Skip to Content

Brute Force: Bring On the Machines!

Well, here I was the other day going on about automated chemistry when this paper was waiting in my RSS feed. It’s from a group at Pfizer, and they’re using an automated microscale flow chemistry rig for reaction optimization. Inspired by this work from Merck, which demonstrated evaluation of 1536 reactions in a plate-based system, this new paper moves from microscale batch mode to continuous flow.

The key objectives were to develop a fully automated system for HTE screening with flow chemistry technology that would (i) integrate inline high-resolution LC-MS analysis for real-time reaction monitoring; (ii) use diverse volatile and nonvolatile solvents; (iii) use ~0.05 mg of substrate per reaction to enable broad parameter space exploration with minimal material consumption; (iv) enable the preparation and analysis of up to 1500 reaction segments in a 24-hour period; (v) establish the capacity of the platform to directly scale up preferred conditions via multiple injections to produce 10- to 100-mg quantities of a specific compound; and (vi) show translation of nano-HTE conditions to both larger-scale batch and flow synthesis.

Objective met, it looks like. They’ve worked up an ingenious system to assemble five-component mixtures (two reactants, a metal catalyst, a ligand for it, and a base) for evaluating Suzuki-style coupling conditions – the five are combined in concentrated form and then injected into a carrier solvent (which can be varied as well), and checked at different temperatures. Injecting in a bolus like this also gives you the ability to send the next one in after a suitable delay, instead of waiting for the first one to get all through the tubing. Five or six dimensions gives you a lot of potential variations, which (unfortunately!) is just what you need for metal-catalyzed couplings. On the plus side, it’s long been a belief of mine that any such coupling can be optimized to a high yield if you’re just willing to spend enough of your life doing so. This setup takes that idea and runs with it.

It’s not a cheap assembly – there are two Agilent systems waiting at the far end of the flow apparatus, to handle the reaction bolus injections that are coming out every 45 seconds. The whole idea took a lot of careful validation as well, to figure out what concentrations to use for the injection stocks, what solvents to have them in, their behavior on hitting the carrier solvent and being diluted by it, the extent to which the injections would spread out as they came down the reaction tubing, the efficiency and accuracy of the LC/MS analysis at the far end, and so on. Without all this groundwork, though, it would have been easy to use the fancy robo-rig to generate pile upon pile of crappy, hard-to-reproduce data, which is a temptation that has to be avoided. Measure fifty-three times, cut once, as the old saying goes.

They ended up screening a total of 5760 reactions, which evaluated Pd couplings across eleven ligands and seven bases (organic and inorganic), all in four different solvents, for most combinations of a matrix of four electrophile and four nucleophile partners. As mentioned above, the machine ran 1500 reactions per 24-hour period, on a 0.4 micromole scale per reaction. A heat map of the results are shown, although don’t feel as if you have to work your way through all of it. You can see immediately, though, that there are some combinations that have a much better success rate than others. Xantphos, for example, seems to consistently underperform the “No ligand” control category for these transformations, whereas good ol’ triphenylphosphine is your friend. The 6-chloroquinoline is a tough customer. The trifluoroborate partner (2c) is just not reactive enough under these conditions, unless you use methanol (probably because it has to hydrolyze before it reacts?).

It’s possible to plot these conditions out in several different ways, naturally – the authors show, for example, that if you’re looking for conditions that always seem to deliver product, no matter the combination, then X-Phos or S-Phos in acetonitrile is the way to go. So if you’re going to turn around and set up a big parallel synthesis run, you’ll be very glad to have scouted all these things to improve your success rate. The paper shows that they could use the exact same conditions (but just injecting the same combination over and over) to provide 50 to 100mg of a specific product within 75 minutes, and translating the same conditions to a larger traditional batch reaction worked just as expected. (They tried a few winning conditions and a few losing ones in batch, actually, and the trend held up every time, which is encouraging).

Interestingly, the group then turned to Pfizer’s traditional parallel synthesis efforts, taking a particularly challenging aryl bromide/pinacolborane combination and optimizing it (the standard parallel conditions for such couplings had given no product at all). Running 576 different combinations across 8 hours showed that there were painfully sharp cutoffs in the reaction landscape. Only two catalysts seemed to give any product at all: CataCXium A was unusually effective, particularly in in THF/water, AmPhos showed some reactivity, but the rest of the catalysts (including the XPhos and SPhos combinations in the main screening run) were completely useless.

That is indeed metal-catalyzed coupling as I have experienced it, and until we get an utterly thorough understanding of the reaction details – don’t hold your breath – the only way to deal with this situation is this sort of brute force experimentation. And this is the best brute-force technique I have yet seen. Believe me, setting up five hundred and seventy-six Suzuki couplings in a row is not fit work for a human being, not when there’s a machine that can do it instead. Great stuff.

25 comments on “Brute Force: Bring On the Machines!”

  1. Hap says:

    I’d dread being the one to maintain the thing (the combination of a robot and a glove box makes me think it’s going to be no fun at all), but I guess I’d dread doing that many reactions even more.

  2. Mad Chemist says:

    Cool stuff. In my (somewhat limited) experience, acetonitrile seems to be the best solvent for most organometallic reactions/catalysis, so I’m not surprised that it did the best here.

  3. HGMoot says:

    This is the kind of paper I like reading… Excellent work, any questions about prep time and money notwithstanding. BTW, the “good old” PPh3 also did a decent job.

    1. anon says:

      hahaha exactly!
      wonder what’s the best, let’s have a look, oh it’s a boronic acid and tetrakis what a revelation.

      1. Derek Lowe says:

        Not with that second example. Tetrakis gives you no yield at all in that case.

      2. JHMoot says:

        you might be right… but was it really a tetrakis or a different Pd-PPh3 species? What do I know…

        1. Derek Lowe says:

          We throws ’em in the flask and we takes what we gets. . .

          1. JHMoot says:

            Yeah, wow, that could be… But I mean from the ligand to Pd ratio given in Fig. 2…

  4. CMCguy says:

    I guess with the preponderance of biaryls out there made by med chemists and combichem using Suzuki reaction (and similar coupling) over the past few decades does support possible value in such work for the optimization although groan when I see such candidates suspecting formulation hurdles. I do wonder if really that practical in real world Process Chemistry where resources typically slim and timelines always short so that faced with meeting material demands on a schedule greatly limits what can be done so must project stumble forward focusing on the largest problems (which probably does not involve aromatic coupling step unless its trouble obtaining the particular partners). IMO most process chemists have limited knowledge and rare training where performing multiple rounds of well thought out DOEs seems would have a higher cost benefit ratio then doing machine enabled Brute force to application for scale-up manufacturing.

    1. HGMoot says:

      ” IMO most process chemists have limited knowledge and rare training…” hm… yet just one study like the one discussed here provides lots and lots of valuable data, that can be simply inferred and applied, with a large benefit/cost ratio. Also, the Suzuki is not only aromatic-to-aromatic, you know…

  5. kjk says:

    Having “diligent and dumb” robots seems to be a much more short-med term viable strategy than fancy AI that crawls the literature. Transcriptic is doing a similar thing for mol-cell bio. But these robots can’t design experiments, and there will always be an oddball case here and there where we need to get our gloves and coats. This will give us a nice balance instead of a tedious slog.

    1. Peter S. Shenkin says:

      ‘Having “diligent and dumb” robots seems to be a much more short-med term viable strategy than fancy AI that crawls the literature.’

      I agree, but fancy AI could be put to much better use scraping the electronic lab notebooks of large pharmaceutical labs.

    2. tangent says:

      Designing “intuitive leap” experiments is a long way off, but directing basic “what happens with this combination” experiments — chosen out of a well-defined parameter space — doesn’t seem so hard. It’s a matter of estimating which are likely to provide progress towards a goal (like a synthesis target). It can be given to a computer as a reinforcement learning problem.

  6. Charlie Kilian says:

    “I’d dread being the one to maintain the thing”

    This reminds me so much of world class astronomical observatories. They have staff dedicated to running the machines. Astronomers submit proposals, the observatory picks the proposals to be run, and each astronomer group sends representatives to help out during the course of their observation run. There’s only so much observation time, of course, and every observatory has differences that make it very good at some things and not so good at others. Naturally, astronomers try to tailor their proposals to the strengths of the telescope.

    It seems to me you could do the same thing here. Build a world class synthesis machine good at making a certain set of reactions. Have a set of technicians on hand to help run the reactions. You submit proposals, and if your proposal is chosen, you get time on the synthesis machine. The technicians are on hand to set things up, and if any ambiguities arise or decisions need to be made, you or someone in your group is there to answer questions and make decisions and generally help out. The syntheses are run, and you go home with the data, which is exclusive to you for a year (or whatever the appropriate embargo period would be). After that, it is released to the public, available for researchers everywhere.

    I’d have to imagine this could be done, and probably for a cost similar to an astronomical observatory; something like LIGO, for example. I imagine the biggest thing preventing it from coming a reality is that the work isn’t as sexy as discovering gravity waves.

    Also, data collection from this kind of “diligent and dumb” robotic approach would be on the kind of level you’d need to make the AI-that-crawls-the-literature a more successful venture. It wouldn’t crawl the literature; it’d crawl the data collected in very well established conditions.

    1. metaphysician says:

      So what your saying is, we need to give our AI’s robotic workers and hands, so they can do stuff in the material world? *eg*

    2. tt says:

      It would make much more sense for a centralized, cloud based labs to have machines like this and run them as on demand, similar to your suggestions. It’s a bit silly for each pharma company to build there own bespoke system. Check out transcript labs and emerald cloud labs as a perfect example of this in biology…now if only they would do this for chemistry.

    3. Ben Deadman says:

      You may be interested in a new research facility at Imperial College London and the Dial-a-Molecule network. The Centre for Rapid Online Analysis of Reactions (ROAR) will provide UK researchers with access to a suite of high-throughput robotic reactors, reaction kinetics investigation platforms and flow chemistry reactors to enable data-rich research in chemical synthesis.

      The access model suggested by Charlie is very close to what we are aiming to do. The facility will have dedicated scientific and technical staff to run it, and we hope to be opening calls for proposals later this year. If anyone here is interested then please check out our website and/or contact us roar[at] to find out more.

  7. Chrispy says:

    I saw a presentation once from Pfizer about doing antibody discovery through well-by-well brute force screening of random antibodies. I couldn’t help but thinking that the sampled space was tiny compared to the possible space, and that the approach itself exhibited the kind of hubris that has made Pfizer the clumsy behemoth that it is today.

    1. yf says:

      I suspect those are cell based assays. Cell manipulations are generally too fragile to be automated.

  8. winampdfx says:

    What a terrible waste of resources! Two bench chemists would get a good yield on multi-gram scale in 4-6 weeks.

    1. Derek Lowe says:

      On all the nucleophile/electrophile combinations?

      1. winampdfx says:

        Most of the information about all possible nucleophile/electrophile/base/solvent combinations is useless. It can potentially be interesting for a small part of scientific community, but that is another story.

        The multinational Pfizer’s team conducted 5760 reactions at a rate of >1500 reactions per 24 hours, wasted a huge amount of solvents and reagents, and at the end managed to synthesize two compounds (3 and 6) on 500 mg scale!

        By the way, they did not test potassium carbonate – a base that typically works very well in Suzuki coupling reactions.

  9. kill me says:

    …and they chose the easiest of lot…Suzuki reaction!! Make a Buchwald amination bro and see how blue you get!

  10. b says:

    Don’t have access to the paper. Curious – how are insoluble solids handled and injected into the system? I understand making concentrated catalyst & substrate solutions for injection, but what about those bases?

  11. Nick K says:

    While I applaud the ingenuity and perseverance of the authors of this work, I feel that an opportunity was missed to discover exactly WHY this reaction was so sensitive to conditions. What else is going on here?

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.