Skip to Content

Chemical News

Here’s What’s Been Done Before

I enjoyed this ACS Med. Chem. Letters perspective on AI and machine learning in medicinal chemistry. It has several good points to make, and it brought up one that I haven’t gone into here before: if you’re mining the literature, you will get what the literature can tell you. At the very best, the high end of the scale (we’re not there yet) the software can tell you things that you didn’t yet realize that the literature was telling you, but all you’ll get is what was there in the first place.

And in some cases, that could lead to trouble. Consider retrosynthesis software:

One potential drawback of this machine learning and pattern matching methodology is that it has the propensity to become self-reinforcing in certain areas. Take, for example, the emerging platforms for reaction planning. . .in suggesting best routes, such systems also often prioritize those routes based on frequency of utilization, i.e., those most used in analogue generation, such as Pd-mediated couplings and amide formations. In doing so, it is not unreasonable to suggest that over time, the utilization of these reactions increases ever further, making them more likely to be suggested by, and therefore potentially reducing the desired creativity and power of the AI systems to help us avoid over-reliance on certain reaction types.

I could definitely see that happening, especially since this sort of software is going to be used for bang-it-out let’s-just-make-some applications, for the most part. We have a lot of those in med-chem. So in terms of “did you get the compound?”, things will work fine, but just being able to make the compound won’t advance the science of synthesis very much – but perhaps this is just an amplification of what already happens. We self-amplify by hand, because as it is now, when it’s time to make some molecule any old way we can make it, we chemists tend to use reactions and routes that we already know and have used before. To see this on a smaller scale, it’s like the way people do palladium couplings: there are better catalysts than the ones that most of us reach for, but since those old favorites do tend to deliver product, we just take what comes out and move on. I remember an email a few years ago from a colleague in the scale-up group, to the effect of “Stop using tetrakis all the time you bozos”, but I’m not sure it did much good.

New reactions do catch on, of course, but they especially do so when they provide some sort of transformation that hasn’t been easy to do before. That’s why redox photochemistry has made inroads – it allows for bond formations that aren’t otherwise accessible. But a great new Pd coupling catalyst will take longer to catch on, because of the inertia just described. This means that there are two particular groups of synthetic customers who will adopt new reactions more quickly: total synthesis people in academia, who need to do difficult transformations in the highest yield possible, and (to a lesser extent) process chemists in industry, who need to maximize yield, reproducibility, cost, and suitability for large scale work. They’ll jump on new reactions (catalytic ones especially) that look more scalable, but they’ll check them out thoroughly to make sure that they actually work the same way every time. Both of these cohorts are outside the “just get some compound, willya” demands of some other fields, and one wonders if that might make them less likely to use the earlier versions of the retrosynthesis programs.

But the piling-on aspect of machine learning could extend further. This is something that (some) people in the area have been thinking about, but those of us seeing ML move into our fields may not have internalized it. If you’ve been involved in handling large data sets in general, these concerns are (or should be) second nature to you, but I’m more worried about the eventual customers who haven’t been doing that. The challenge will be to keep such systems from becoming “conventional wisdom machines”. Sticking with the retrosynthesis example, perhaps there could be separate “exemplification” and “novelty” scores available if you want them. You could imagine setting it along a scale, generating routes that range from “bound to work, everyone’s done this” to “potentially a lot easier/shorter, if these papers are real”.

16 comments on “Here’s What’s Been Done Before”

  1. tlp says:

    Historical prevalence of Pd cross-coupling and amide couplings could also have been reinforced by the fact that med chemists were designing compounds with these reactions in mind. If one designs compounds in reaction-unbiased fashion, it could be less of an issue.

  2. Chemist turned Banker says:

    Reminds me of the recent AI efforts by Amazon in recruitment- they fed it ten years of applications and the conclusion of the algorithm was….hire men rather than women. Not quite what they wanted and it was quietly binned…

  3. Anonymous says:

    I’ve mentioned some of this before, but here goes. Hendrickson’s Syngen program was designed to generate EVERY retrosynthetic bond set. (I.e., an 18C steroid skeleton would be cut into all of the 9+9, every 10+8, … 17+1 pairs.) The linkages to the starting material catalog and literature could rank them. Or the user could view them all and pare things down manually. To me, the most interesting suggestions were the screwy, weird, unprecedented suggestions. And I doubt that management would be interested in developing anything that new and risky.

    Also, most AI synthesis programs I am aware of to not incorporate and recommend rearrangements. Maybe the occasional Claisen or Cope, but not a slew of others — including the unpredicted ones — designed to confound students on their exams and orals. Nevertheless, they can be quite useful.

    (more later)

  4. Wavefunction says:

    The self-reinforcing nature of reaction planning is of course not limited to AI, since as the paper from Walters et al. (J. Med. Chem., 2011, 54 (19), pp 6405–6416) showed, humans also prioritize things like coupling reactions because of their ease and robustness, and these results turn into a self-fulfilling prophecy: for instance, you think that kinase inhibitors made using coupling reactions work well, but that’s because other techniques simply haven’t been used, and you just happened to pick the few successful results that worked.

    What I find most valuable in these approaches is what you are suggesting lies at the high end of the scale; recognition of patterns that are not new per se, but that are too complicated and scattered through the literature for humans to easily spot.

  5. AVS-600 says:

    The big-picture version of this doesn’t really seem like that much of a problem in and of itself. We do lots of amide bond formations and Suzuki couplings because those are fairly robust, fairly general reactions. It would be sort of weird (and arguably a failure of the program) if a synthesis that used those reactions WASN’T suggested. If/when AI retrosynthesis takes off, the real challenge will be getting out of the mindset that we need to be designing all our compounds to contain biaryls and amides in the first place.

  6. Retro sinner says:

    I remember sending a pot of Pd118 and a preferred ligand to a discovery colleague in med chem as an alternative to tetrakis. He replaced 50% ‘catalyst’ loadings with 1-2%. He raved about it but couldn’t give it away to any of his neighbours doing similar work and with the same issues. I still don’t understand why.

  7. a degenerate chemist says:

    Would you do a 9-step synthesis that has a 90% chance of working, or a 3-step synthesis that has a 75% chance of working? Where would you draw a line?

    1. DrOcto says:

      3 steps, without hesitation.

      If it’s chemistry you’ve never done before, then 1 step per week is a ball park figure (one test, one repeat, one scaleup, and analysis).

      When I personally choose reactions it’s the ‘just get it made’ route. Exotic catalysts, extreme conditions, weird reactions, radical chemistry will all get binned straight away if there is an alternative option with shelf chemicals, even if the yield is lower.

    2. doc says:

      3 steps, no doubt.

      At a first glance, looking at the question as a question of risk of failure/loss of time, one might think as follows. Taking one step per week as a reasonable time, 21 days x 25% chance of failure overall = 5 days of hypothetical cost. 63 days x 10% = 6 days. On that basis, maybe it’s a wash. Except that’s not how chemists think, nor is that the way to look at he problem.

      The problem with that analysis (and the overall question) is that it’s just wrongly stated. It’s a series of chained events, conditional probabilities if you will. Most of us have some sad stories about that. A 90% chance of success over 9 steps implies a slightly less than 99% chance that each step will work. (0.99^9=.92), or at least work well enough to go on to the next step. A 75% chance of success overall implies that each step has a somewhat better than 90% chance of working well enough to go on (0.9^3=0.73). At the bench, 90% and 99% look pretty much the same, unless one has run the exact reaction before. With untried starting materials and conditions, though- not different.

      Three almost certain steps, vs nine certainties, but never having done any of them? Most, I suspect the vast majority, of experienced synthetikers would take the shorter path and bet we could bludgeon it into submission in four or five weeks. Or, at least, fail fast and then try the alternate route.

      The problem with trying to make this decision quantitatively is that I’m not good enough to predict success with anything better than maybe quartiles (probably won’t work, maybe, probably will work, certainly will work).

      There is still plenty of art to synthesis. Given that, quick and dirty beats long and likely every time.

  8. hn says:

    AI is only self-reinforcing if that is how it’s designed. One can set weights for different kinds of training data. There are generative AIs that look for patterns in data and try to suggest new models that fit those trends. They can be more “creative” by adding randomness, mixing in different types of data, inclusion of physical laws, or user input. We are just getting started.

    1. gwern says:

      Many of the techniques used for generating new chemicals to sample, like Bayesian optimization, inherently avoid the OP problem because they quantify the uncertainty of each possibility. If there are a lot of pathways in the literature using a specific reaction, it will be estimated accurately, and be of little value compared to trying something with only a few examples. Even greedy methods of tree search which one might apply to retrosynthesis will unavoidably cause some exploration, simply because they will be optimistic in thinking that some undersampled pathways are better than those pathways actually are (due to sampling or model error) and try to exploit them.

  9. hn says:

    I don’t think there are that many AI researchers working on reaction planning. That is probably not the best application of AI in chemistry/drug discovery right now. It’s also not as interesting to those of us who are not synthetic chemists (sorry!). Most of the AI chemistry people come from backgrounds in physical/computational chemistry.

    1. Some idiot says:

      Don’t be sorry! 🙂 But I am just curious as to _why_ it is not considered to be an interesting problem… Lack of perceived complexity, or something else? I say “perceived” because as a chemist it is my opinion that this is indeed a complex question… And I am really, honestly curious as to why this area is not considered interesting! I look forward to finding out more! 🙂

      1. LF Velez says:

        “Interesting” sometimes means “what I understand well enough to put into a program” or “what is easiest to get via my available pool of research subjects”. [See also: lab mice are a well-established model, so we keep using them..] Early AI research looked at activities like chess [finite set of rules] using expert/novice comparisons [the strategies used by experts were turned into algorithms for a computer program, and novice players would be coached using the same rules to see how things turned out]. Grad students in psychology and in English were sometimes the handiest research subjects/sources of information [and eventually, they would go off and found their own research agendas, sometimes with an AI focus]. Crunching chem literature might not be seen as an accessible data set because fewer chemists participated in such projects. [I am guessing, but based on my experiences at Carnegie Mellon in the 80s and 90s]

        1. Some idiot says:

          Many thanks! Yep, sounds/feels real… 🙂

  10. Somhairle MacCormick says:

    I guess the power of AI will be greater when huge research organisations use it on their own reaction data. Because the lit. is basically everything that has worked, whic is informative, but adding the info on what hasn’t worked creates a much more colourful picture. And AI can use all that info to greater effect.

    Although not all chemists were created equal and somethings just don’t work in one persons hands.

    Maybe the AI could suggest the chemist to do each reaction for it…….

Leave a Reply

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

Time limit is exhausted. Please reload CAPTCHA.