Why do our reactions fail? Why do our darn reactions fail? Every experimental chemist wonders this, because we all have set up reactions that we thought would work (why else would we run them, eh?) only to have them sit there and do nothing – or worse, do everything and turn the color of used lawnmower oil. Every chemical reaction is a tightrope walk. In fact, if you picture them traversing an energy surface it’s a literal tightrope walk, and you can see the narrow path that leads to the product, along with the various pits, slopes, and gulches that things can slide off into instead.
The only way to figure this sort of thing out is to run a lot of careful attempts under well-controlled conditions: scientia est experentia. In the drug industry, process research chemists do this because the stakes are high. People in academia who are trying to get crucial total synthesis steps to work are often in the same position, and (needless to say) so are the folks who are trying to invent new synthetic methods entirely. Most of the rest of us, though, will try a few variations while our thoughts turn to whether there are any different reagents or reactions that might be less complicated and just produce some product already. The reaction landscape is large and complex, and mapping it out in useful detail is actually rather painful.
Here’s a new paper that is trying another automated attack on all this. A group at Merck (with collaborators at Bruker) reports another miniaturized approach to reaction screening, this time using MALDI-TOF mass spec as the analytic end of the process. It would appear to be later work from the same initiative that produced this earlier work on reaction optimization, and for more industrial chemists trying this, see here for some recent work from Pfizer. That paper investigated Suzuki-type couplings, and this new one is going after Buchwald-Hartwig ones (the carbon-nitrogen palladium-catalyzed reaction, famously persnickety about catalysts and conditions). For a pain-in-the-rear reaction, though, it sure does get run a lot. It’s a very worthy place to set the machines loose.
The Merck group looked back through the company’s electronic lab notebooks, and found plenty of such coupling reactions. But they were run under so many different conditions, and with such irregular coverage of chemical space, that it was hard to extract anything useful out of the data set. So they made their own. But they did it differently than the Pfizer group (who used flow chemistry, which requires homogeneous reactions and thus often higher dilution conditions):
. . .Rather than accepting such constraints, we systematically engineered and validated new plate-based nanomole synthesis tools with a general ability to carry out a wide variety of typical synthetic reactions. We identified effective chemically compatible glass microplate reactors, fast 384-tip dosing of reagent solutions in volatile solvents, and designed aluminum sealing blocks that can retain volatile solvents on heating. In addition, we used LabRam resonant acoustic mixing to both agitate reactions and to create milky slurries of solid inorganic bases that can be added to reactions by parallel liquid handling. Finally, we created a nanomole scale photochemistry tool to enable reaction evaluation in the rapidly growing field of photoredox catalysis.
This was not the work of a few idle afternoons. (A lot of information on how this was done is available in the supplementary information file of the paper, by the way). The decision to use MALDI/TOF mass spec was motivated by the time it takes to do LC/MS work – the syringe has to come over, pull up some sample, inject it, get washed and move to the next sample while the first one is making its way down the column, etc. It’s true that you get more information from the LC trace, but the speedup in the solid-phase MALDI technique (where a laser beam blasts the sample off the surface and into the mass spec) is hard to ignore. The group did some optimization there, looking for conditions that would give reliable readouts across a range of substrates, and worked out a procedure where aliquots of each reaction were spotted into the wells of a MALDI analysis plate (with internal standard added – without that, you’re at sea) and the appropriate matrix compound (alpha-cyano 4-hydroxycinnamic acid) was added afterwards in solution, followed by evaporation. A 1536-well MALDI plate could be analyzed in about ten minutes, with automated data collection and analysis on the back end.
A first run of the system looked at four different coupling conditions (iridium/nickel and ruthenium/nickel photoredox, along copper-catalyzed and palladium-catalyzed conditions) to search for functional groups that could poison the reaction conditions (as in the work of the Glorius group on reaction robustness). These four reactions were run with well-performing reactions partners in the presence of 383 other polar molecules with a wide range of functional groups. As always with the automated chemistry work, one’s lip starts to quiver at the thought of doing this sort of thing the traditional way. That is just a lot of work.
What they found was that each of the reactions is subject to poisoning by specific functional groups. For example, the copper-catalyzed reactions are hammered by anything with a thiol group, which makes perfect sense, while the iridium reactions were not only intolerant to SH, but to acids, phenols, oximes, and nitro groups as well. Which is annoying. And when such functional groups appeared in larger, more complex additives, they were almost always just as bad. To add to the fun, there were other large molecules that contained several by-themselves-innocuous functional groups that together managed to poison the reaction as a team. Yep, this is metal-catalyzed coupling chemistry all right. (It’s worth noting that the copper reaction actually had the most permissive profile, thiols aside).
The group then turned to variations in the coupling partners, with a list of 192 aryl halides and 192 secondary amines. If you did all of those, you’d be looking at 36,864 possible products, and across the four reaction types that gives you a cool 147,456 experiments to set up, which was enough even to give these guys pause. They tried a sampling of that space by setting up each bromide with the most simple amine in the set, and each amine with the most simple bromide – that takes you back down to 1536 reactions, a 1% scattering fairly evenly across the possibilities.
This is a much harder row to hoe. Adding an internal standard is necessary, but a lot trickier to work out across so many reactions and structural types. The group had run that previous experiment in traditional LC/MS and found a good correlation in the results (once you normalized to the internal standard) but these 1536 experiments had a lot more scatter in them. Backing down to a pass/fail on the MALDI data (with an arbitrary cutoff) gave a much better correlation with the more comprehensive LC/MS data, so you can at least search reaction space in a “what’s bad?” mode, with analysis of the properties of each compound for steric hindrance, number of nitrogens, cLogP, hydrogen bond donors, etc. Rather particular structural features turn out to be reproducibly important for success and failure in the Cu-catalyzed versus the Pd-catalyzed reactions, and several of these cannot predicted by our current understanding of the reactions.
The paper refers to the “dark matter” of chemical reactivity, and there’s some of it for you: things that you wouldn’t have known would affect the reaction, and do so for reasons that aren’t yet clear. These coupling reactions are particularly tricky to work out, but there are plenty of other examples in all fields of synthetic chemistry. (Going back to my grad school work, I can give examples of small changes across carbohydrate structures that totally change their reactivity, in ways that I often discovered in the most painful ways possible!) Every class of compounds and every class of reactions has things like this, results that just make you throw up your hands. If there’s a way to start mapping out the vast “Just One of Those Things” spaces, we’ll learn a lot more about the chemistry we’re supposed to already understand by doing it.