I've had a number of people call my attention to a recent Twitter thread involving drug discovery and AlphaFold, so I thought I'd go into the details in longer form, rather than string together a long series of tweets. Russel Kaplan's series of posts starts off here, and I'll put them together:
Today I saw the impact that AlphaFold is having on speeding up drug discovery firsthand:
1/ A friend runs a biotech startup designing drugs to fight cancer. In prior work, they found that tumor cells make a protein that binds to two receptors in the body. Binding to just one of them would inhibit the tumor’s growth, but binding to both makes the tumor grow faster.
2/ If they could design a new protein that binds to only one receptor and not the other, this mutant protein might be a potent cancer drug.
3/ Before AlphaFold, finding such a protein would take ~1 month: order lots of mutated DNA sequences, insert them into cells, filter the cells which bind to one receptor but not the other, and sequence those cells’ DNA.
4/ AlphaFold unlocked a different approach: it found the 3D structure of the existing protein & receptors, which was unknown. With the structure + another ML model, they saw how binding affinities would change with different mutations. This led to an ideal candidate in 8 hours.
5/ Animal trials with this new protein will start soon. Still a long way to an approved drug, but it’s exciting to see drug discovery go faster with a neural network in a Colab notebook.
This got a lot of excited reactions, naturally, because people are always ready to hear that drug discovery has hugely accelerated (and with good reason!) Some of the other reactions were along the lines of "OK, sounds like they saved a month here" or "Sounds like they sped up slightly on what's not a rate-limiting step", and as you can imagine, I incline more toward those points of view. But read on - I'm going to try to give everyone their due.
Taking those tweets in order, though, there are some steps missing that I don't think a lot of general readers noticed. The first two conflate the process of "designing drugs to fight cancer" with "design a new protein", and those are unfortunately not the same thing. The process described above (making a selective receptor ligand) is exactly the thing that medicinal chemists spend a great deal of time doing with small molecules; it's a perfectly sound strategy if you have good reason to think that you need that selectivity. The tweets are describing doing that protein ligand as opposed to a small-molecule one. But you look at the entire landscape of proteins as therapeutics, you will find that the great majority are antibodies, and the next biggest category are replacements of human signaling proteins like insulin. There are very few examples of "design a new protein to make a new drug", and that's because that's a very difficult thing to do.
The "find/design a new protein" part is generally not the one that kills the project. It's the "turn your new protein into a drug" part. Proteins are (for the most part) not very stable on dosing. You generally don't even bother with oral administration, because your gut digests them like it would digest a bite of a hamburger. But even i.v. has plenty of problems. That's why there are so many antibody drugs, because in that case eons of evolution have already provided us with a protein template that we know is relatively stable in the bloodstream. Your new beast will almost certainly not be. This is also why, among the non-antibody-based protein drugs, so many of them are known proteins engineered to have longer lifetimes in the bloodstream than their original forms had. And your new beast may also turn out to be immunogenic - that's a worry with any protein-based drug, antibody or not. Our immune systems are constantly on guard for "Not Invented Here" proteins (thus all those zillions of different antibodies floating around) and if your new therapeutic idea sets off one of those tripwires, your new idea is in trouble. Adding to the complication is that (because of the variability of that immune system from person to person) it might be only a smaller group of your intended patients that react, and there's no good way to find that out for sure other than by dosing patients.
Another thing that's not clear from the tweets (which are, after all, not meant to be a full report!) is how well the AlphaFold predictions worked out in a real assay. The line about finding an "ideal candidate" in eight hours probably confuses people on that point. What it means to me is that they found a nice candidate for testing in the primary assay in eight hours, which is fine, but what it does not mean is "ideal drug candidate". You need to see if your calculated protein really does what you thought it would, and no, you can't skip that step. AlphaFold or not. We're still getting our bearings on how accurate these structure predictions are, and there are some pitfalls. You need to make the actual protein and test it against the actual receptors, and that takes more than eight hours. As the later tweet says, there's still a long way to go. But even there, another complication is that saying "still a long way to go to an approved drug" makes it sound like the wait is for human clinical trials and regulatory reviews - it doesn't quite get how there's still a long way to go before you can even imagine starting any sort of clinical trial at all, and how most lead compounds collapse before they get that far.
One of those steps towards human trials is the animal testing that's mentioned in the last tweet. The first things you'll want to do are make sure what the blood levels and half-life of your new protein are, and that it's not obviously toxic. You do that in normal animals - calling these "animal trials" as the tweet does will give some people the idea that you're going straight into mice-with-cancer, because it sounds like the direct analog of human clinical trials, but that's not how it works. And about those mice with cancer: the state of animal models in oncology is (and always has been) rather messy. Testing mice with actual mice tumors is often not so useful, because the molecular targets can be a bit off what they are in humans. So there are human xenografts, done with tumor cell lines in immune-deficient mice. Those are OK as far as they go, but they don't go that far: if your new oncology drug candidate doesn't work in those, it's probably a lost cause, but if it does work, that still doesn't mean it's going to be any good in humans. The latest improvement in that area are "patient-derived xenografts", using actual human tumor tissue rather than cell lines, and those tend to be more meaningful.
And this brings up another risk factor in any of these projects: your disease hypothesis. The initial tweet says about the two receptor targets that "Binding to just one of them would inhibit the tumor’s growth, but binding to both makes the tumor grow faster", but that makes things sound more straightforward than they are. There are all sorts of ideas that have worked at that level that turn out to be pretty useless in actual human disease, sadly, and we often don't quite know what went wrong. Remember, the failure rate in oncology clinical trials is up around 90%. The other thing to remember is that when you hear "the tumor" that this means "this specific tumor for this specific type of cancer". Some people reading this series of tweets may have assumed that ideas like this apply to any sort of tumor, but there's almost nothing out that that applies to any sort of tumor. Cancer is really several thousand diseases, all of which share the phenotype of unregulated cell growth. And there are a lot of ways for cells to slip into that state.
All of this takes us back to that rate-limiting step idea. What I've been trying to get across is what I was saying back when the AlphaFold results came out: the biggest problems in drug development are not really solved by having faster access to protein structures. We can have problems with getting lead compounds against targets, true, and we have preclinical issues with things like blood levels and half-live (especially with protein candidates). But we have far bigger issues with target selection (does this idea of ours really do anything against the disease, in the end?) and with unexpected toxicity (what else does our candidate do in real animals and real humans?) Those last two are where that 90% clinical failure rate comes from.
It looks to me like the AlphaFold software saved these folks about a month, as the tweets say. Which is fine (assuming that the predictions worked out). But now all the parts of drug development that are not accelerated by protein structure knowledge start, and those take just as long as they ever did. And the failure modes that sink almost all of our projects are still waiting, just as before.