At a previous company some years back, I was interested in getting a “covalent fragment” collection going, and did to a small extent. It got screened against some antibacterial targets, but never became all that popular. That was partly because fragment-based screening was a younger field, and combining it with covalent drug discovery (also just beginning to pick up renewed interest) seemed a step too far for many people. But there were practical difficulties as well: for starters, there were only a limited number of new targets coming along per year, and not all of them even needed a screening mode like that one. As a practical matter, we didn’t always have the mass spec capability available to characterize the behavior of the fragment collection, either (and since you’re covalently adding mass to your proteins of interest, that can be an important step to check). So screening would sometimes be done with a functional readout, which meant that you’d only be seeing the fragments that did a reasonable job of shutting down the target’s binding site.
And how often does that happen? That’s the key question, especially as you move down to fragment chemical space. You can picture some little acrylate or the like just sort of banging into nucleophiles all over a protein, and if it’s too hot an electrophile it might just react with several of them, particularly surface residues where it has a variety of geometries to choose from. Alternatively, if it’s a small fragment without much other functionality, it might find a lot of low-affinity binding sites (via the non-warhead part of its structure) without necessarily hitting one that is really favorable (because it just doesn’t have a distinct enough structure to make some particular site a really stand-out binding fit). That makes activity-based screening of such a collection tricky (I’ve written about some of these issues here before).
So breaking it out into quadrants, this means that if you have a small fragment with a weak electrophile, you may never see it hit at all – its binding affinities will rarely if ever be high enough to compensate for its intrinsic low reactivity. A small fragment with a red-hot electrophile, on the other hand, might be more likely to just wander around labeling all kinds of surface stuff, which probably isn’t too useful either. Meanwhile, a larger and more complex ligand with such a reactive electrophile will probably get hits, although perhaps not all of them will be taking advantage of all of the compound’s structure (since it might still go ahead and react promiscuously). And a larger, more complex structure with a weak electrophile will hardly hit at all unless it gets a really good binding event that puts its warhead right up against an appropriate protein residue. The odds of that are not high, although when it does happen it’s a very high signal/noise event that’s worth paying attention to.
All these thought exercises would push a person towards trying to balance chemical complexity with warhead reactivity, and towards avoiding extremely reactive electrophiles in general. Here, though, is a look at a collection of about 1000 electrophile against ten reference proteins (commentary here at Practical Fragments). Some of their functional groups (such as chloroacetamide) are definitely towards the hot end of the scale, and you can see that even with this functional group you can get discrimination – in fact, promiscuity of protein labeling didn’t correlate all that well with intrinsic reactivity. So that mental model I was just laying out needs some adjustment when confronted with the real world, at least on the intrinsic-reactivity end of it. The compound-complexity part, though, I think holds up (see below).
For example, here’s another look at the topic – like that previous link, it has a lot of useful information on intrinsic reactivity of various warhead groups towards different amino acid residues, via an NMR assay. Their assays are done in isolation, so binding-pocket effects are removed, and the order of reactivity (for the groups that have been studied before) makes sense compared to other work. The paper emphasizes that these reactivity rankings can change quite a bit with pH, and that is of course something that’s going to change a lot inside a binding pocket itself (insofar as it makes sense to talk about a bulk quantity like pH in such an environment, of course – but it’s for sure that things are going to be deprotonated to greater and lesser extents in there).
But one thing to note in that paper is that one of their compounds is a beta-lactam penicillin derivative, and it reacts with plain serine at a rate too low to be measured. Such compounds, of course, actually exert their biological effects by labeling a particular active-site serine in a bacterial transpeptidase enzyme. But that’s a very different situation that just labeling serine by itself out in an assay tube, clearly. I’ve seen that effect personally as well: covalent compounds (of reasonable complexity) that will indeed bind to and inhibit some particular target, while not performing well in simpler reactivity screens, and whose reactive groups didn’t show much activity against that target when reduced to fragment size, either. That’s even more the case with the weaker electrophiles: you can find plenty of examples in the literature (this is a classic one) of not-very-reactive compounds that are just waiting for their moment in one particular binding site.
Now, finding such things is not so easy, which is why you usually see the problem approached from the other way around. That’s the case with this paper from earlier this year, where the authors take a fairly promiscuous acrylate-containing kinase inhibitor and use it to survey targetable kinases (and their various targetable cysteine residues) in that whole enzyme class. The expectation is that SAR development could then lead you towards selective chemical matter, now that you know that the covalent labeling is possible to start with.
And looping back to where I started this post, that’s the hope with covalent fragment collections in general, that you’ll come across something that gives you a starting point for further development. There’s no promise (or at least there shouldn’t be!) that you’re doing some sort of comprehensive survey of what can react against a particular protein or any particular amino acid side chain on it – you’re just looking for a handhold so you can get climbing, and the fragment-screening route is a useful way to do that.
If you run such a fragment screen across some protein and don’t really see anything, you definitely shouldn’t conclude that it’s not targetable by covalent mechanisms. You can’t make that leap at all. Your protein might just be waiting for the right dance partner to come along. I think that’s where covalent fragments and “regular” fragments part company, actually: if you do a fragment screen against some target with a large, diverse set of compounds in an assay that you know is performing, and get nothing. . .well, I think you can conclude that small-molecule binding sites on that protein target are going to be mighty hard to find and exploit. You’ve explored a lot of “primitive” simple binding modes, and your target doesn’t want any of them with decent affinity (see yesterday’s post, though, for pushing into “indecent” affinity territory!)
But if you do a covalent fragment screen and come up empty, I think that it’s not so definitive. Forming a covalent bond, it seems to me, is a more stringent process than plain binding, because of the angles and distances involved in nucleophile/electrophile bond forming reactions. It’s as if a regular noncovalent fragment collection could only read out if it formed hydrogen bonds, with all their picky spatial requirements, instead of being able to access all the other modes of compound interaction to show affinity. Covalent collections are just intrinsically more finicky?