One of the reasons that people in or near this business can write such gaudy press releases is that it has so many moving parts. That lets everyone claim that the part that they’re addressing is Crucial. Think of a car: the wheels are indeed key to mobility, but so is the engine. As is the oil, the power source (be it gas tank or battery), and any number of other parts. Or you can use the human body as a metaphor: there are parts of it that you can cut off or out, but there are quite a few parts of it that you can’t. All of the latter have a plausible claim to being crucial.
Let’s consider, though, this claim which is the headline of a Stat story from the last couple of days: “Amazon, Google, and Facebook are using AI to find protein structures”, it says, “The tech giants are becoming increasingly active in deciphering protein structures, a crucial step in helping to find drugs” We heard a lot about how crucial this was during the last protein-folding competition as well. But is this true?
I’m going to sound like a heretic (at least to the people doing that work!) and say no, not generally. There are cases where such structures can be very helpful, but there are cases where they don’t do you that much good at all. I would not want to do fragment-based drug discovery without access to ligand-bound structures, for sure, although it’s certainly possible if you don’t mind wasting a lot of time feeling around in the dark. But if you have a solid primary assay against a target you really believe in, and an animal model with real translatability, then no, you can push right ahead. Don’t you need a defined target to get a drug to market? No, you don’t, actually, although it certainly can help. What you absolutely need are safety and efficacy data. And those you need successful human clinical trials, and to get to those you need for the FDA to allow your IND application. The best way to get one of those approved is to show activity in a relevant animal model and clean toxicology studies in at least two species, and you can do those without knowing the protein target at all, let alone its structure. There aren’t enough reliable animal models in the world for this to be a common route, of course, but it’s certainly possible (and many years ago, before my time, it used to be the rule).
Even as you read through the Stat article, this point becomes apparent (to their credit). Protein structure determination simply isn’t a rate-limiting step in drug discovery in general. If we were (far) better at modeling potential drug candidates to such structures, that would make things a bit different, but we’re really not at that point yet (cue the AI people who are doing such virtual screening!) Both of these fields – prediction of protein structure and prediction of small-molecule binding – are advancing, but neither of them are ready to be the backbone of a company’s drug discovery efforts yet. They’re tools, and they’re sometimes very useful and sometimes a waste of time and effort, which you can say about most of the other tools as well.
I welcome the big-tech folks to the protein structure party, of course. We really do need more insights in that area, and if such predictions are every going to be generally useful we’re going to need all of the insights we can get (and all the processing power we can get, too, most likely!) But if someone is telling you that protein structure prediction is going to lead to a big leap in drug discovery efficiency, hold on to your wallet. What would lead to such a leap? Off the top of my head, I’d say better prediction of useful drug targets, more translatable disease-predictive cell and animal models, and earlier assays that are more predictive of human toxicology. Those, as far as I’m concerned, address the real killers in the whole process. Protein structure just isn’t on that list.