I’ve very much been enjoying this paper, on fragment-based chemogenomics in whole cells. That’s the sort of blurb, I admit, that’s probably going to make you immediately want to read the rest of the paper, or immediately go do something else (all the way up to “podiatrist appointment”). And I understand those impulses. But if you’re into new targets and drug discovery, you had better read on. It’s quite a ride. (Update: here’s a post on it at Practical Fragments as well).
It’s a collaboration from the Cravatt group at Scripps, the University of Lausanne, Bristol-Myers Squibb, and the Salk Institute, and what they’re doing is, in a way, the more demanding follow-up to the covalent fragments work that’s been published before. In that case, you have at least labeled the proteins that you’re hitting, which gives you a handle to figure out what they are. But how do you determine that with a “regular” library of compounds or fragments? The interactions will take place within the cell, and you know that they’re going on in there, but how do you find out what’s been happening?
Well, under current technology there still has to be a covalent bond formed somewhere, so what’s going on in this case is the popular photoaffinity/click-label combination. The paper investigates 14 simple chemical motifs (a benzofuran, a phenylpiperidine, a benzyl N-methylpiperazine, etc.), and each of these fragments have the same short alkyne-and-diazirene combo chain appended to them. If you soak these compound into cells for a bit and then hit them with UV light, the diazirene will do its thing and covalently label any nearby protein side chains. That’s not a perfect process, as chemical biologists well know, but it’s one of the best we’ve got in that line. You’ll get false negatives, from things where the labeling geometry or timing were a bit off and nothing happened, but the positives tend to be largely real. At this point you can lyse the cells and use some sort of azide-containing reagent to pick out the proteins that got tagged with your alkyne fragments.
What they found, in a pass through the well-studied HEK293T cells, is that each of the 14 fragment probes gave a different protein labeling profile. Timing didn’t seem to be too important – five minutes on the cells gave about the same results as an hour. Washing the cells before the UV exposure decreased a lot of the labeling (but interestingly, not all of it). And running the same experiment with a control probe, which had the same diazirene alkyne but was substituted only with a methyl group at the fragment end, picked out the ones where it was the labeling groups doing the binding, rather than the fragments (and there weren’t too many). Looking at the resulting gels, it seems that the benzofuran picked up the fewest proteins, and to the least degree, while a benzhydrylpiperazine species labeled things like crazy. (I would really enjoy extending this work to some other structural motifs, making sure to throw in some of the groups that tend to make people raise their eyebrows (SF5 aryls, etc.) just to see how they all compare).
So far, this section by itself would make a very nice publication, honestly, but this isn’t one of those papers. What I’ve described so far is the first paragraph of the Results section; this thing just keeps on going. The next part gets quantitative, using the SILAC mass spec technique. For those who don’t know it, that’s an ingenious use of isotopic labeling. In this case, you grow the cells on isotopically enriched media, so that the proteins get heavier by incorporating stable isotopes like 13C. You treat a batch of normal “light” cells with a fragment probe of interest, while treating a batch of “heavy” cells with the control methyl-only probe. You then lyse both sets, combine the two lysates, and then process them together – click labeling, cleanup, etc. Now when you run these things down the LC, you’ll get peaks of labeled proteins, and the light and heavy ones will have the same retention times (the peaks will overlap). But when you do the mass spec on them, you can see how much of each peak is labeled light protein, versus labeled heavy protein, and that tells you, quantitatively, what the labeling is for your probe over the background. They followed this up (with appropriate control experiments) with several probe-versus-probe comparisons, in addition to the probe-versus-methyl control ones.
So what showed up? Both soluble and membrane-bound proteins were labeled by the various probe fragments, and as mentioned above, each compound’s labeling varied, with each of them picking out some proteins that the others didn’t. Interestingly, only 17% of the proteins labeled have known ligands, and most of those were enzymes. The others tended to be transcription factors, regulatory proteins, channels, transporters, and a generous helping of proteins of unknown function entirely. I find this sort of thing very hopeful; it’s like getting a quantitative printout on things we haven’t even tried to do yet.
By digesting these proteins and looking for the labeled pieces, it was also possible to localize the sites where the fragments had bound. For the proteins that have known structures, these overlapped significantly with known pockets and binding sites (but keep in mind that most of the hit proteins have no structures worked out yet at all). Overall, there was only about a 15% overlap with the hits from the previous reactive fragment screening work, indicating that these compounds are picking up on a truly different set of sites.
They went on to pick out an enzyme (PTGR2) and a transporter (SLC25A20) that came up as hits with specific fragments, because neither of them are reported to have useful small-molecule binders yet. Doing some quick SAR on the isolated recombinant proteins led, in both cases, to analogs with improve activity that fit the definitions of useful chemical probes (and indeed, possible leads for drug discovery efforts). Again, this sort of thing is extremely encouraging.
So we’re up to about three or four papers worth of results now, but there’s more. The team realized that since the fragment screen had worked so well, that slightly more elaborated molecules might furnish a collection for outright phenotypic screening. So they made 465 compounds, all with the diazirene/alkyne tail, and set out to do just that. For most of them, they also synthesized the “plain” version, where there was just a propyl chain in the place of the diazirene/alkyne, as comparators. The phenotypic screen they chose was a well-worked-out one: differentiation of 3T3 cells into adipocytes, which can be measured quite easily by looking at Nile Red staining of the lipid droplets in the actual adipocytes. 9 compounds showed greater than 3x activity versus controls, and none of them turned out to be PPAR-gamma ligands (which is one of the first things you’d think of; the positive control for the assay was in fact the PPAR-active drug rosiglitazone).
Working off the most active compound of the bunch, which also promoted several known differentiation protein markers (and did not have such effects on non adipogenic cell lines), they confirmed that the non-reporter-labeled “plain” analog was also active (several structural analogs were also identified as inactive, which is reassuring). Using the SILAC technique to quantify and identify the proteins that were labeled, they tracked down PGRMC2 (progesterone receptor membrane component 2) as the likely target – it was, in fact, the only protein that made it through all the control experiments. You can be forgiven for not having heard of it, because despite the name, its function is very poorly sketched out so far in the literature. They tracked down the exact site of labeling to a predicted cytochrome-binding domain, which I might add is not exactly the first place you’d go looking for a small-molecule binding pocket.
So what’s the compound doing? The first guess would be that it’s messing something up, and that this loss of function is promoting adipogenesis in the phenotypic assay. That’s just because most of the time our small molecules are gumming up the works somewhere. But shRNA experiments showed that knocking down PGRMC2 did not set off differentiation into adipocytes, and in fact blocked the effect of the compound once it was added to those cell lines. This suggests that it’s acting as a gain-of-function ligand, which is even more interesting. (They go into some further experiments to suggest how this might be happening, just for good measure).
I’ll let the authors sum up:
We have described herein a chemical proteomic method to globally map small-molecule fragment-protein interactions directly in human cells. More than 2,000 fragment-binding proteins were discovered, only a small fraction of which was found in DrugBank, highlighting the broad and still largely untapped ligandability of the human proteome. We demonstrated that the discovered fragment-protein interactions can be further advanced to generate selective ligands that modulate the functions of proteins in cells. That the case studies investigated herein include enzymes (PTGR2), transporters (SLC25A20), and poorly characterized transmembrane proteins (PGRMC2) for which selective ligands were previously lacking, underscores the versatility and scope of chemical proteomics for accelerating the discovery of small-molecule probes for diverse categories of proteins.
Yes, yes it does. People have been using the photoaffinity/click combination for some time, but this is the biggest, widest application of it that I’ve yet seen, and it points the way into a huge amount of useful research. If you’re into phenotypic screening, difficult targets, and early-stage drug discovery in general, you have to pay attention to this work. Read it now! I’ve been saying for years that fragment-based screens and phenotypic screening are two worlds that never intersect, but here is the pathway between them. The fragment portion of the paper demonstrates that the compounds are engaged in real, meaningful, and often selective interactions with proteins in living cells, and the phenotypic screen in the latter half – the business ends of whose structures are still well within fragment parameters – show the power of this technique to identify useful, functional ligands.
This is one of the most interesting, useful, encouraging and thought-provoking chemical biology papers I’ve ever seen; it really does not get any better than this. You’d have to go to another planet and hope that someone over there is working at this level.