how NOT to do a method comparison.) I won't dismiss this out of hand, but there is still a lot of work to be done here to move it into a "front line" screening technique.

Another NMR Tool...

There are many things which aid in the successful prosecution of fragments.  Most people would agree that structural information is one of those things.  However, in many cases there is no structure, nor any hope of obtaining one.  Many different methods have been developed to try to address this gap.  Oftentimes they are impractical, sometimes they are useful.  In this paper, Gregg Siegal, Marcellus Ubbink, and co-workers from his academic lab present a new NMR-based structural tool.  [Editor's Note: I used to have a business relationship with Gregg's commercial side.]  So, is this a practical or impractical tool?  You can skip down to the bottom for the answer, or keep reading and follow me down the rabbit hole.

Their approach is not to generate high-resolution structures, but low resolution models of how initial fragments bind to the target. To accomplish this, the use pseudocontact shifts  (PCS)induced by paramagnetic ions.  To those of you whose eyes just glazed over, let me explain.  We typically only use diamagnetic atoms in NMR, because paramagnetic atoms cause line broadening, sometimes to extinction.  For ease of explanation, the PCS is similar to any dipolar coupling, it is a way to relax between atoms, like the NOE, but with a longer distance dependence r^-3 (PCS), vs. r^-6 (NOE).  However, with good decisions like the choice of the ion, the placement of the ion, and so on, you can get subtle effects on your ligand, rather than wiping it out. In the end, you need to know a few things: the actual fraction of ligand bound, the structure of the target (or a good homology model), and the PCS tensor (see below).  This work used rigid, paramagnetic ion binding tags attached to the target via engineered disulfide linkages (CLaNP).

 In total, they made three different tagged proteins and used Yb3+ as the paramagnetic ion and Lu3+ as the diagmagnetic ion. 

This data represents a mixture of bound and free ligand, so using the experimentally determined Kd and the known concentrations of ligand and target, the % bound ligand can be determined.  This can then be converted into PCS of only the bound state. 

However, the authors then tried to calculate the tensor, which is necessary to calculate the orientation of the PCS tensor.  When compared to the orientation of the ligand determined by NOE, there was an 4.7 A RMSD.  This approach only gives the relative location of the binding site.  When they formally calculated the PCS tensors they were able to get a better match of the PCS-derived orientation compared to the NOE-derived, but still not perfect agreement.  That is expected for different methods which can be considered orthogonal.  There ends up being a lengthy discussion of the shortcomings of this method and why it could be possibly better than NOE-based methods, in particular it does not need labeled protein.  However, I would argue if you are not producing your protein in E. coli it is likely being made in insect cells or mammalian cells.  In the case of insect cells, why would you wait two months, to get ligand orientation information on an initial hit?  The project has come and gone on the initial screen hits by that time.

While this is a interesting approach academically, it is really impractical.  Why?  As the authors state, this method is best for ligands with high micromolar to low millimolar affinity.  This positions it firmly in the very early stages of FBHG.  You need to have the structure of the target, or a good homology model.   You need to generate multiple mutants (they do state you can get by with only two positions, but three is better).  You need to do some seriously involved computation; something that is not routine at all.  This would be a much better tool if it could be robustly used at late hit expansion/early lead generation, but that doesn't seem likely.  So, you have what is largely an academic tool for generating models of ligand-target binding with fragments, but not something that would be routinely used.  

NMR as an Impractical Tool, Again

When I was in grad school, I was faced with the choice of two NMR labs to join (after starting as a organic chemist and flirting with enzymology).  Both used NMR, but with very different goals.  One lab used NMR and found systems to study using NMR.  The other studied interesting problems.  The PI would say if NMR is most appropriate than using, but don't be a slave to it.  I have taken that attitude my entire career.  In industry, it also has to be the mantra: best tool for the problem.  Academics tend to have the opposite mindset: let's make my tool work for anything.  

The Krimm lab has been cited here, here, here, here, and here on the blog and I hold their approach to academic tool creation for drug discovery in good regard.  In this paper, they present a combination computational/NMR method for determining if a fragment induces conformational changes in the target.  In their own words: 

The approach relies on the comparison of experimental fragment-induced Chemical Shift Perturbation (CSP) of amine protons to CSP simulated for a set of docked fragment poses, considering the ring-current effect from fragment binding.

Sometimes good people do bad things. 

I am not going to get into the details, but rest assured the science is sound.  Their approach is to evaluate H-N (you could also use H-C, why not) chemicals shifts from titration data to simulated CSPs.  When they did compare experimental with calculated CSPs they could not explain some of these shifts, even when they included ring current-induced shifts.  To further investigate this phenomenon, they used Residual Dipolar Couplings (RDCs) to further explore these unexplained CSPs.  It does.  

What are my problems with this paper?  Practicality, primarily.  Is this another anti-compchem rant?  Nope.  The problem here is that everything they propose to do, and they do it well, relies upon a whole sh!tpile of a priori knowledge: 1. the structure of the protein (typically from X-ray) and 2. the assignments of the protein (not trivial).  Additionally, the RDCs require acquiring two sets of data, aligned and unaligned.  RDCs are wholly impractical.  All of this should red flag this paper as a "Impractical" approach.  It also does not present a method for interrogating structural changes induced by ligands that is any better or more robust than the current standard of analysis.

Not to get too meta on you

If you read this blog, you should be aware that I  think you do not need structure to prosecute fragments.  So, when a paper comes along titled "Toward Rational Fragment-Based Lead Design without 3D Structures" you would think I would plotz.  After reading it, plotz I did, but not from excitement.  

Henen and colleagues (in Robert Konrat's lab in Vienna) present their vision for a structureless future of FBDD.  This future (dystopian for many, utopian for me) is predicated on their meta-structural analysis.  Meta-structural analysis relies on "higher level of abstraction" and "incorporating 3D structural information."  This seems like unfocusing the lens to get a sharper image.  I won't go into the fine details I am sure Peter Kenney has already or would do a much better job if he hasn't.  The general gist is (and trust me I may be misunderstanding this) for a target you predict its secondary structure (because the primary sequence has too low homology to build a reasonable model) and then search for proteins (in DRUGBANK) with known structure and similar secondary structure motifs.  You then get a list of "meta-structurally" similar targets and you make sure they have experimentally-verified ligands. To demonstrate this, the authors show the results of the meta-structure search for lipocalin Q83.

Red structure shows similarity from Q83 while orange shows similarity from the homologs, in this case the beta-barrel structure (A)Strepavidin; (B) FABP, (C) chorismate lyase.  The authors note that chorismate lyase only shares half of the beta-barrel structure.  The further demonstrate this approach, they then chose to go with chorismate lyase and its ligand vanillic acid.  It shares structural similarity with the known Q83 ligand enterobactin.  

A comparison of the solution structure of Q83 (where is my utopian strucutureless future?) and chorismate lyase shows that the two proteins bind their ligands in similar fashion.  From this they concluded that Q83 binds vanillic acid in a 2:1 stoichiometry.  They then used 1H-15N HSQC and ITC to verify the binding of vanillic acid to Q83.  Not surprisingly, it binds. And because they have the assignments (and structure) they found that it binds in the enterobactin binding site.  ITC confirmed that it binds in a 2:1 ratio.  Using the plethora of structural information they have, they wanted to test their ability to link two vanillic acid moieties and went with a "Analog by Catalog"  using the structural insights they have from the meta-structure (and NMR structure) approach with this beast ->.   [To their credit, the authors admit this is not in anyway a reasonable molecule.  I think its most appealing property is that it was available from Sigma-Aldrich.]  They confirmed that this bound to Q83, showing similar chemical shift perturbations as vanillic acid and confirmed by fluorescence quenching.  As their conclusion to this part of the paper, the authors state

Most importantly, it demonstrates that the information needed to rationally improve molecular fragments, found in a first iteration of an FBLD program, is eventually solely provided by meta-structural data without the requirement of a highly resolved crystal structure.

To quote the greatest movie ever, "Opinions vary." [Warning VERY NSFW and very potentially offensive].  As far as I can tell, they didn't have a crystal structure, but a NMR structure.  In my, admittedly biased eyes, that is still structural data. 

In the second part of this paper, the authors repeat this approach with beta-catenin.  But here, they want to show the AFP-NOESY (adiabatic fast passage-NOESY) method they developed can be used for epitope mapping.  The concept here is that:

sizable spin diffusion effects, as a result of the existence of dense hydrogen networks or hydrophobic clusters, lead to measurable shifts of the zero passage toward larger tilt angles.

This seems like "Another (impractical) NMR Method".  STD can deliver the exact same data, has been accepted in the hit discovery  community for a long time, and so on.  I don't see why AFP-NOESY is better.  Anybody care to change my mind?  They end with an example of dynamic combinatorial chemistry to select for a combination of two fragments with improved binding. 

So, did the authors deliver on the promise of their title?  Not in the least.  This is homology modeling by another name.  They still are using structures, just not of the target of interest.  Could it be useful?  Maybe. 

Another (impractical) NMR Screening Method

NMR has a checkered history in drug discovery. In the 90s, it promised to deliver structures just like X-ray. Strike 1! After that, especially after the advent of SAR by NMR, it promised to deliver boatloads of hits from screening. Strike 2! After that, pharmaceutical NMR worked hard to make sure that it was impactful and value-added. It found niches in which it thrives, e.g. a variety of -omics. In drug discovery, NMR still needs to realize it is living with two strikes. How can NMR survive and even thrive? Quite simply. NMR needs to provide rapid, robust, and easily understandable data to medchemists that leads to decisions. Data that results in no action has no value.

In this paper, Salvia et al. present a ligand-based NMR screening method using "long-lived states (LLS)" of the ligand to boost the sensitivity of ligand-based screening. This new method provides 25x better signal-to-noise than established (T1rho) methods and uses less protein. One of the benefits of this method is it allows NMR to study interactions as tight as 100nM and up to 1 mM.

The graphical abstract (above) shows that while this method is very similar in concept to other ligand-based methods (TOP: equilibrium between NMR differentiated states) it requires much more work than these other methods (Bottom: titrations of ligands). The data (Below) does generate quite satisfying curves, and as noted by the authors, are in agreement with previously published values.
I think this work, while an interesting application of Long Lived States, has really no practical value to the screening world. The strength of the binding can be too strong, making the bound lifetime too long, and thus there is a practical floor for Kd. Of course, because it is based upon kinetics, it can be very different for every system.

If you want to determine Kds for a complex < 10uM there are better, far more robust methods (SPR, for one). The amount of time and effort required to generate Kds from this method seems to run contrary to the tenets I described above (rapid, robust, and (most importantly) easily understandable). To me, the title of the paper simply does not deliver. This method is NOT a screening application. A screening application is one experiment (NMR or otherwise) from which you can determine whether a compound is binding or not, ideally from a mixture of compounds.

I would be curious to see in the comments if anyone (especially our NMR savvy readers, you know who you are) think that this method has practical applications.