Skip to main content

AI and Machine Learning

Lab! Of! The! Future!

This is a good article at C&E News on the “lab of the future”, and I’ll go ahead and make the standard comment that this has been the lab of the future for quite a while now. The idea is to have mechanical automation and experiment-evaluating software fitting together to “close the loop” and make experimentation more of a turn-it-loose-and-watch-it-run process. It’s a worthy goal, and people have been trying for many years to realize it.

But there are a lot of ways for this vision to go off the rails. On the mechanical end, chemical experimentation is notoriously hard to automate because it can involve such a range of physical actions. Powders, liquids, gases (all of which are measured by different means), orders of addition, inert atmospheres, widely varied temperatures, stirring techniques, and a huge variety of solvents and reagents no matter what. Some of those evaporate on you if you don’t move quickly, some of them schlork up water from the air and turn from powders into sticky goo, some of them precipitate from solution and clog up your needles and transfer lines, and some of them have a tendency to burst into flame. So any attempt at wide-ranging automation is going to need some troubleshooting.

On the software end, the optimization of reaction conditions and of product activity can still devolve into a semi-science. Both of these have notorious cliffs and discontinuities that will only become apparent when you stumble over them, and these things tend to be poorly handled by computational models. Sometimes a principal-components-analysis approach like DoE can work well, but other times it struggles, and there is often no way to know which of those situations you’re getting into. But to be honest, most of us would settle for an automated “make me a whole bunch of stuff roughly like this” or “try a whole bunch of catalysts and solvents and so on around this transformation”, with human brainpower to be applied afterwards. That’s a point the article stresses:

These labs aren’t perfect yet, and much work and convincing are needed before research managers in the drug and chemical industries embrace them. Observations vary on where the bottlenecks lie in the clocklike loops—some point to the data; others, to the robots. But the innovators are unanimous regarding the importance of the third element in the loop: the human researcher.

“This idea of the clockwork laboratory is, in the long term, not the strongest approach,” Cooper says of the notion of a self-sufficient research machine. “The strongest approach is to have the clockwork laboratory with a very permeable interface so that the human knowledge can be captured as well.”

Exactly. And that’s why I don’t run around shouting for people to man the barricades because the robots are coming. I want the damn robots to hurry up and get here already! We’re a long, long way from hitting a few buttons on the screen to have the Synthotronic 5000 make us a new antibiotic or a new transparent conducting polymer. What we can use help with, and what the current systems are finally able to start to do, is to say things like “Hey, take this library of 100 amines, couple them onto this intermediate one by one, and ring this bell when you have the set of purified samples for testing” or “Hey, vary these crystallization conditions across this big matrix of temperature/time/additives and get back to me with what happens”. In short, the kinds of things that you would turn over to some idealized imaginary group of highly competent and technically adept lab assistants who work constantly and dive right into any task you give them.

There is an awful lot of this kind of work to be done across all fields of chemistry, and all of us have put in our time doing it, too. And if some of it sounds tedious, well, that’s because it darn well can be, and some tireless mechanical help would be much appreciated. I’ve written several times (links therein) about this sort of thing over the years. And yes, I realize that we already have more of that tireless mechanical help than we used to, what with autosamplers and all, but allow me to propose a corollary to Parkinson’s Law: work expands to fill the automation available.

I have a presentation that I give about automation and AI, and one of its main points is that it will actually give chemists more to do, not less. But that work is going to migrate to higher levels as what we call “grunt work” is redefined by machines. Trivial (but real) examples from my own career: in grad school, I cut my own fractions as I ran hand-packed silica gel chromatography columns, using everything from 1mL vials (or smaller) up to 250 mL Erlenmeyers. In fact, folks in my lab used to wire up various sized vials into little 5×5 racks that could then be rinsed and re-used, and which eliminated those nasty elbow-scythe and roll-a-strike-at-the-bowling-alley incidents on the benchtop. I had two or three of the ones made out of 20mL scintillation vials, and boy did they get some use. Making those, collecting fractions in them, and rinsing them out afterwards was considered (by both me and my boss!) a perfectly good use of my time, but now such things are automated, and thank goodness.

Even in my first job in the drug industry, the NMR machines had no autosamplers on them yet. There was a signup list by each NMR, with names and office phone numbers, and you went down there with your tube(s) when you got the call and used up your 15-minute time slot however you wished. Then you called the next person on the list – it was generally someone that I knew well, so my standard message was “Get your worthless butt down to the NMR, you lazy slug” or something like that. Overnight set of samples? Sure, if you wanted to sit there in the NMR room and run them overnight. One by one.

Eventually, much of the work we’re doing now will seem as primitive and pointless as that kind of thing, and I’m honestly OK with that. I can get all nostalgic for those days, as long as I’m not actually having to sit there packing columns, waiting to retrieve my NMR samples, and standing next to the prep HPLC to make sure that it actually works. There are better and more interesting things to do with our time!

27 comments on “Lab! Of! The! Future!”

  1. David E. Young, MD says:

    I think that the idea of organized arrays is fascinating. But as a physician with just a modest memory of chemistry, I wonder if the chemistry that applied to most of the array will apply to all of the array. If you have a human cell culture in 5,000 microcells and your automated machine puts a 100 micromolar solution of one of 5,000 <500-dalton chemicals in each of the cells, will the chemical behave in the same way in each cell? What if one is too volatile? What if one degrades at room temp? What if one forms pairs with a brother molecule? And does sorting this out take lots of time? My examples may well be off, but you get the idea. My admiration for those who can do these arrays and be confident with the results.

    1. Dr. Victor Frankenstein says:

      The simple answer is that it does not matter that much in a high-throughput screen.

      Sure, some compounds will be unstable, but if you screen a sufficiently large number of compounds there will be something that comes through in the end.
      The unstable compounds are probably not the ones you want as lead compounds anyway.

  2. Jason says:

    These stories recycle the same themes every year. They cite the same professors from MIT and one from University of Toronto (who used to be at Harvard), because those are the ones all the journalists get hits on in their story lead feed.

    However, go to Pittcon, SLAS, AIChE and any number of other industry conferences and you will see all the major pharma companies (GSK, Pfizer, Vertex, Lilly, Novartis, and a good number of Boston-Cambridge companies) sharing much more practical presentations. You will also see a different group of professors, and a rather diverse group at that, presenting on a lot more realistic stuff. My advice is to not read these stories but go to the conferences themselves. For a grad student stuck in the lab, this is what is accessible but sadly these stories do not show the real picture.

    1. Anon says:

      It’s the age of marketing and hype. Some people do it really well.

  3. Ross Presser says:

    I have nothing else to say but that I really like that you used the phrase “schlork up water from the air.”

    1. tangent says:

      +1. As a professional professional I do much schlorking myself.

  4. anon the II says:

    I just read the article in C&ENews. About 2 paragraphs in, I got out my bullshit bingo card. By the time I was finished, I had won down, across, and diagonally, multiple times, with no need for the free space in the middle.

    1. Rhenium says:

      OK, now I’m really curious to know what is on your Bingo card!
      I stopped re-newing my ACS membership some time ago so alas I don’t have access.

  5. Peter S. Shenkin says:

    “The strongest approach is to have the clockwork laboratory with a very permeable interface so that the human knowledge can be captured as well.”

    True, but the same problem plagues traditional hands-on organic synthesis. If a step in the procedure fails to achieve its anticipated yield, often a set of experiments is devised which might involve altering solvents or reaction conditions. Even when success is achieved, my sense is that very little time is put into understanding why that set of conditions was successful when none other was. There is just a shrugging of shoulders and movement onto the next step of the synthesis.

    I can’t help but believe that a much better job could be done of trying to “capture human knowledge”, though “understanding” would be a better word. It might be that the poor grad student has to get through the synthesis to get their doctorate, and his advisor needs him to succeed so he can get his next grant. Still, I can’t help but think that additional experiments to test hypotheses about why this set of conditions worked but others didn’t would move the field forward the area of “human understanding” and that that understanding would have practical benefits as well.

    In fact, if robotic synthesis is successful, perhaps even greater understanding could be achieved than is now the case, by performing the additional experiments in an automated fashion in the background while the poor GS goes on to the next step in the synthesis.

    Caveat: IANAOC.

    1. Process Chemist says:

      Understanding is nice, but not always necessary. Due to time constraints I might be totally fine with getting reaction conditions from some kind of automated black box and then my job is just to figure out how to make the chemistry reliable on commercial scale. If I have time to look at all the data later to obtain some deeper understanding and publish something that’s nice, but it’s not the end goal in my case.

      1. Peter S. Shenkin says:

        Yes, obviously, that is what I was referring to. For the *current* practical purpose, understanding might be necessary, but for the long-term advancement of the field, gaining that understanding would be highly useful; one would hope, by providing greater predictability.

        It’s a classic conflict between achieving short- and long-term goals, and short-term usually wins. For obvious reasons.

        However, if the type of robotic synthesis that Derek describes is indeed short on providing understanding, my main point is that the Edisonian procedures that are invoked fairly regularly in manual synthesis also is short in going the extra distance to provide human understanding. So I don’t see that the automated procedures as any worse in this particular regard.

        But I cannot help but think that somewhere, there’s a place for investigations of the type I described. It might be that the automation under discussion would actually help run the grueling extra experiments that might provide human understanding without the chemist at the bench having to go through the exhausting experimental procedures manually.

        1. Peter S. Shenkin says:

          Whoops. I meant:

          “For the *current* practical purpose, understanding might NOT be necessary…”

        2. Process Chemist says:

          I think you’re exactly right the machines are currently best utilized doing the grunt work of running experiments and generating data. At least, they will become extremely useful once material handling and auto sampling are fully fleshed out. It’s hard for me to imagine a future where humans aren’t the ones asking the relevant questions and coming up with hypotheses to be tested, but I’m sure that scenario is closer than anyone can imagine.

    2. anon says:

      Understanding is nice, but it’s not what I’m getting paid to do. I get paid to deliver a certain amount by a certain date. If there were a year-end bonus that depended on understanding, then that’s a different story.

  6. Bob Seevers says:

    I am from the same grad school generation as Derek. What he did not mention is the early automated fraction collectors for those hand-packed silica columns often chose, for reasons of their own, not to rotate to bring the next empty collection vessel under the column. Thus, the all-nighter babysitting the chromatography column to ensure that I did not come into the lab the next morning to find a single collection vessel overflowing onto the floor with the rest of them sitting empty because the gremlins got to the fraction collector.

  7. Derek Freyberg says:

    @Anon the II:
    As an ex-chemist, my attention went also to: “IBM went on to launch Watson as a commercial product; its first application was in decision management for lung cancer treatment at Memorial Sloan Kettering Cancer Center. The firm also customized Watson for industrial research markets, including chemicals, in which companies including the big German firm Evonik Industries have deployed it.”
    As I recall, Watson crashed horribly in lung cancer management: see an IEE Spectrum article “How IBM Watson Overpromised and Underdelivered on AI Health Care”,

  8. sgcox says:

    “There are better and more interesting things to do with our time!”
    Indeed, I am too busy now arguing with antivaxxers here while robots run assays in my lab.
    What can possibly go wrong ?

    1. John Wayne says:


  9. Some idiot says:

    I like your corollary:

    Parkinson’s Law: work expands to fill the automation available

    although I would suggest the slight variation: work expands to fill the useful/appropriate automation available.

    As a process chemist who is very open-minded (but sceptical, in a good way, and like testing and trying), I have seen quite a variety of different automation possibilities, and yes, they all tend to fall into the general pattern of “some are good for some purposes, some for others, and very many are purchased for the wrong application…”

    But I like the corollary, because it emphasises that my experience is that introducing automation does not lead to people losing jobs, but rather increases the productivity of the people around it (assuming that the automation equipment purchased is fit for purpose).

    In my area (process chemistry), the needs of automation vary (in my opinion) very significantly from early route scouting to final optimisation or QRM experiments. My experience is that (generally) in the early stages pretty simple, semi-manual is most appropriate, but automation in analysis is extremely important (auto samplers in HPLCs changed my life, as did those on NMRs!!!).

  10. Reverend Lasher R-127 says:

    “This crusading spirit of the managers and engineers, the idea of designing and manufacturing and distributing being sort of a holy war: all that folklore was cooked up by public relations and advertising men hired by managers and engineers to make big business popular in the old days, which it certainly wasn’t in the beginning. Now, the engineers and managers believe with all their hearts the glorious things their forebears hired people to say about them. Yesterday’s snow job becomes today’s sermon…prophecy’s a thankless business, and history has a way of showing us what, in retrospect, are very logical solutions to awful messes.”

  11. exGlaxoid says:

    When I see a robot run a Grignard and then do a 1,4 addition to an eneone by the constant addition of copper iodide, I will be truely impressed. Until then, my favorite uses are very simple machines that do only one or two things really well, Gilson 215 and similar liquid handlers, vial weighing machine with barcoded vials, automated HPLC or ISCO type chromatography machines, and other simple workstations are great. Simple vial/tube rack systems and multiwell plates are also great ways to do things in parallel, with out without automation.

    I have seen quite a few very complex automated synthesis machines, from $100,000 to milions in cost, and most were a complete waste of time and money. A few had some clear uses, but most were mostly just way to complex, and had leaks, clogs, glitches, or took a year to program.

    1. anon says:

      A first-year graduate student or an intern is a lot cheaper than a robot.

    2. Anonymous says:

      Holy cow, are Gilson 215s still used? That’s awesome. I bought me first one in … 1996? Loved those little guys. How about Foxy fraction collectors?

  12. kjk says:

    My hope is that robots will get each *elementary step* reliable. One such step would be “pour x mg of said powder from container into rotovap flask, x is 10-100 (for x100 that would be a different step).” There are dozens of such steps, but each one doesn’t have all that many permutations.

    Each step will have options such as “how much do we tap the weight boat to remove the powder from it?”. Robots aren’t blind or deaf, so we can include feedback such as “if bubbles appear, slow down pipette pressure”. In fact, the robot may sense things we can’t such as the amount of static charge on powder reagents (which can make it fly everywhere). Defaults are provided, but tweaking is necessary if the sample has unusual physical properties or is time-sensitive, etc.

    These steps are further combined into successively larger protocols, with each level being a python function that calls lower down functions, again with defaults that occasionally are overridden.

    Doing all this is hard work, but hard in a much less tedious way. It requires much more clairvoyance to write flexible code that can handle common situations and alert the human to the rare ones. It requires *knowing* how deep into the nuts and bolts of what is going on you need to be rather than “being there with every datapoint”.

  13. Project Osprey says:

    I feel this is all directed very much towards pharma or other ‘blue chip’ groups. In the last 10 years I’ve worked in R&D at chemical companies which couldn’t afford their own NMR machines and where reaction tracking was still by flame ionisation GC. Even where I am now we still pack our own silica columns (on the odd occasion when nothing else works). I feel that for most chemist this sort of automation is decades off.

  14. Ya’ll want a preview of the *biomolecular* version of the Lab! Of! The! Future!???? Then hurry on over to and All! Will! Be! Revealed!!!!!!!

  15. Melodie says:

    Derek, thanks for highlighting the troubleshooting necessary to develop high-quality chemistry automation protocols. I was trained as a process chemist prior to transitioning into lab automation and found that scale-up principals also apply to to scale-down, parallelization and automation. If care is put into their development, automated systems can generate excellent quality data. We just need to stay vigilant. Closed loops are becoming more and more practically feasible, thanks to advances that address some of the challenges you bring up. However, as you highlight so eloquently, there is nothing wrong with designing human-in-the-loop systems. In fact, this is likely the most practical and effective arrangement.

Leave a Reply

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

Time limit is exhausted. Please reload CAPTCHA.

This site uses Akismet to reduce spam. Learn how your comment data is processed.