07 June 2021

A minimal fragment library for maximal coverage of pharmacophore space

Last week we described a fragment library built with the aid of machine learning and designed to contain privileged fragments that should produce high hit rates. Unfortunately, only about a tenth of the library members are commercially available, so it will be some time before we know whether the design was successful. We continue the theme of fragment libraries with a just published Nat. Commun. paper by György Keserű (Hungarian Research Centre for Natural Sciences) and a large group of multinational collaborators (see also here for a nice summary by György).
 
The researchers started by analyzing more than 3300 crystal structures of protein-fragment complexes in the protein data bank. Fragments were defined as having 10-16 non-hydrogen atoms, and the computational approach FTMap was used to ensure that fragments were binding at hotspots as opposed to spurious, less ligandable sites. This exercise yielded 3584 fragments, but many of them were identical or very similar to one another. The researchers used a series of computational tools to cluster similar fragments (or pharmacophores) and choose a set that would maximize diversity. This ultimately led them to assemble a library of just 96 fragments, purchased from five vendors.
 
This SpotXplorer0 library mostly follows the rule of three, with 7 to 17 non-hydrogen atoms, MW 100-250 (or 280 for bromine-containing molecules), ≤ 3 hydrogen bond donors, ≤ 8 hydrogen bond acceptors, and ≤ 3 rotatable bonds. In addition, all members have 1-3 rings, no more than a single halogen or sulfur atom, and no PAINS. Despite the small size, this library covers most of the pharmacophores identified in the larger set, and considerably more than the F2X-Entry fragment library we highlighted last year or the top five commercial library vendors we noted here.
 
The researchers then screened this library against eight targets. Three GPCRs (the serotonin receptors 5-HT1A, 5-HT6, and 5-HT7) were assessed in a cell-based radioligand displacement assay with fragments at just 10 µM. Despite the low concentration, 4-11 hits were found. Biochemical screens conducted at 800 µM against the proteases thrombin and Factor Xa yielded 7 and 8 hits respectively. Further analysis revealed that the SpotXplorer0 ligands sampled a majority of the pharmacophores found in published fragment hits against theses five targets.
 
Next the researchers screened their library against the histone methyltransferase SETD2, an oncology target with few known attractive ligands. An enzymatic assay yielded two hits, with IC50 values between 300 and 500 µM.
 
Finally, the SpotXplorer0 library was part of the XChem crystallographic screens against the SARS-CoV-2 main protease (Mpro) and Nsp3 macrodomain, which we discussed here and here. For Mpro, just a single hit was found. This is only half the overall hit rate for noncovalent fragments in the crystallographic screen against this target, but the hit is functionally active and has a high ligand efficiency.
 
The screen against NSP3 yielded five hits binding at two different sites, for a hit rate of 5.2%. The overall hit rate against this target was 8%, but that encompasses screens against two crystal forms of the protein. The crystal form used for SpotXplorer0 had a hit rate of 21%.
 
In summary, SpotXplorer0 is new fragment library that gives high coverage of experimental pharmacophore space. Laudably, structures of all 96 fragments are provided in the Supplementary Information. But the jury remains out on how hit-rich the library will be. Interestingly, the F2X-Entry library we highlighted last year gave considerably higher hit rates of 21% and 30%, albeit against two different targets. SpotXplorer0 is being screened crystallographically against multiple targets at XChem, and it will be interesting to see how it performs in the long run.

01 June 2021

New fragments suggested by machine learning

Machine learning has become a hot new thang in drug discovery, attracting massive attention and investment. While easy to parody, artificial intelligence techniques are becoming increasingly powerful. A new paper in J. Chem Inf. Mod. by Angelo Pugliese and colleagues at the Beatson Institute applies the methodology to generate a new fragment library.
 
Machine learning entails collecting large amounts of data, passing that through various neural networks, and obtaining recommendations. In this case, the researchers wanted to generate “privileged fragments” that would hit in multiple assays. (Of course, the idea would be to make genuinely privileged fragments, such as 4-azaindole, rather than PAINS.) The researchers used a training set of 66 fragments that hit in at least three of 25 screens done at the Beatson, for which the average hit rate was 2.18%.
 
First though, the researchers needed to teach their model how to generate chemically valid fragments in the first place (for example, fewer than 5 bonds to carbon). To do this they used both SMILES (simplified molecular-input line-entry system) and chemical fingerprints from a set of 486,565 commercially available fragments. They then combined this model with the privileged fragments. Extensive details are provided; as they go well beyond my expertise I won’t even attempt to summarize them. (For example, “the classifier for the smi2smi model comprised sequential 64-unit and 32-unit dense ReLU layers followed by a single sigmoid output neuron.”) At the end of the exercise, and after triaging by medicinal chemists, the researchers came up with a set of 741 fragments.
 
What are their overall properties? For one thing, generated fragments tend to be more planar (as assessed by PBF) and have lower Fsp3 values than the nearly half-million fragments used for training. The researchers acknowledge that this could reflect the historical composition of the Beatson fragment library, although as we argued here it could also be true that flatter fragments just give higher hit rates.
 
Molecular complexity is a fundamental but poorly defined aspect of fragment-based lead discovery, and the researchers have come up with their own metric, called feature complexity (FeCo), which incorporates information on rotatable bonds, numbers of halogens, hydrogen bond donors and acceptors, charged groups, aromatic rings, and hydrophobic elements, all normalized by the number of heavy atoms. Hopefully this will be explored more fully in a dedicated publication.
 
What do the individual fragments actually look like? Five examples are shown in the paper, and nearly 200 more are provided in the supporting information. Below are seven chosen arbitrarily from that list (sampling every 30 structures).
 

Of course, the question remains as to whether these fragments will truly turn out to be privileged. As might be expected given the vastness of chemical space, only 78 of the 741 are commercially available. The researchers note that they are acquiring some of these, and it will be interesting to see how they perform in the screens to come.

28 May 2021

Sotorasib: the fifth fragment-derived drug approved

Today the US FDA approved sotorasib (AMG 510) for a subset of patients with non-small cell lung cancer. This is the fifth fragment-derived drug approved.
 
Last year Practical Fragments outlined the discovery of sotorasib, tracing its origins to two independent fragment-based efforts. (Disclosure: I was part of this work at my previous company, Carmot Therapeutics.) Today's approval illustrates another aspect of the story: the speed with which the program progressed. From the Nature publication demonstrating that KRASG12C is ligandable it took just five years to deliver sotorasib to the clinic, and less than three to demonstrate sufficient safety and efficacy to win approval.
 
This is fast for any drug, and all the more so for a target previously deemed undruggable.
 
In addition to commending the scientists and clinicians, credit is also due to Amgen, which cleared all bureaucratic hurdles to push this program forward. Every company with even a passing interest in cancer saw the 2013 Nature paper, but many were put off by the covalent nature of the molecule. It took organizational vision - in addition to great science - to succeed.
 
It is also worth noting that of the five fragment-derived drugs now approved, two are for difficult targets; venetoclax blocks a protein-protein interaction.
 
Of course, the raison d'être for all these efforts is to help patients, and the emerging clinical data for sotorasib clearly demonstrate this. Of 126 patients with advanced, previously treated non-small cell lung cancer, 43 had partial responses, and 3 had complete responses. For those not familiar with cancer research this may not sound impressive, but these are patients with no other options. And this is just the beginning for sotorasib, the first drug to work via this long-sought mechanism. Combination therapies are already actively being pursued, and these often show dramatic improvements as physicians figure out how best to use new drugs.
 
Congratulations - and thanks - to everyone involved in bringing this new gift to humanity.

24 May 2021

Sixteenth Annual Fragment-Based Drug Discovery Meeting

Last week Cambridge Healthtech Institute held its Sixteenth Drug Discovery Chemistry (DDC) meeting virtually for the second year in a row. The eight tracks, four in parallel, spanned three days. There were occasional kinks, and the 6:30 morning start time was painful for those of us on the Western edge of the US, but for the most part things went smoothly. Attendance was up 25% over last year, to 750 participants. Attendees from outside the US comprised 40%, similar to last year and up from 2019, while industry representation remained steady at 70%. As always I’ll just touch on a few broad themes – please feel free to add your impressions.
 
Methods
These meetings are always great venues to learn about new techniques, and one that stood out is “mass photometry” (MP). This label-free method allows visualization of single proteins in solution. The technology, which builds on interference reflection microscopy and interferometric scattering microscopy, rapidly provides information on the molecular weight of proteins to an accuracy of about 2 kDa. Stefan Geschwindner (AstraZeneca) used the method to follow a change in the monomer-dimer equilibrium of the protein PAD4 in response to ligand-binding. He noted that the instrument, sold by Refyn, provides measurements in about 10 seconds for proteins from 35 kDa to 4 MDa in size.
 
On the opposite end of reductionism, Bryan Roth (University of North Carolina, Chapel Hill) discussed customized high-throughput cell-based screens for 320 of the 358 GPCRs in the human genome. We’ve previously described how these assays have been used to screen molecules prioritized from massive computational screens. If you’d like to profile your own molecules, you can submit compounds here for free with no IP-entanglements.
 
Cryo-EM was the subject of our April 1 post last year, but as we discussed here the technique is no joke. Pamela Williams discussed how Astex has formed a consortium with several other companies to advance the technology. She noted that of the 270 membrane protein structures released last year with resolutions better than 3.5 Å, 167 came from cryo-EM while only 103 came from crystallography. Unlike in crystallography, resolution can vary over the protein; local resolution around the region of interest may be better or worse than global resolution. And there is still much to be done to truly industrialize the approach, particularly in sample preparation (specifically making the grids).
 
Speaking of crystallography, James Fraser (University of California, San Francisco) discussed a high-throughput crystallographic screen against the SARS-CoV-2 macrodomain protein, which we covered earlier this year. A fragment-linking program called “Fragmentstein” was used to generate molecules with low micromolar affinity.
 
At the last DDC meeting Frank von Delft (Diamond Light Source) described a high-throughput crystallographic screen of the SARS-CoV-2 MPro protein, and this year he provided an update on how the open-source COVID Moonshot effort has advanced hits to non-covalent 30 nanomolar binders with 100 nanomolar antiviral activity in cells. Current efforts are focused on improving metabolic stability and oral bioavailability, but Frank lamented the “astonishing inertia” on the part of funding agencies to support the effort. Although the spectacular success of vaccines should contain the outbreak, there are still occasional breakthrough infections, and not everyone can be vaccinated or mount a successful immune response. Small molecule drugs will be useful, and it would be nice if more people were working on them.
 
Success Stories
I’ll never turn down a crystal structure, but it is important to remember that fragments can be advanced even in the absence of structural information. As a case in point, Paolo di Fruscia (AstraZeneca) described the discovery of allosteric binders of MEK1; we wrote about these earlier this year.
 
On the other hand, Justin Dietrich described AbbVie’s discovery of TNFα inhibitors – an effort he said would not have been possible without structural information. As we discussed here, the team struggled with the lipophilic nature of some of their initial molecules; Justin noted that “if a protein wants grease you have to find a way to get it grease.” The physicochemical properties can then be improved during optimization.
 
Rod Hubbard (Vernalis) described the discovery of LpxC inhibitors. He emphasized the importance of using biophysical methods to carefully characterize and optimize the target protein, and noted that NMR can identify fragments that crystallography might miss. Also, as we noted when we wrote about this campaign last December, fragments can provide multiple starting points, and it can be useful to optimize fragments before embarking on a fragment-to-lead effort.
 
Jus Singh (Ankaa) provided a brief history of targeted covalent drugs. Some folks are still nervous about the potential for covalent molecules to cause idiosyncratic toxicity; indeed, this was one of the reasons Frank von Delft gave for pursuing non-covalent MPro inhibitors. However, Jus noted that five out of seven approved covalent kinase inhibitors are not associated with liver toxicity, while some non-covalent kinases inhibitors are. Jus also noted that all three FDA-approved BTK inhibitors are covalent, despite longstanding efforts to develop non-covalent inhibitors.
 
Finally, Micah Steffek (Genentech) described the discovery of LC3 binders, the first step in making an autophagy-based version of PROTACs. An NMR screen yielded hundreds of hits, but despite obtaining multiple crystal structures the team struggled to obtain molecules with better than mM potency. However, combining fragment information with results from a DEL-screen, akin to the story we described here, led to submicromolar inhibitors. Intriguingly, these covalently modified a lysine residue.
 
Odd and ends
Micah’s talk led to some discussion in a speaker panel as to the meaning of hit rates. Do you consider the primary hit rate or the confirmed hit rate? And what level of confirmation is required? As Rod Hubbard noted, different techniques may give different answers for good reasons, and you don’t want to throw away potential hits on the basis of the worst performing or least sensitive technique.
 
Mads Clausen (Danish Technical University) described his shapely library of fluorinated fragments (see here). Hit rates seem to be similar to flatter compounds, though with only four targets screened it is perhaps too early to draw conclusions.
 
But perhaps there is a limit to shapeliness: Justin Dietrich could not recall seeing any fully unsaturated Fsp3 = 1) fragments coming up as hits, despite being present in the AbbVie libraries. Andreas Lingel said he had occasionally seen these at Novartis, but they were never pursued. Similarly, no one could recall advancing any acyclic fragments. A quick glance at the 131 fragment-to-lead success stories captured in five annual reviews revealed that all of the fragments had at least one ring and one sp2-hybridized carbon.
 
And with that I’ll close. The Seventeenth installment of DDC is scheduled for April 18-21 of next year in San Diego, and the more biology-focused Discovery on Target is scheduled to take place both online and in-person in Boston from September 27-30. Hope to see you there!

17 May 2021

Understanding fully-functionalized diazirine tags

Four years ago we highlighted work out of Ben Cravatt’s lab describing “fully functionalized fragments,” or FFFs (see here for a schematic). In addition to the variable fragment portion, FFFs contain a photoaffinity tag that can react with proteins plus an alkyne moiety that allows the labeled proteins to be captured for identification by Western blots or mass spectrometry. In the 2017 paper FFFs were used to identify fragments binding to proteins in cells, and more recently other researchers have used the same approach to find fragment hits against isolated proteins and even RNA. In all these cases the photoaffinity tag used has been a diazirine. But how do diazirines react with proteins?
 
As more researchers work with FFFs, it is important to consider what affects the reactivity of the probes. This question is addressed in two new papers.
 
The first, by Christina Woo and collaborators at Harvard, Dana-Farber Cancer Institute and Jnana appears in J. Am. Chem. Soc. The researchers looked at alkyl diazirines (exemplified by LD below – the most commonly used photoaffinity tag) as well as aryl fluorodiazirines (exemplified by Ar below). When treated with ultraviolet light, both moieties form diazo intermediates that can lose nitrogen to generate highly reactive carbenes. However, alkyl diazo intermediates can also react with carboxylic acids, while aryl diazo intermediates do not.
 
The researchers assessed the reactivity of alkyl and aryl diazirines with individual amino acids. In aqueous solution, the specific aryl diazirine tested did not react with any amino acid, while the alkyl diazirine reacted with Glu, Asp, and – at higher concentrations – Tyr and Cys. Moreover, reaction with Glu and Asp was pH-dependent; at higher pH, when the carboxylic acids were deprotonated, the reaction did not occur. A similar pH effect was observed with the model protein bovine serum albumin and the alkyl diazirine probe.
 
Next, the researchers tested a panel of 32 FFFs in intact cells or cell lysates (typically at 10 µM, with 60 s photoirradiation). Positively charged probes tended to give better labeling, suggesting that these FFFs were binding near acidic Glu and Asp residues on proteins. Positively charged fragments yielded an average of 50 unique binding sites, while neutral FFFs produced an average of 14 and negatively charged FFFs gave only 5. A closer look at some of the specific proteins revealed patches of electronegativity in the regions labeled. Interestingly, and consistent with earlier work, membrane proteins were particularly enriched, possibly because Glu and Asp residues in membrane proteins often have elevated pKa values and are thus likely to be more reactive with the diazo intermediates.
 
The second paper, in Chem Sci. by Christopher Parker and collaborators at Scripps Research Institute in Jupiter, FL and Bristol Myers Squibb, explores five different diazirine tags. In addition to the conventional LD probe, they examined an aryl and three alternative alkyl diazirines. All five (LD, Ar, BD, DF, and Tm) were appended to either a simple phenyl substituent (controls), a positively charged lipophilic fragment (FFFs), the promiscuous kinase inhibitor staurosporine, or the bromodomain ligand JQ1.
 

All the controls and FFFs could modify proteins in cells in a dose dependent manner after treatment with UV light. The Ar control appeared to give more non-specific binding, and the LD, Ar, and BD-tagged fragments produced more robust labeling than the Df and Tm fragments.
 
As for the target-specific probes, LD-, BD-, and Tm-tagged staurosporine each labeled between 529 and 836 proteins, but only 10-21 kinases – far fewer than the hundreds of kinases staurosporine inhibits. (To be fair, the photoaffinity tag did reduce the affinity of the probes for protein kinase A and abolished it entirely for the Ar probe.) Among the JQ1-derived probes, only the BD- and Tm-tags pulled down one or two bromodomains.
 
There are many critical but subtle details in both papers. For example, the DF and Tm tags require shorter wavelength illumination (300-310 nm) than the more conventional tags (~360 nm). Changing the wavelength can shift the balance between diazo intermediates versus carbenes. The diazirines can also react with water and buffers or generate olefins, though the latter reaction can be disfavored by deuterating the methylenes on either side of the diazirine.
 
So what does all this mean? To state the obvious, this is complex stuff, and small changes to the tag or conditions can completely change the outcome of the experiment. As the kinase and bromodomain examples show, a negative result does not mean your ligand does not bind to a given target, while a positive result says nothing about the strength of the interaction. Perhaps the “understanding” promised in the title of this post has some way to go. But photoaffinity tagging and chemoproteomics make a powerful combination, and these papers contribute to helping us figure out how to use this tool.

10 May 2021

Fragments vs MAT2A: AstraZeneca’s turn

Last week we highlighted work out of Agios leading to a clinical compound for the oncology target MAT2a. But given the appeal of new cancer targets, Agios is not without competition. In a recent (open access) J. Med. Chem. paper, Claudia De Fusco, Marianne Schimple, and colleagues at AstraZeneca describe their efforts.
 
The researchers started with fragment screens using thermal shift, crystallography, and SPR, with the latter providing the most potent and efficient hits. In particular, compounds 1 and 2 stood out and were subsequently found crystallographically to bind at the same allosteric site discussed last week. Interestingly, despite its higher affinity, compound 2 was inactive in a functional assay while compound 1 had an IC50 similar to its dissociation constant.

 
Initial efforts to improve the affinity of compound 1 were unsuccessful, but merging the two fragments yielded compound 5, which had comparable affinity to compound 2 and was also functionally active. Attempts to tweak the dimethylamine moiety were mostly unsuccessful, and a high-resolution (1.1 Å) crystal structure combined with quantum mechanical calculations revealed that the substituents around the aromatic quinazolinone ring system were twisted somewhat out of the plane.
 
Moving to the other side of the molecule, addition of the phenyl ring originally present in compound 2 gave a satisfying boost in affinity for compound 28, which was attributed in part to displacing two “unhappy” water molecules.
 
Compound 28 is active in cells and has good pharmacokinetics in rats. Unfortunately, it is cleared more rapidly in mice. Subcutaneous dosing in mouse xenograft models led to tumor stasis, though weight loss was also observed, suggesting toxicity. Though that likely precludes more intensive in vivo work, compound 28 could still be a useful chemical probe.
 
This paper is a nice example of fragment merging and growing. Notably, compound 28 is relatively small and very ligand efficient. Perhaps combining information from both this and the Agios series will lead to even better molecules.

03 May 2021

Fragments in the clinic: AG-270

A promising oncology approach is to target “synthetic lethal” proteins that are required for cancer cells but not for ordinary cells. Methionine adenosyltransferase 2A (MAT2A) is a metabolic enzyme that produces S-adenosyl methionine (SAM). Cancer cells lacking another gene, methylthioadenosine phosphorylase (MTAP), appear particularly dependent on MAT2A. In a recent J. Med. Chem. paper, Zenon Konteatis and colleagues at Agios, Viva, and ChemPartner describe the first clinical inhibitor of MAT2A.
 
The researchers began by screening >2000 fragments in pools of 20, each at 50 µM, using ultrafiltration. The 31 hits were tested in enzymatic and SPR assays, and compound 1 confirmed in both. Testing 54 commercially available analogs led to compound 2, with low micromolar activity, and this molecule was profiled intensively.
 

Kinetic studies revealed compound 2 to be non-competitive with respect to substrates ATP and L-methionine, and crystallography confirmed that compound 2 binds in a previously discovered allosteric pocket. The molecule makes polar and hydrophobic contacts to MAT2A and also displaces several water molecules. SAR studies found a preference for aromatic moieties at the 2-position of the central core, but the phenyl off the 3-position could be substituted with more shapely moieties such as the piperidine in compound 9.
 
Closer examination of the structure of compound 2 bound to MAT2A revealed a protein-bound water molecule, and displacing this with the phenol in AGI-24512 led to a satisfying boost in biochemical potency as well as cell activity. However, the molecule has poor oral absorption and a short half-life in rats. Metabolite identification studies pinned the blame on the piperidine, with the phenol no doubt doing no favors. Medicinal chemistry ultimately led to AGI-25696, which despite its lower biochemical activity was active in cells, metabolically stable, and showed efficacy in a mouse xenograft model when dosed orally.
 
Despite these favorable properties, AGI-25696 has very high protein binding in human plasma (>99.9%) as well as high efflux, which would likely necessitate a high clinical dose. The researchers proposed that, due to the weakly acidic nature of the molecule, it could tautomerize, and each tautomer could bind to different plasma proteins. Simply methylating the N-H led to decreased plasma protein binding but also lower binding to MAT2A. Thus, the researchers sought to shield the N-H by forming an intramolecular hydrogen bond. After appending more than 70 heterocycles they eventually arrived at AG-270, which has a more respectable plasma protein binding of around 98.5%.
 
Extensive characterization of AG-270 revealed it to be potent with good pharmacokinetics and oral bioavailability. It is relatively clean in a panel of 95 potential off-targets and showed tumor growth reduction in xenograft models. But it is not without warts: low solubility necessitated a spray-dried dispersion for animal dosing, not surprising giving its high lipophilicity. Nonetheless, the molecule has entered a phase 1 clinical trial in patients with MTAP loss.
 
This is a lovely fragment-to-lead success story with several lessons. First, although enzymes are sometimes considered “easy,” this is not necessarily true. Indeed, at an ACS meeting in 2018 Anil Padyana mentioned that metabolic enzymes in particular often have shallow, polar active sites. Targeting allosteric sites, as done here, can be a useful alternative.
 
Second, it is striking how the core of the initial fragment remains intact in the clinical compound, a reminder of the power of fragments to efficiently explore chemical space.
 
Finally, this story is another important reminder that affinity is often just the beginning of a long journey. It took considerable effort to optimize the pharmaceutical properties from AGI-24512 to AG-270, including a decrease in ligand efficiency. In the end the team has succeeded, and Practical Fragments wishes them – and the patients – luck in the trials.

26 April 2021

STD NMR on putative SARS-CoV-2 main protease ligands

Over the past sixteen months SARS-CoV-2 has infected more than 146 million people worldwide and killed over 3 million. Highly effective vaccines are now available, but not everywhere, and how long the vaccinations will last as new variants arise remains unknown. COVID-19 will likely be with us indefinitely, necessitating drugs as well as vaccines.
 
More than a year ago we highlighted the COVID Moonshot effort, which began with a crystallographic screen against the essential main protease (Mpro) to find starting points for drug discovery. In a new open-access J. Biomol. NMR paper, Ioannis Vakonakis and collaborators at University of Oxford and University of Patras have attempted to characterize some of the hits using saturation transfer difference NMR (STD NMR, see here for a brief description).
 
The researchers had access to a 950 MHz NMR (jealous much?). Samples were screened at 10 µM protein using irradiation of a Mpro methyl group with a chemical shift at 0.5 ppm. To try to minimize differences in relaxation parameters among different ligands, only the strongest STD signals in aromatic moieties were examined.
 
Of 39 non-covalent ligands discovered in the crystallographic screen, five either did not produce an NMR signal or the spectra were inconsistent with the expected structures, suggesting the ligands may be insoluble or unstable in aqueous buffer. The remaining 34 compounds were nominally screened at 0.8 mM each, but the reference spectra differed in intensity by as much as 15-fold, suggesting dramatic differences in concentration. Since the strength of an STD signal is related to both affinity and concentration, this could obviously complicate interpretation of results. As we’ve written previously, careful curation of your library is essential.
 
Of thirteen active site ligands, only four showed strong STD signals. Dose-response titrations between 0.05 and 4 mM revealed dissociation constants of 1.6-1.7 mM for two of them, with the other two being too weak to accurately measure. Molecular dynamics simulations starting with the known structures were consistent with these results, with the tighter binders tending to maintain their positions more than the weaker binders.
 
The researchers also characterized 650 elaborated molecules from the COVID Moonshot, some of which had been reported to be nanomolar inhibitors. Disturbingly, 35 gave no NMR signal and another 86 yielded weak signals. Among those remaining there was a weak correlation (R2=30%) between IC50 in an enzymatic assay and the STDratio (integrated signal intensity of peaks in the STD spectrum over reference spectrum). STD NMR is not appropriate for molecules with Kd < 10 µM, so the researchers also used a competition experiment in which four putative high-affinity molecules would compete a weaker “spy” fragment. This exercise confirmed two ligands but not two others, calling into question their mechanism.
 
STD NMR is often used as part of an assay cascade prior to attempting crystallography, but as crystallography throughput increases there is a case for starting with crystallography, as we argued five years ago. Results from the 39 crystallographic hits perhaps gives pause to that notion, or at least emphasize the need for confirmatory assays. It is easy to be seduced by a high-resolution structure, but because of its sensitivity crystallography may identify ligands so weak as to be unadvanceable. As for the 650 elaborated molecules, it’s too early to draw conclusions, though it’s good to always be on the lookout for false positives.
 
Hopefully the COVID Moonshot will ultimately lead to drugs against SARS-CoV-2. But even if it doesn’t, the intensive focus of multiple techniques on a few proteins is providing useful guidance and best practices that will be applicable to other targets.

19 April 2021

Fragments vs KEAP1: deconstruction and merging

One of the more challenging protein-protein interactions targeted by drug hunters is the interface between the transcription factor NRF2 and its repressor KEAP1. This is part of the cellular defense against reactive oxygen species; increasing NRF2 activity may be useful for treating a variety of diseases. Unfortunately, the binding site on KEAP1 that interacts with NRF2 is large and has a predilection for carboxylic acids. Thus, many of the molecules reported as inhibitors tend not to be druglike. Anders Bach (University of Copenhagen) and a multinational team of collaborators sought to do better, and have just published some of their journey in J. Med. Chem.
 
The researchers had previously tested 19 reported small-molecule KEAP1 inhibitors, of which only nine confirmed. (This is a salutary reminder to take any individual publication with a large grain of salt.) The nine fell into six chemical series (two shown below), and the researchers decided to fragment some of these molecules into 77 fragments. The fragments were then tested in four assays: fluorescence polarization (FP), a thermal shift assay (TSA), saturation transfer difference (STD) NMR, and surface plasmon resonance (SPR).
 
Primary hit rates were generally high, from 25%-64%, but long-time readers will not be surprised that the overlap was not great: no fragments hit in all four assays, and only eight hit in three. As the researchers point out, this could reflect differences in sensitivity, conditions (from 3-8% DMSO and from 0.5 to 8 mM fragment), and different types of false positives and false negatives. Interestingly, and in contrast to previous work, overlap was good between STD NMR and SPR.
 
Crystal structures of seven hits were solved bound to the protein, and compounds 4c and 1m (from different precursor molecules) were merged to provide compound 8, with low micromolar affinity. Compound 8 was the subject of considerable medicinal chemistry, with five different vectors chosen for growing. Despite being structurally enabled, the researchers struggled; changes that improved affinity in one context did not do so in another. After considerable effort, the researchers obtained compound 77o, with mid-nanomolar activity.
 

Compound 77o is stable in human plasma and mouse liver microsomes. Unfortunately, and unsurprisingly given the two carboxylic acids, it has poor permeability. Indeed, a fragment-derived KEAP1 inhibitor we described previously has only a single carboxylic acid, as does precursor compound 7. As the researchers themselves acknowledge, “the physicochemical properties of our compounds are not favorable for membrane permeability.”
 
Nonetheless, this paper is a lovely example of fragment-based deconstruction reconstruction (FBDR) and is well worth studying for the thorough descriptions of fragment screening in orthogonal assays and structure-based design. Another lesson may be that despite considerable effort, the final molecule is far from a chemical probe, let alone a drug. Perhaps some targets truly are undruggable. Or maybe – as for other seemingly undruggable targets – a change in strategy is needed.

12 April 2021

Fragment merging on c-MET

Fragment-based inhibitors of kinases are legion, particularly those that bind in the so-called hinge region where the adenine of ATP normally sits. However, even among these there are many different flavors of inhibitors. In particular, about 10 kinases can adopt a “folded P-loop” conformation, in which the phosphate-binding loop collapses into the ATP binding site. This was the focus of a recent open-access paper in ACS Med. Chem. Lett. by Gavin Collie and colleagues at AstraZeneca.
 
The researchers were interested in the oncology target c-MET. A ligand-based NMR screen of 1150 fragments (in pools of 6 at 200 µM each) yielded a 6% hit rate, of which 20 confirmed by SPR. Crystallography was attempted unsuccessfully on most of these, but compound 1 was found to snuggle into the active site with the protein in the folded P-loop conformation.
 
A computational similarity search of AstraZeneca’s internal library identified compound 2, which crystallography revealed to bind in a similar manner, with two hydrogen bonds to the hinge region and the benzyl group buried in a hydrophobic pocket. A second similarity search of the library – this time based on compound 2 – identified compound 3. Crystallography confirmed that the core azaindole moieties of compounds 2 and 3 overlay, and thus fragment merging was attempted.
 

The resulting compound 5 bound as expected. This prompted yet another computational search of the internal library, and after a bit of medicinal chemistry compound 7 was identified as a mid-nanomolar inhibitor with low micromolar cell-based activity. Crystallography revealed that it too binds to the folded P-loop conformation of c-MET.
 
Because the folded P-loop conformation is rare among kinases, the researchers hoped that the resulting molecule would be selective. Unfortunately, when profiled against a panel of 140 kinases at the low concentration of 100 nM, 27 of them were inhibited by at least 60%. This is perhaps not surprising given the 7-azaindole core, which has been found to bind to more than 90 kinases, though some compounds containing this moiety are selective.
 
Nonetheless, this paper is a nice example of structure-guided fragment merging. A cynic could point out that had the researchers screened the entire AstraZeneca compound collection they likely would have identified molecules very similar to compound 7 anyway, but this may have cost more and would not be an option at smaller organizations without million-compound libraries. And the approach is useful for more difficult targets for which high-affinity molecules may not exist – yet.

05 April 2021

A general fragment-based approach to… targeting RNA?

This is taken from the title of a recent open-access paper by Matthew Disney and collaborators at Scripps Research Institute Jupiter and Florida Atlantic University in Proc. Nat. Acad. Sci. USA. RNA has long been a target of FBLD: Practical Fragments first blogged about it in 2009, and a 2002 paper reported using fragment linking to obtain a low micromolar binder. So how general is the new approach?
 
The researchers describe chemical cross-linking and isolation by pull-down fragment mapping (Chem-CLIP-Frag-Map). This involves using photoaffinity probes that can crosslink to biomolecules such as RNA. The probes also have an alkyne tag that can be used to isolate bound molecules using click chemistry. We’ve written previously about such “fully functionalized fragments” (FFFs).
 
Earlier work had resulted in the identification of compound 1, which binds to a specific site on pre-miR-21, the precursor to a non-coding microRNA linked to cancer. An FFF version of compound 1 was shown to crosslink to pre-miR-21 after irradiation with UV light, and the site of modification could be mapped using a reverse-transcriptase-mediated primer extension, which stalled at the modified bases.
 
Next, the researchers screened 460 FFFs at 100 µM and found 21 that crosslinked to pre-miR-21. They were ultimately looking to link fragments with compound 1, and thus competition studies were used to eliminate fragments that bound at the same site. This left three fragments, and primer extension studies confirmed that these bound near but not at the binding site of compound 1.
 
Next, the researchers attached these three fragments to compound 1, with or without various linkers. Some of the resulting molecules had improved affinity, and compound 9 showed the tightest binding according to microscale thermophoresis (MST). Mutational and competition studies confirmed that the molecule binds to the expected site. Importantly, compound 9 not only bound to pre-miR-21, it also blocked processing by the enzyme Dicer. Moreover, it showed activity in cell models consistent with inhibition of pre-miR-21.
 

This is a nice paper, but there are several limitations. First, compound 9 is still a fairly modest binder with lackluster ligand efficiency. Indeed, while potency can be overrated, I would love to see a fully synthetic low nanomolar RNA binder. Second, while the approach may be general, it is not necessarily easy, and it requires specialized fragments. And as we noted last year, there is no relation between crosslinking efficiency and affinity. I wish the researchers had tried linking some of the non-selected fragments to see whether these were false negatives. Indeed, given the complexity of the approach, I wonder if the researchers would have been better off simply making and testing an anchor library around compound 1, in a similar fashion as described here.
 
But whether or not Chem-CLIP-Frag-Map turns out to be the solution to targeting RNA, I wholeheartedly agree with the conclusion: “It may be time to describe biomolecules that are perceived to be challenging small molecule targets as ‘not yet drugged’ rather than ‘undruggable.’”