Here’s another paper for the automated med-chem files. A group at Merck (Boston) reports a combination of very small-scale automated synthesis with a screening assay in situ (no purification). You may be wondering how that works, or how it can possibly work, especially when you hear that the nanoscale reactions are transition-metal catalyzed. After all, we usually spend a good deal of effort trying to make sure that our assay candidates are cleaned up and (most especially) free of metal contamination.
The key is the type of protein assay. It’s not a functional readout, which is where the metals can give you all sorts of false positives, but rather just a binding assay, done by mass spec. That’s generally done in an is-anything-there binary mode, but in this paper, they’re actually titrating in the amount of protein to get a rough order of affinities. Still, one strike against that sort of detection has always been “But you’re only measuring binding, not real inhibition, etc.” That’s still true, but now you can always consider going the targeted protein degradation route, in which case all you need is some sort of binder. As mentioned here in another recent post, the same consideration applies to DNA-encoded libraries, which also just give you a binding affinity readout – the Merck authors don’t mention this aspect, but it’s a real consideration these days.
In this case, they did a test with 20 reactions each using amide formation, Suzuki couplings, and Buchwald-Hartwig C-N couplings (which certainly cover a lot of real-world medicinal chemistry, for better or worse). They used the small-scale reaction setup to screen conditions for the latter two, in a smaller-scale version of this sort of thing. In each case, they also synthesized the products on a 20mg scale and purified them by traditional means, in order to compare them with the nanoscale assay results. (These all produced compounds in known kinase inhibitor space, which I’m afraid is also pretty realistic.) And the correlation was strong: the affinity-mass-spec assay could indeed pick out the potent compounds (as it should) without giving false positives along the way (since you’re looking for the particular mass signature of the product associated with the enzyme.
With this confirmation in hand, the group then took one of the intermediates and coupled it in each individual well of a 384 plate after an automated reaction conditions screen. They got 345 products and rank-ordered them by titrating in CHK1 protein, with a mass spec readout each time, and identified three new potent binders. These reactions are done on about 0.05mg of material each, so you can plow through quite a few of them without plowing through a lot of starting material, and generate a lot of SAR data rather quickly. The reaction screening is a key part of the process – taking robust but generic coupling conditions across that same 384 set only gave 158 products (and thus missed several nanomolar inhibitors). Overall, to optimize the 384 set and assay them took 3114 reactions, consuming about 120mg of the starting material. You are not going to be able to do that by hand. I don’t care how much of a machine you are in the lab, you’re not more of a machine than a machine is.
As the authors note, machine-learning algorithms may eventually make it easier to navigate both reaction condition space and SAR space, but for now, we really have to attack these things empirically, and this (in general) seems like the way to do it. Small scale, highly automated, optimized, high-throughput: you’re not going to be able to sort things so thoroughly any other way. And since we often have to resort to brute force, we should be letting the machines take that on whenever possible, because brute force is exactly what they do the best. That was proven on the high-throughput screening end of the business many years ago, and it’s true on the compound synthesis end of things, too. Bring ’em on.