Skip to main content

Drug Assays

Aggregator Aggravation, In a New Way

Experienced drug discovery folks, particularly those that work early on in the process, will tell you that aggregation is one of the most common sources of false positive “hits”. This happens when the molecule in question bunches up with others of its kind and makes a larger species, particles of something that has different properties than the individual small molecules themselves. Generally, it’s thought that you get into colloidal territory, with the protein target you’re screening against binding to the newly-formed gunk. Your protein (perhaps many molecules of it) ends up tangled in this pond scum and perhaps even denatured by it. (And that’s what gives you the false-positive readout, of course, since from the assay’s perspective the protein, and the signal coming from its activity, has been inactivated, just not the way that you were imagining it).

But this new paper, from Janssen (J&J) shows another way aggregation can trip you up. The authors were screening for small-molecule inhibitors of TNF-alpha, which is a challenging target. Biologics targeting this protein and pathway are out there making billions of dollars, but no one’s ever been able to go after it successfully with traditional medicinal chemistry. When you’re in that sort of space, you have expect that the great, huge, heaping, vast majority of the hits you get from a small-molecule screen are not going to be real. That’s sad, but it’s a fact. If it were straightforward to get real small-molecule hits against TNF-alpha, you’d have seen more of them by now, even failed projects that got published, and that goes for the insulin receptor, or cMyc, or KRAS, or any of the other superstar targets that no one’s been able to get to work with small molecules yet. These things have seen all the standard screening collection stuff and come up wanting. As a side note, that means that you should be very suspicious when you see a paper where someone’s got a hit against one of these things after screening (say) 5,000 commercially available compounds, because the odds are excellent that they’re wrong.

So here’s JNJ525, which really did look like a TNF-alpha-blocking hit. That’s not the most beautiful molecule in the bunch, but none of its parts are ugly. The only thing to worry about is that it’s pretty bulky for a screening hit, and it has the chain-of-sausage-links geometry that’s common to the synthetically common manipulations of aryl groups and amines (SnAr displacements, Buchwald-Hartwig and Suzuki couplings, N-benzylation, etc.) On the other hand, you ‘d have to think (as I’m sure the Janssen folks did) that anything that actually does hit a target like this might well look a bit large and odd, because it’s a protein-protein interface that has never had any evolutionary pressure to bind small molecules.

This compound came in at about 1 micromolar potency, which is certainly worth paying attention to for a screen like this. It was a FRET-based assay, so the first thing to check is that it’s not causing some sort of assay interference, but the compound passed (you could easily imagine a structure like this causing fluorescence artifacts, on the face of it). But the next thing checked was the killer: adding Triton X-100 detergent to the assay affected the binding of positive control a bit (the TNF/antibody  fusion protein Enbrel (etanercept), which is one of those compounds that’s been out there raking in the cash all these years). But it certainly didn’t kill it. On the other hand, detergent hammered the binding curves of JNJ525, which is a Bad Sign.

The group went on the characterize the binding of the compound to TNF-alpha, which is very interesting, but makes me wonder a little bit if they did this characterization first, thinking that it was a real hit, and then later went back and found that Triton killed it. I have a suspicious nature, or it might be that I’ve done it in that order a couple of times myself over the years – just throwing that out there. At any rate, ultracentrifugation suggested that what formed was a dimer of TNF-alpha protein, along with 13 molecules of JNJ525, which is a heck of a lot more defined than the usual aggregation hit. Even more unusually, they actually got a crystal structure of the bound species, and I  again hope that they didn’t do all that before running the detergent assay, but who knows.

In the crystal structure, they can see a single TNF-alpha protein binding to a conglomerate of at least five separate molecules of JNJ525, which are making various contacts all over it (and to each other). The protein’s quaternary structure is altered by all this, as well it might be, and thus the inhibition seen in the assay. It’s real inhibition, but it’s just not really useful.  The paper also mentions SPD-304, first reported in 2005 as a putative small-molecule hit for TNF-alpha (and sold as such by a number of suppliers), but notes that it, too, has its binding killed by addition of detergent. It seems to be having a similar effect on the protein’s structure through a very similar mechanism. There is a crystal structure of SPD-304 bound to TNF-alpha, but the paper mentions that there’s some unresolved/unassigned electron density running around in that structure in the same region as what they see under better resolution.

So aggregation is even more fun than we’d thought – not only do you get huge driftnets of crap forming, you get small, well-defined crap lumps that can send you down the wrong path just as surely. Run detergent controls, is the answer. And if your assay can’t take detergent controls (and some can’t) then you’d better come to terms with the chances that you’re taking if you rely on the assay data that you generate.

33 comments on “Aggregator Aggravation, In a New Way”

  1. not again says:

    JNJ525 would appear to be a poster child for Mothers Against Molecular Obesity (MAMO). Repent J&J, for the wrath of the holy PFI is upon ye!

  2. Peter Kenny says:

    When considering why compounds may behave badly in assays it can be useful to distinguish interference (assay leads to incorrect conclusion as to whether or not compound affects target function) and undesirable mode of action (compound affects target function but not in the way that one wants). Quenching/scavenging singlet oxygen in an AlphaScreen assay would fit into the first category while aggregation would fit into the second category. SPR is arguably the ideal method with which to follow up screening hits since association, dissociation and stoichiometry can all be observed directly. Given the rather imprecise definition of PAINS, I am somewhat mystified as to why aggregators are not classified as PAINS. I have linked the ‘PAINS and editorial policy’ blog post as the URL for this comment.

    1. Hap says:

      Maybe there’s not a consistent/semi-consistent structural type leading to aggregation?

      1. Peter Kenny says:

        Aggregation, like crystallization, is a self-recognition process and therefore I’d expect that it’d be difficult to make useful predictions using substructural patterns. Also there may be variations in the nature of the aggregates formed and the target protein may also influence the extent to which aggregate formation takes place. Ionization under assay conditions (the compound shown in the post could potentially form a dication)also needs to be considered because an aggregate of charged molecules will need to incorporate counter-ions if it is to be stable.

        1. Hap says:

          I just assumed (mistake, I know) that PAINS were substructures that were proposed to have activity in lots of assays. If you can’t assign a substructure likely to cause nonproductive aggregation in assays, then wouldn’t it not be a PAINS anymore? The compound’s activity in assays would not have been sufficiently determined, but you wouldn’t be able to generalize further.

          1. Peter Kenny says:

            The PAINS acronym translates as Pan Assay INterference compoundS so strictly it should only be applied to compounds. That said, it also gets applied to substructures and it is not always clear if compounds asserted to be PAINS have actually done nasty things in many assays or are simply believed to do so. It is worth asking whether compounds matching PAINS filters but not hitting in any assays in the PAINS assay panel should still be considered to be PAINS. When discussing bad actors in assays, it can be difficult to know whether one is dealing with fact, prediction or belief. This is a theme explored in the blog post linked as the URL for this post.

        2. Druid says:

          I think dimerization can occur favorably when the pH is near the pKa of a base. One molecule is protonated and the other neutral molecule forms a strong hydrogen bond to the protonated amine, while the lipophilic parts of the molecule align. No counterion is needed. It is not so simple in larger aggregates, or course, but aggregation starts with dimerization. This process can often be followed by changes in the UV absorption spectrum as the concentration increases or the pH is adjusted.

      2. Liqc says:

        An intuitive correlate of solubility and aggregation propensity could be melting point. Of course it’s rarely known until a crystalline form of a candidate is designed.

  3. Steve says:

    I don’t run a lot of high throughput stuff but it seems to me that any such assay should always incorporate some detergent just to avoid nonspecific hydrophobic interactions in the first place. I do that with low-throughput assays all the time.

  4. NJBiologist says:

    “Even more unusually, they actually got a crystal structure of the bound species, and I again hope that they didn’t do all that before running the detergent assay, but who knows.”

    Maybe they had the detergent data and the unusually-defined stoichiometry, and made the crystal to figure out what on earth was going on?

  5. Curious Wavefunction says:

    I sometimes wonder whether aggregation based binding and dissociation might have played a role during the origin of life.

  6. Konrad Koehler says:

    Perhaps this is not a total loss. Only two of the six ligands in the complex form extensive contacts with the protein and one of these forms three hydrogen bonds to the receptor. If the ligand could be modified to disfavor ligand-ligand contacts …

    It is also curious that the crystal structure was deposited in the PDB (accession code 5MU8), but the PDB reports that it has been withdrawn.

    1. Konrad Koehler says:

      Opps. Misread the PDB. The structure is merely unreleased. Not withdrawn.

  7. anonymouse says:

    TNF trimer can certainly bind suitable (drug) molecules within trimer. Crystallographically verified at near-atomic resolution, no artifact. Potency is not great, however.

  8. Barry says:

    A decade after Shoichet’s work on aggregators, what assay formats still “can’t take detergent controls”? Are they irreplaceable? A rescreen in the presence of Triton X-100 seems like a cheap experiment that should be routinely run before things like co-crystallization.

    1. HTSguy says:

      Cell-based assays, for one example. And yes, aggregators do foul up cell-based assays for membrane proteins:

      doi: 10.1021/jm301749y

  9. Anonymous says:

    Puzzled by the observation of 5 compounds bound per dimer of dimers in the crystal structure… from the methods it seems the co-crystallization was setup with less than the 13:2 stoichiometry that the centrifugation suggested (3.3 mM of compound to 48 mg/ml protein). If the solution state of the aggregate is prone to forming a 13-mer that binds the protein, then why does it switch to 5 in the crystal? Also, 48 copies of the protein in a P1 space group would not be a fun structure to try and solve!

  10. sgcox says:

    TNFalpha is 17 KDa protein so 48 mg/ml is 2.7 mM (if I am not mistaken). Pretty close to 5 comp to 4 (dimer of dimer) in the crystal cell. Something me and others usually do – mix about equivalent amount of protein and inhibitor in the first go and see what happens 😉

  11. Mr. Magoo says:

    Does the Hill coefficient not do a good job of identifying aggregators?

    1. Chris Swain says:

      Since the aggregation is a concentration based affect, they can give quite nice dose-response curves unfortunately, with reasonable Hill slopes.

      1. Peter Kenny says:

        Chris, the recent ACS editorial ‘ The Ecstasy and Agony of Assay Interference Compounds’ (I’ve linked it as the URL for this comment) makes the following comment on aggregators:

        “Noncompetitive inhibition with high Hill slopes. There are classical reasons for noncompetitive inhibition and for cooperative binding, but the latter is rare in early discovery and the two together suggest aggregation”

        I do have more than a few issues (e.g. do frequent-hitter results from a proprietary data set count as ‘known’ assay interference) with this editorial but that’s what the Editors are saying.

  12. Chris says:

    Slightly irritatingly whist the paper is published the X-ray structure is not yet released

  13. Pharmacist says:

    Derek (or anyone else): Could you say a little about detergent controls? What is the purpose and what sort of information do you get from them? Thanks.

    1. Peter S. Shenkin says:

      These so-called “promiscuous aggregators” are generally bound by means of hydrophobic interactions. So a little bit of detergent breaks up the aggregates, exactly the same way dishwashing detergent solubilizes grease. So if the inactivation of the protein is due to its sequesterization (in some sense) in or by such an aggregate, detergent will restore the activity. If you are looking for inhibition, this of course means that you have not really found an inhibitor.

      What is unusual in this case is that the aggregate formed is specific enough to be crystallized; usually, such aggregates are thought of as being rather ill-defined structures, similar to micelles, But then again, we don’t really know, because it’s not clear than anybody else has tried to crystallize these things. Maybe they’re all like this! (“If one has seen one tiger in the woods, might the woods not be filled with tigers?”)

      There are a few subtleties that have long been known but can easily be overlooked. Shoichet pointed out that there are known drugs which also act as aggregators in screens. I don’t think the implication of this is that aggregation is the mechanism of their action, but rather that they are strong inhibitors as well. Usually, the therapeutic dose is orders of magnitude lower than the concentration of the compound in a lead-finding experiment such as HTS. So compounds that are aggregators at high concentrations can still be real inhibitors at therapeutic doses. This has implications for lead-discover. Even mild inhibitory activity that persists with detergent, if real and specific, might make a compound worth bringing forward to the optimization stage, despite the fact that it’s a promiscuous aggregator.

  14. Sandro says:

    Interesting paper, which brings some more questions. Is this an isolate case or not? Among aggregate-forming compounds, how any act via sequestration / protein denaturation and how many via such a mechanism?

    An interesting comment from the authors: “all ligands show different contacts to the protein or other ligands, which precludes a straightforward evaluation of structure −
    activity relationships for TNFα binding and aggregation”. This suggests that more compounds were made & evaluated. It would be interesting to know if anything was observed, for instance around the pyrimidine, which does make some H-bonds in the crystal structure.

    1. DrugHunter says:

      “Homogeneous aggregate” – no longer an oxymoron I suppose. I take some solace in the observation that apparent ligand-ligand and ligand-protein interaction specificity suggests this should indeed be a very rare event.

      1. STACking says:

        It may not be as rare as you think. There are other examples in the literature and there would probably be more, except, aside from big pharma, most institutions don’t have the resources or time to co-crystallize would-be aggregators. Take SIRT1 activators for example – their mode of interaction appears to be driven by dimeric association/aggregation (see Figs. 2d and 3e). Imagine trying to do structure-based design around that motif. Notably GSK sirtuin R&D has gone very quiet.

  15. Daniel Barkalow says:

    Given that they got a paper out of a false positive, I’m not sure you can say they were wrong to investigate how the compound generated a hit even after finding that detergent broke it up. If you’re going after TNF-alpha, the best you may be able to realistically hope for is an interesting failure.

  16. LiqC says:

    Fun times. My own brief experience in medicinal chemistry involved looking for allosteric inhibitors of a viral protease. Those allosteric sites were discovered by soaking protease crystals in a large collection of concentrated solutions of small fragments, of which one wiggles its way into the crystal. A computational screening effort suggested a number of directions in which a molecule can be grown. Nobody cared to pay attention to the fact that this very shallow site existed only in the crystal, at the protein-protein interface between lattice neighbors.

    I recused myself upon receipt of a crystal structure containing molecules of my “allosteric” inhibitor in the active site. Four of them.

    Derek, since it FBDD is your bread and butter, can you please confirm that if you see crystal structures with molecules in shallow pockets, you don’t go chasing after analogues until you check how proteins pack in the crystal?

    1. Derek Lowe says:

      Oh, you’re absolutely right. You have to make sure that the “binding pocket” isn’t an artifact of the crystalline state, and high-concentration soaks are prone to giving you stuff like that. Proceed with caution!

  17. FormerXtalGrower says:

    Always worth bearing in mind that crystallographers produce “A crystal structure” never “THE crystal structure.” That electron density doesn’t assign itself and a crystal structure is simply the scientist’s interpretation of the map (refined of course with cross-validation and calculation of the requisite statistics to measure structure quality). Those statistics describe the whole of the structure, but the devil is always in the details, right alongside all the most interesting scientific conclusions. Don’t expect to use R-free as an indicator of local problems, like how well a ligand is fit to the electron density (to their credit the PDB has done an excellent job recently by bringing these issues to the fore with standardized validation reports, ligand reports, etc). But in this case, ask your crystallographer colleagues what they would think of a structure at 3 Angstrom that reported an Rsym of “3 6.8%” (sheesh, i hope that’s a typo) and refined 800 kDa worth of a 17 kDa protein in an asymmetric unit where all the equivalent protein chains show nothing more than non-crystallographic symmetry. There is actually more electron density shown in the paper for the previously published structure SPD-304 than there is for the ordered aggregate of JNJ525. Just like any other crystal structure, this too merits a close look.

  18. SPQR says:

    Derek, you write “because it’s a protein-protein interface that has never had any evolutionary pressure to bind small molecules.”. Excuse me if it is a stupid question, I am a biologist, not chemist, but are you saying that proteins have generally evolved to bind small molecules?

    I always imagined that drug binding capability was more of a happenstance due to irregular structure in folded peptides than actual evolved function.

Comments are closed.