Data Release and Updates: 2020-06-12

Hi all,

We’ve been busy collecting data! I’ll try to give an overview on recent progress.

In case you are wondering, I was just looking at the graph below, and we just yesterday hit the mark of over 500 compounds assayed. We also now have over 100 structures from Diamond UK of ligands bound to the protein.

And as the graph shows, many compounds to come! Compounds are mainly being made at Sai Life Sciences (India), Enamine (Ukraine), and WuXi (China), then shipped for testing to Oxford, UK (Diamond UK + Oxford U.) and Israel (Weizmann Institute). This all shows the incredible international nature of this community, which also involves many of you submitting ideas (now over 7000 designs) from other diverse parts of the world.

Currently Promising Compounds

The compounds

@edgriffen created a very nice slide of 5 potential starting points so far (Tour of Learning from Docking, modelling and crystal structures), which I will shamelessly copy below. Note that the slide is slightly outdated. For example, we have since realized that the sulfone-piperazines with a chloroacetamide warhead are much more reactive than other piperazine-chloracetamides (I’ll explain more below).

As you may notice, there are two IC50 values listed in the slide. This is because we now have assays up and running at both Weizmann (Fluorescense) and Oxford (RapidFire MS)! Even better, they generally tend to agree, as shown in this plot at https://covid.postera.ai/covid/activity_data:

Structures of these Compounds

As the slide also shows, we have structures on four of these compounds, and we do have an Ugi structure, just not the compound shown. @markc has posted an excellent overview of how these ligands fit in the binding pocket here Tour of Learning from Docking, modelling and crystal structures. I will again shamelessly copy some of the slides below.

First, Mark helpfully orients us in the binding pocket by using the structure of the N3 ligand (PDB: 6LU7) from https://www.nature.com/articles/s41586-020-2223-y_reference.pdf

an Ugi structure

The first structure I’ll look at is an Ugi compound with an acrylamide warhead
LON-WEI-b8d98729-43
, which is actually a pretty weak binder (23% inhibition at 50 uM). @MarkC provides a nice overlay of the structure (x2694 on Fragalysis) with the N3 ligand:


a chloroacetamide structure

LON-WEI-8f408cad-3 (x0689)

chloracetamide_identity

@MarkC shows that this piperazine with a chloroacetamide warhead (likely very reactive) reaches into the P1’ pocket, where the Ugi compound (also shown) does not reach.

Obviously, understanding the reactivity of both the chloroacetamide and acrylamide warheads, and how they contribute to reversible/irreversible inhibition is vital, which I’ll go into more a bit later.

A non-covalent 3-aminopyridine structure

TRY-UNI-714a760b-6 (x2646)
Fluorescence IC50: 24.57 uM
RapidFire IC50: 13.90 uM

This is one of the compounds that most of the med-chemists of the group immediately zeroed in on due to its relatively good ligand efficiency.

As you can see from the structure, the compound spans the P1-P2 sites aminopyridine_identity

The overlay with the Ugi compound shows that there is still a lot of room to explore in orthogonal directions.

a non-covalent quinolone structure

MAT-POS-916a2c5a-2 (x2910)

This one is also non-covalent and spans the P1-P2 pockets.

IC50 (µM) - Fluorescence: 5.83
IC50 (µM) - RapidFire: 3.46

The overlay with the 3-aminopyridine above shows the clear difference in the linker binding.

I’ll also note that this one was dug up as I was reading through a 2009 high throughput screen of the original SARS-CoV main protease https://pubchem.ncbi.nlm.nih.gov/bioassay/1890 . It’s nice to know that the activity we observe is quite similar in this main protease.

In related news, @londonir has done a great job of setting up a new HTS at Weizmann for this MPro. It’s about halfway done, and some initial hits (still need confirmation) can be seen here

one last Beta-Lactam structure

BAR-COM-4e090d3a-39 (x2562)

This one is actually a weak binder (22.7% inhibition at 50 uM), but was of interest because the lactam binds closer to the p5 pocket.

I’ll mention that this was a whirlwind overview of the structures. It’s worth exploring them (and the many more available structures) yourself on Fragalysis. Getting into the details, you should also check out @RGlen’s excellent points here Tour of Learning from Docking, modelling and crystal structures. He knows this structure really well, and there are some nasty, intricate details that he covers.

Expanding on the Initial Hits

We’ve actually been able to very quickly do quite a lot of exploration of the 3-aminopyridine and Ugi scaffolds – just by exploiting quickly made compounds. See all the results so far at https://covid.postera.ai/covid/activity_data

@londonir’s lab has led the exploration of the Ugi scaffold. I looked through it, and we have so far assayed ~120 Ugi compounds. In order to better understand what was happening, I made a table like this of the activity data, which may be useful to some people (3 most potent compounds shown)


The full table is available here

You all have partially led the initial exploration of the 3-aminopyridines – and many more compounds are on the way. JOR-UNI-2fc98d0b-12 is down to 3 uM; however, we are yet to see any massive gains in potency over TRY-UNI-714a760b-6. But we have some SAR along the way!

Also, if you haven’t yet, definitely check out @pwalters blog post where he explores some of these relationships.

I have to say we didn’t do the best job of making people mention which series they were expanding upon. Therefore, I had to write some nasty SMARTS strings to grab the current expansion assay data, but here it is for the amino-pyridines if you want it! (will also be up on the site soon):

amino_pyridine_activity.xlsx (498.2 KB) (more strict 3-aminopyridine scaffold)
amino_pyridine_like.xlsx (536.7 KB) (slightly less strict scaffold)

We also have very preliminary exploration of the quinolone scaffold, which I am attaching here if you want to play around with the data yourself:
quinolinone_activity.xlsx (84.9 KB)

Lastly, there are quite a few piperazine-chloroacetamides at the top of the activity data list https://covid.postera.ai/covid/activity_data The improving potency despite a relatively constant warhead structure did hint that some improved recognition was happening. However, with these chloroacetamides understanding the reactivity is key. Which brings me to my last point… we finally have some very preliminary K_inact/K_i data, as many of you have been asking about for a while

K_inact/K_i

First, a big thank you to @Wal-Ward, who reached out to offer his enzymology expertise in analyzing some of this data. [Update: he has kindly posted his initial analysis here Biochemical assay kinetics] I’ll reiterate that these results are very preliminary.

@londonir had his group measure K_inact/K_i for 13 of the most potent covalent compounds.

As is expected, the chloroacetamides are quite reactive (if we take K_inact as an imperfect proxy for “reactivity”). The sulfone-piperazines as especially reactive. The Ugi acrylamide has by far the lowest K_inact, as expected – though the K_i still could be greatly improved. Warheads aside, some of the med-chemists pointed out that it also looks like potency is being driven by lipophilicity in a few of these compounds.

For reference, @londonir provided the K_inact values for some approved irreversible kinase inhibitors:

5.60E-03 Acalabrutinib
2.70E-02 Ibrutinib
1.80E-03 Dacomitinib
2.40E-03 Afatinib
1.10E-03 Neratinib
1.1E-01 Osimertinib

These aren’t terribly different values from some shown here. The difference is that the K_i in those drugs is orders of magnitude better. Room for improvement.

In the next couple of days, hopefully we will have more K_inact/K_i data and also will share the full protocol.

Last thoughts

We are now constantly pushing new data through the assays and getting back new structures (the Diamond UK beamlime is back up soon); therefore, I will try to produce these data releases more frequently!

Also note, that these results mainly represent the testing of the earliest compounds that were sent off for synthesis. We should have lots more results both exploring new, more complex scaffolds, and expanding on these early hits in the near future.

Requests
If you want to buy compounds for testing / send already made compounds for testing: Please see How to send moonshot compounds for testing Some of you have already had your compounds tested!

Questions
I’m sure many of you have questions. Please reply to this thread if so, We’ll try to respond.

Thanks
And I just want to thank once again all of the incredible people involved in this project – behind all of this data is a lot of dedicated people working around the clock.

-Matt, and the rest of the PostEra team helping out on this project (@alphalee, @aaron.morris, @ajajack, @milan.cvitkovic)

4 Likes

Docking Benchmarking

Oh, and something else I should share! We have received a lot of design ideas using docking and/or other computational methods. There also seems to be a lot of new papers out almost every day using these methods for finding new drugs/ repurposing. @alphalee and I thought about how one might evaluate these methods, and came up with the following proposal for benchmarking.

There are a lot of SARS-CoV 3CLPro inhibitors out there, which have also shown pan-coronavirus activity. Others have also shown that the SARS-CoV and SARS-CoV-2 3CLPro active sites are very similar (see this thread by members of the community, for example, Similarity between CoV proteases and their inhibitors). Thus, we think it is important to understand:

(a) does docking successfully prioritize active molecules for SARS-CoV

(b) does FEP reproduce rank ordering of affinity against SARS-CoV 3CLPro,

© if b is answered in the affirmative, does going from SARS-CoV to SARS-CoV-2 changes the ordering of analogues.

More precisely, we suggest the following proposal:

  1. Dock (or use method of choice) molecules tested in this HTS against SARS-CoV (https://pubchem.ncbi.nlm.nih.gov/bioassay/1706). What is the AUC/PRC if we use dock score as a classifier between active/inactive?

  2. Run relative FEP (or other method of choice) to compute the relative IC50 of the attached molecular series against SARS-CoV 3CLPro (https://www.rcsb.org/structure/1UK4]). We have collected 6 non-covalent series - series 3, 4, 5, 6 are non-peptide small molecules.

  3. Run relative FEP (or other method of choice) to compute the relative IC50 of the attached molecular series against SARS-CoV-2 3CLPro (structure from Diamond)

Internally, we have had a few members of the Moonshot computational community obtain preliminary results. There seems to be some dependence on what SAR-CoV structure is chosen, but understanding this variability is likely important if we are going to trust results on these proteins. We are interested in others’ results on this benchmarking project (both good and bad). Hopefully, this may be a way forward on evaluating the quality of the methods and their applicability to this protein.

molecular_series_sars (1).xlsx (15.6 KB)

1 Like

Fantastic work, everyone. Special thanks to Matt for pulling this all together.

Regarding the covalent compounds characterised by Nir and colleagues, some of these (eg CVD654 & maybe others) appear promising for further optimization, with small size and low lipophilicity giving scope to improve affinity (Ki) and possibly reactivity (kinact). Having crystal structures of complexes puts the project in a good position.

Regarding the kinact values listed for kinase inhibitors. Afatinib has an initial Ki = 0.16 nM, which is crucial for its activity, especially as it has a GSH half life of only 14 min (t1/2 = 0.693/k). The 55 nM figure stated for osimertinib actually is the initial Ki. The kinact is 0.11 s-1, with a GSH half-life of 121 min.

I strongly recommend measuring the rate constants for stability in reduced GSH (kGSH), because it gives insight into stability and intrinsic reactivity of the compounds (ideally not too high in order to avoid off-target reactions and ligand depletion by reaction with intracellular GSH). Values for kGSH allow calculation of the enzyme rate enhancement factor (EREF =kinact/kGSH), a useful metric, which reflects the proximity effect coming from specific binding to the target enzyme. This value is only 2.9 for afatinib (the low Ki probably is responsible for efficacy) and higher at 1200 for osimertinib (more dependent upon specific reaction with the target).

Take care! Wal

1 Like

Thanks @Wal-Ward, great comments – and I believe a kGSH assay is in the works. I have also edited my post to reflect the correct value of osimertinib Kinact. Thanks for catching that!

Regarding benchmarking the docking, wouldn’t it make more sense to (also) create an internal active/inactive/decoy set? (using internal data + for instance the DUD-e webserver)

Hey @bart.lenselink, yes that is perhaps a good idea. We still only have a few active scaffolds , but we could start to try to put it together.

1 Like

I’ve taken a look at the feasibility of linking ‘reversible’ warheads to x2646 and prepared these notes:

Using the amide nitrogen of x2646 as a synthetic handle has implications for trans/cis preferences of the amide group and there are some highly relevant measured data in this article (ref 9 in the notes):

I have submitted structures 3 and 4 from the notes as designs:

https://covid.postera.ai/covid/submissions/bbe8d7ff-23b7-4848-a830-4c2ce8fa64e9

1 Like

Hi Matt and rest of PostEra team

I think that it would be better to separate the reversible and irreversible inhibitors in data summaries. Structures of irreversibly-bound, covalent inhibitors are not especially relevant to their inhibitory activity and I would recommend generation of transition state models if the plan is to take the irreversible inhibitors forward. This could be in a number of ways and one approach would be to build a structurally-prototypical transition state and then to elaborate it structurally (i.e. add the substituents) in order to generate the transition states corresponding to the different inhibitors.

It would also be an idea to think about what the ideal inhibitor ‘profile’ with respect to k.inact and K.i would be. For example, would the profile for a kinase inhibitor be appropriate (or achievable) when targeting a catalytic cysteine residue? There seems to be an error in the unit (reciprocal micromolar) for K.i in the table.

I’ll link my ligand efficiency article again since the approach suggested for defining efficiency could also be used to normalize k.inact, K.i or k.inact/K.i values (with respect to both molecular size and lipophilicity):

I’ve posted a message on this thread about potential structural elaboration of the x2646 fragment-derived inhibitor (3-aminopyridine) and it is important to be aware that synthetic elaboration at the amide nitrogen is likely to impact on the cis/trans conformational preference of the amide. The 3-chlorophenyl group of this inhibitor occupies the S2 subsite but is something of a dead end because it does not allow access in the direction towards the S5 subsite. My guess is that it will be necessary to replace the aromatic ring with something that does allow access and it may be worth looking at the P2 substituent SAR of known peptidomimetic inhibitors.

I’m guessing that the 2-methoxyphenyl substituent in the quinolone x2910 may also need to be replaced in order to provide access in the direction of the S5 subsite and it will be more difficult to incorporate a reversible warhead than for x2646. The catalytic cysteine sulfur appears to directed at amide carbonyl carbon of x2910 and it may be worth taking a close look at the electron density to see if there is evidence for a productive interaction. The quinolone NH does not appear (at least in the modelled structure) to be donating a hydrogen bond to the protein and there are ways in which the nitrogen could be moved elsewhere in the ring.

Something that makes me a bit nervous about the quinolone is the (electron-withdrawing) amide carbonyl group at C4 (analogy with fumaramide). In general, I think that it is a good idea to check the reversibility even for non-covalent inhibitors.

1 Like

Hi Wal,

It may also be useful to think about how to use glutathione reactivity measurements might be used to evaluate reversible warheads. For example, can we measure equilibrium constants?

Is kGSH a pseudo first order rate constant or a second order rate constant?

If progressing irreversible inhibitors, we need to be thinking about how to assess selectivity with respect to other cysteine proteases and other enzymes that use cysteine catalytically. Some of the cathepsins are considered to be therapeutic targets but there will be a number of enzymes that we won’t want to hit.

1 Like

Hi Pete
The following reference gives information on using GSH reactivity in the optimisation to osimertinib. You may find it useful. They use a pseudo-first order rate constant for reaction with excess GSH. Ward, RA et al, 2013, J Med Chem 56, 7025-48.
All the best
Wal

1 Like

Thanks @pwkenny, I will get this info to the people who need it. And I will try to adjust my summaries in the future to make clear the different strands of inhibitors.

Just some notes/questions on a few of your points.

Thanks! We have seen your designs PET-UNK-bbe8d7ff, which were of interest to the team. There is a bit of a debate to as whether one starts by building more potency then attach a warhead at the end, or to attach warheads now. But I believe these designs will be tried (I have to confirm).

TRY-UNI-2eddb1ff-7 seems to be a new merge building on the 3-chlorophenyl group with a beta-lactam likely reaching towards P5. We’ll need to see a structure for sure, but this compound (4 uM IC50) did show improved potency.

Very good point, I will raise the issue with some of the people who’ve looked closely at this one.

I’ll also point them to this post for much more info outside those highlighted points. Thanks again for the detailed response.

Hi Wal,

Thanks for the reference (I recognize some of the authors). I certainly agree that it would be a good idea to measure rate constants for reaction with reduced GSH. Hopefully the data can be used to assess the viability of chloroacetamides and other irreversible warheads and to establish guidelines for design.

Looking at exposure in terms of a competition between target and GSH for the compound raises questions about how to account for the concentration and location of target within the cell. I’m guessing that intracellular concentration will be greater than Ki for some of the kinase inhibitors but will be less than Ki for the irreversible inhibitors from the Moonshot project.

While industry efforts against cathepsins have typically focused on reversible inhibition, irreversible inhibitors, such as the vinyl sulfone K777 feature in the antiparasitic literature. I believe that K777 was being readied for clinical development and there may have been some tox although I’ve not seen any details. If the team has not done so already, it might be an idea to try to make contact with people who have been involved with (at least) preclinical development of irreversible cysteine protease inhibitors. What is the team’s view with respect to potential anti-targets for irreversible cysteine protease inhibitors?

1 Like

Hi Matt,

Thanks for the update and please keep me posted on status of the designs. I’m guessing that we will ultimately need to cyclize in order to present a reversible warhead to the catalytic cysteine in a low energy conformation although the designed compounds should give us an idea about potential benefits of targeting the catalytic cysteine in this manner. This information may be transferable to other structural series. Specifically, we would get an idea of nitrile/aldehyde affinity differences (I’m currently assuming that the nitrile will be significantly less potent although I would be delighted to be proven wrong).

I believe that the substituent will need to make contact with a concave region of the molecular surface of the protein (like where a DMSO molecule binds in the pdb:6y2f crystal structure) if growing in the direction of P5 is to result useful gains in affinity. I don’t think that this will be feasible if you need to maintain the penetration of the S2 subsite by the chloro subsituent. The beta lactam ring would be regarded as potentially reactive by many medicinal chemists and I think that it would be important to have a strong rationale for using it as a molecular recognition element.

1 Like

Hi Pete

Thanks for your comments. My simplistic view is that reactivity with GSH is a very approximate surrogate for many off-target effects, including depletion of compound by reaction with intracellular GSH and other proteins (eg serum albumin), together with a crude measure of propensity to react off-target to give undesired effects.

I also am simplistic in my thinking about effects of target concentration. A typical drug has a total peak plasma concentration of around 1 uM, of which about 5% or 50 nM is free. I assume that the intracellular concentration is similar to that in plasma (not always the case, in part due to P-glycoprotein and other transporters). When the concentration of free compound in the target cytoplasmic compartment is depleted by covalent or noncovalent binding, then I assume a degree of re-equilibration occurs, so that the drug is effectively replenished to some extent from the pool of compound outside the cell. I accept that equilibration may not be complete or instantaneous and that PK effects will have significant influence, but overall, I don’t expect there to be a big issue due to depletion by binding to target in the cystoplasm. I think off target effects are much more likely to compromise efficacy by depleting compound, simply because there are more off target molecules. Of course biosynthesis of fresh target protein of could compromise efficacy and duration of action…

I hope these musings are useful. Best wishes

Wal

2 Likes

To add to this PPB is an important parameter to consider (usually plasma protein binding <95% desirable) which can differ if other meds are used (highly likely for a covid patient!)… and drug drug interactions. This is very early stage, i know, but having ideas of some lead-like molecules soon in terms of CYP isoform selectivity (usually IC50>3-10 uM)) would be useful to see how these compounds would work when a very ill patient has several other drugs in their system that could also affect their concentrations? We can go on with PAMPA, efflux ratios. everything… it’s a true minefield!!!

1 Like

We have PPB in the pre-in vivo testing plan. Generally I’m with Genentech in that it’s important to understand, but not optimise against (https://pubs.acs.org/doi/abs/10.1021/jm5007935) particularly as very highly bound compounds are at risk of high inter and intra patient variability and therefore represent a toxicity risk. For the LO phase there is also the risk of issues as we move between species as very different PPB between species can lead to overdosing / underdosing and concomitant toxic effects or lack of expected PD effect.
We also have CYPs in the plan - we’ll screen for them as an issue once we have good enough leads - the consensus view amongst the medicinal chemists on the team is CYPs are always something you can optimise away from if you have an issue.

1 Like

Super, sure. Aldehyde oxidase is another one that’s poorly dealt with. We’ve been using the Baran Litmus test and it works a treat (CHF2 substitution correlates with AO susceptibility). Are triazoles A.O. substrates? Obviously to bear in mind @ a later stage.
BTW, latest J Med Chem is very useful I’m sure; it;'s on drug tox and lead optmisation strategies;

June 25, 2020 Volume 63, Issue 12
Pages 6249 - 6574
DOI: 10.1021/acs.jmedchem.9b00917.

Hi Wal,

I think that we are in broad agreement on the value and interpretation of the GSH stability assay and I agree that what happens to a compound outside cells is likely to be much more important, from the perspective of clearance, than what happens to a compound inside cells. I would guess that we should be able to use measured kGSH and typical GSH plasma levels to calculate a lower bound for the clearance of a compound that reacted irreversibly with GSH.

When I was working on cathepsins, the view was that problems associated with covalent drug-protein adducts would go away if a sufficiently low dose could be achieved. The covalent cathepsin inhibitors were reversible (nitriles) and the equilibrium constants (not measured) would have been more relevant than the rate constants that were actually measured. While I would guess that reactivity of chloroacetamides with the main protease can be improved significantly, it is likely to be more difficult to reduce the GSH reactivity of chloroacetamides since the warhead is peripheral within molecular structures that incorporate it.

My view is that one needs to have a strong rationale (e.g. essential for target engagement) to justify pursuing irreversible inhibitors given the additional complexity that these bring to design and PK/PD modelling and the general undesirability of irreversible drug-protein adduct formation for anything other than the primary targets. I would not regard the success of irreversible kinase inhibitors as adequate justification for pursuing chloroacetamides, It can be argued that the benefits of the warhead in the kinase inhibitors are more due to the recognition of the cysteine residue that to the irreversible nature of the drug with this residue. Reversible covalent bond formation might even be the better option for the kinase inhibitors (although I’d conjecture that this would be more difficult to achieve with non-catalytic cysteines).

1 Like

Hi Ed,

I would agree with monitoring PPB rather than explicitly optimizing it. My understanding is that PPB is not really an issue for drug-drug interactions unless doses are sufficiently high as to saturate the PP. Also, PPB needs to be seen in the broader context of distribution (an increase in fraction unbound from 1% to 5% is not necessarily going to lead to a 5-fold increase in unbound concentration). This article may be helpful if you’ve not seen it already:

https://doi.org/10.1038/nrd3287

In addition to the generic anti-targets (CYPs, hERG, etc), I would recommend that the team think about which cysteine proteases (and other enzymes that use catalytic cysteine) should be considered as anti-targets. This is especially important if targeting the catalytic cysteine (whether irreversibly or reversibly)

1 Like