Hi, I'm Corin, CEO and co-founder of Rowan, and in this video we're going to talk about ADME/tox prediction. ADME/tox is a shorthand that's used in medicinal chemistry and drug design to refer to some of the drug-like properties that aren't directly tied to binding affinity. So A stands for absorption, D stands for distribution, M stands for metabolism, and E stands for excretion. So those pertain to how drugs get into the body, what they do in the body, and how they get out of the body. And then tox refers to toxicity—the worry that drugs will be bad. You know, will this drug cause one of a litany of different negative side effects if it's administered to a human being?
And these properties, unlike most of the other properties and workflows currently on Rowan, are not directly computable from first principles, at least not in a straightforward way in most cases. So thinking through predicting metabolism requires modeling the effect of the liver, the kidneys, all of this sort of stuff on drug molecules, and this is not the sort of thing we can just load and run a DFT calculation on. There are ways of course that DFT calculations like bond-dissociation energies can inform thinking about where and how fast metabolism will occur, but in general end-to-end metabolism or toxicity prediction from first principles just isn't very feasible because it involves something as massive and complex as the human body.
Instead, what's been done is to build empirical machine learning models to model and sort of guess what the effect of the human body will be on a given drug. Instead, what's commonly done is to build machine learning models trained on experimental data that provides a quick guess as to what a drug will do once it's in the body. This is what we do here on Rowan. So to run an ADME/tox prediction, we just come here and click “new ADMET prediction,” and we load in the molecule that we care about. So in this case, we'll choose a simple molecule just to show off some of the basic features of the workflow. And of course, you can come here and at no cost just make a free account and try out your own molecules and see what you think of these predictions. So we'll draw nitrobenzene just because we expect this to be a pretty interesting result. We'll click save, we'll say “nitrobenzene,” and we'll say “submit ADMET.” So here we go.
On every ADME/tox page we have this warning: global models of ADME/tox are not substitutes for experimental measurements. This is really important to note because especially with properties as complex as these, it's really important to know that these predictions are useful, but they're not as good as running an experiment. And if you are actually in a drug design campaign and you need to make mission-critical decisions, you should actually pay the money and measure these properties yourself with a CRO and a vivarium and any of these things, because these are really better thought of as guesses. They're not as good as experimental data. And there's plenty of articles we link here pointing out that this is the case.
All that being said, there are a lot of ways in which these quick experimental measurements can actually be very useful in encoding some of the intuition that we as chemists have about molecules. And so I'll walk through the rest of this without making any further caveats. So the first page that we see here on the ADME/tox prediction screen is this one—we can see the physicochemical properties. This just tells you simple things like how big is this molecule and molecular weight, how greasy is it, how many hydrogen-bond donors or acceptors it has. And on this page, as with all of the ADME/tox pages, we have everything sort of highlighted here. So green means that this will probably make an effective drug based on human encoded heuristics. Yellow means you're sort of in a gray area: maybe it'll be an okay drug, maybe it won't. And then red, which we'll get to later, shows that this is a little concerning, you might want to take a closer look at this and figure out if you should make some modifications to the structure. And again, all of these are just estimates, and there's, of course, exceptions to everything here.
If you want to see exactly what the colors actually correspond to, we can hover over this info icon. So here it explains logP. So logP is the partition coefficient, which is the ratio of solubility between octanol and water. And then it shows what good values are: so a good value for logP generally is considered to be about zero to five, you can have a little deviation from that and it's fine. But then really, really greasy or really, really hydrophilic compounds are both unlikely to make great drugs because they probably won't be absorbed to get into cells the right way. Okay, so this is physicochemical properties.
Here if we go over to absorption properties, this gives us things like cell permeability, so PAMPA, Caco2 permeability, and then, you know, a variety of other things like logD, so distribution coefficient, solubility. And here, actually, this is a pretty good metric, because solubility here is predicted to be not so good. Actually, nitrobenzene generally is not miscable with water, at least as a bulk substance, so it forms little droplets in water, and so this is a correct flag that this might not be the most water-soluble compound on Earth.
If we go over here to metabolism, you know, we have here most of the CYP inhibition is predicted to be okay, but there is predicted to be inhibition of CYP1A2. So that, you know, has sort of a high predicted inhibition, which could lead to bad drug–drug interactions, and is generally something people worry about. And this molecule is predicted to be a substrate or possibly a substrate for CYP3A4, so that's something else to think about. Here we have, you know, this is predicted to get into the blood–brain barrier. This is true, I believe nitrobenzene does cross the blood–brain barrier quite quickly, so that's true and then there's some other metrics here like volume of distribution. Now we come to excretion so this is predicted to be cleared and excreted quite quickly, so that might be bad for a drug.
And now we come to toxicity. So here we can notice that we're setting off some flags so there's a lot of different toxicity warnings here for various different mechanisms of toxicity or individual targets. hERG is a pretty famous potassium ion channel, which often causes issues with basic amines. The Ames test for mutagenicity is another good one. And here we can see that nitrobenzene is predicted to cause drug-induced liver injury and skin sensitization. And, you know, this is not surprising at all because nitrobenzene and nitro-containing compounds in general are actually quite toxic. So this is an acutely toxic compound which is one of the reasons I chose it for this: I believe it is bad for the liver, it's also bad for the nervous system which doesn't make it into here, as I can recall I think it's also bad for your skin. It's really bad for most parts of your body, so if you're exposed to this compound you should definitely seek medical attention acutely. So this sort of shows that this is a legitimate concern: we've flagged a compound that's bad and it does indeed seem to be bad.
So that's an overview of the ADME/tox workflow. Another thing we can do here is we can submit another ADME/tox prediction. Maybe we'll say, okay, nitrobenzene wasn't a very good molecule, maybe we can add some steric shielding here. Maybe this will make it a better drug. So what about 2,6-dimethyl nitrobenzene? So nitrobenzene, we'll say 2,6-dimethyl nitrobenzene. You know, personally, I doubt this will have a huge effect on the predicted or actual toxicity because a nitro group is pretty bad no matter where it is. But nevertheless, we can look and sort of see, well, at least the model thinks this will cause less liver injury, but will still be bad for the skin. I don't know if that's actually true. I don't know how toxic this compound is—that'd be interesting to measure.
But one of the things we can do is now actually compare these two by selecting multiple compounds and see them side by side. So we can see here, you know, we've obviously changed the lipophilicity. So we've added more methyl groups, we've made this slightly greasier. We've changed the molecular weight, we've changed, we actually haven't changed the total polar surface area, because we've not added any polar functionality. And we've mitigated a lot of other things. So you know, the CYP profile is predicted to be different, we're predicted to be less toxic to the liver, we're predicted to still be a skin sensitizer, etc. And we can download these results as a CSV if we want to for further analysis. That's how to analyze ADME/tox properties on Rowan.