Skip to main content
Menu

Chemical News

Predicting – Or Not Predicting – New Materials

We chemists would love to be able to do just a tiny bit less chemistry now and then and just let models and simulations tell us what would happen instead. Only every once in a while – you wouldn’t want to obtain such a perfectly accurate picture of chemical and physical interactions that there was no point in running any experiments at all, now would you. (Fortunately that is a rather. . .distant. . .possibility at best). But it would be nice to predict what a new compound’s melting point might be, or what crystal form it might take, or what any of its important solid-phase properties might look like.

And you’d think that we’d be able to do a better job of that, but sadly, that ain’t the case. For the most part (the very most part) we can’t really do any of those things. Not reliably, anyway, not from a standing start. If you know the properties of compound A, you can make predictions about what will happen to them as you start modifying A’s structure, and those can be fairly decent (up to a point), but if you just start drawing some new structure up on the board and ask “What’s that compound’s melting point?” or “How many polymorphic forms will that have?” or “What conditions should I use to make that crystallize, and what space group will those crystals be in?”, all anyone can do is shrug. “Make some and find out” is the only realistic answer.

That goes even more for more exotic materials with more exotic properties. That’s the point of this recent piece in Nature, which concentrates on a recent field at the intersection of solid-state physics and materials chemistry, the investigation of topological materials. A topological insulator, for example, is indeed an insulator in the bulk (3-D) sense, but can conduct electricity (is in fact more or less forced to conduct electricity) along its two-dimensional surface – and only in one direction, at that. You rapidly get into the weeds of band theory, symmetry, and quantum-mechanical electron behavior when reading up on this field, but suffice it to say that there are some really unusual effects of electron behavior (spin states, etc.) overlaid on the fundamental geometries of solid materials.

A very hot area is the prediction of such properties, or at least the computational narrowing-down of what materials to look at for signs of them. But as that Nature article points out, there are a lot of ways that such predictions can fail. There have been structures proposed that would have very interesting topology, but are just not thermodynamically stable compared to other readily accessible geometries. Others just aren’t in the predicted crystal form, or have magnetic properties that either weren’t properly predicted or were left out of the calculations entirely. And some of the predicted materials have (and require) very high symmetries, which are states that are often difficult to realize in the real world, since there are all sorts of symmetry-breaking defects that have a way of sneaking into actual samples.

That reminds me of something a crystallographer explained to me a few years ago, about what happens as you get better and better crystal structures. What looks at first like a reasonable crystal with a reasonable unit cell might, on really close inspection with better data, turn out to be more complicated. The “real” unit cell, once you know things to much higher resolution, might actually be an assembly of many of those simpler ones, each of which are deformed subtly in a new symmetrical arrangement which then repeats as a much larger “supercell”. This is just the sort of thing that’s going to be very important in figuring out topological material properties, and it requires both high-quality data and high-quality crystallographers to interpret it.

This all may sound rather removed from the concerns of medicinal chemists and drug developers. But the same problems apply across the field. Compound crystal structures (and melting points, etc.) are of great importance in formulations work, and polymorphs are notorious in drug development. More broadly, the calculations needed to try to predict solid forms of materials overlap with the techniques we’d like to use to predict small-molecule binding and protein structures: when you get down to it, it’s all atoms and molecules interacting with each other. And in that short phrase are uncounted amounts of work, along with Nobel prizes, complete disappointments, breakthroughs and duds and entire careers, and things we haven’t even imagined yet.

7 comments on “Predicting – Or Not Predicting – New Materials”

  1. myma says:

    Ah, polymorphs. You can tie yourself in knots thinking about polymorphs. QC can tie you in knots forcing you to think about polymorphs.
    We had a semi-crystalline ingredient for a liquid sterile injectable formulation, about 80-some% crystalline, and two polymorphs within there. The molecule has a floppy area, which just will never 100% crystallize. For the drug product, it had to be dissolved, fill+finish including sterile filtered. There was an insane amount of work our formulation people had to do to characterize dissolution rates and filtering and content pre- post-filter to satisfy our QC people. And this was for a phase 1.

  2. Imaging guy says:

    Don’t worry. Machine learning will solve all these problems. Just throw in whatever independent variables/parameters you can measure (even if there are measurement errors and the variables are not known to be remotely related to the properties you want to predict) and the algorithms will predict whatever properties (dependent variables) you want to know. Trying to understand something deeply is for the losers.
    https://xkcd.com/1838/

  3. Marcello says:

    Schrodinger the company is not going to like that paper

  4. Peter Kenny says:

    Something that the predictive chemistry folk need to get better at is assessing the difficulty of prediction problems. For example, I would guess that (passive) permeability will be easier to predict from an arbitrary molecular structure than aqueous solubility. Given all the other things that one measures for during lead optimization, is it such a big deal to measure aqueous solubility? If you’ve got aqueous solubility measurements for a well-chosen subset of project compounds (i.e. spanning full range of lipophilicity and molecular size rather than just the most potent compounds) then you may be able to derive a useful model based on one or two physicochemical descriptors. If you fail to observe any correlation between aqueous solubility and lipophilicity for your project compounds then do you really want to be making decisions based on a model that might not have ‘seen’ anything like your project compounds?

    1. Druid says:

      Does anyone have a rule of thumb for the drop-off in solubility from first synthesis, which is probably “amorphous” or a mix of things, to later batches as they become progressively more stable crystalline forms? A lot of data comes from the first batch which can’t yet come from predictions, so there is a lot of value in the first few mgs, but if you could estimate problems that are going to arise in development, you might select a different candidate. Has anybody mined the data?

  5. Uncle Al says:

    I want empirical answers prior to investing in synthesis. If that means gutting HR personnel and stringing said obtentions across clothes lines to observe which carrion eaters arrive…it’s a viable method. MY time is valuable.

  6. Hood Rat says:

    Hot damn!

Leave a Reply to Uncle Al Cancel reply

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

Time limit is exhausted. Please reload CAPTCHA.