Hey folks, it's Corin here, co-founder and CEO of Rowan, and I'm excited to talk about physics-based methods for predicting blood–brain-barrier permeability. So the blood–brain barrier is very important for anyone who works in central nervous system targeted therapeutics. Basically, there's this barrier called the blood–brain barrier that blocks arbitrary compounds in our blood from getting into the brain so that if we eat a poisonous substance, all the poison doesn't go to our brain and kill us.
However, as you might imagine, if you're designing a drug that's supposed to act against brain cancer, and you need to get this drug into the brain to deal with the cancer, this will be a big problem because your drug probably won't be able to get through the blood–brain barrier. And so although there are structural modifications you can do, and there's usually a way to get sufficiently small or lipophilic molecules through the blood–brain barrier, it's not always easy to know what we'll get through without doing a ton of expensive animal testing.
So people would really like cheap and effective computational methods to predict what compounds will get through the blood–brain barrier, maybe predict it even before you have to synthesize it to save you a lot of time and money. But historically, this has been super, super hard. Machine-learning methods just don't perform well, there's not enough data, the trend is too hard to ascertain, and there haven't really been easy physics-based methods to do so.
So in 2023, scientists from Schrödinger came up with this really clever approach based on the energy of solvation. And specifically what they're trying to do here, this is a Drug Hunter article on this research, is predict this property called Kp,uu. And this is sort of a mouthful, but this is the unbound brain-to-plasma partition coefficient. And essentially, this is just a way to quantify how much a compound can get through the blood-brain barrier. So it's the ratio of how much is in blood plasma versus how much is in the brain. So this turns out to be a really handy way to quantify these matters.
And so what the scientists at Schrödinger were able to do is they can run a conformer search on their molecule. They can get the energy in the gas phase, and the energy in water. And from that, they can sort of back out the energy of solvation, so how much the compound is being stabilized in water. You know, this only works for neutral compounds. So for compounds which have, for instance, basic amines, you have to add this state penalty using the pKa to correct for neutralizing the compound and sort of the free energy costs there.
But intuitively, I think this makes sense, right? So compounds that much more prefer to be in water are going to be much less likely to get through this greasy blood–brain barrier membrane. They're going to be much less permeable and much less likely to actually get into the brain. And it turns out when you actually do this, you get really, really good predictive accuracy here, pretty much unprecedented predictive accuracy for actually predicting which compounds will get through and which ones won't.
And so just to go down to some case studies, here's a series of compounds. These are imidazopyridines, so lots of sites. And it actually turns out that just along this series, one, two, three, you can see this compound has a Kp,uu of 0.08. So that's pretty low. We want it to be higher to show that more is getting into the brain. And then as we sort of take away hydrogen-bond-acceptor sites and add inductive electron-withdrawing groups here, we see that the Kp,uu increases, showing that these will be more likely to actually get into the brain.
So this work is super cool, but it's also very computationally intensive. You have to do all of this conformer searching and DFT calculations, and it ends up taking about 40 hours of CPU time per compound. So this isn't really suitable for high-throughput screening, and this is going to end up running quite the computational bill if you want to do this on a library of even just 100 compounds.
Here at Rowan, we just published a different approach that uses neural network potentials and our new Starling pKa model to generate very similar results in just a fraction of the time. So here's our ROC analysis of the accuracy of our method. And we can show that we're actually quite good at separating true positives from false positives. So this is an AUC score of 0.85. And we get about a 75% predictive accuracy, which isn't perfect, but it's pretty good for a blood–brain-barrier-permeability prediction.
What does this look like in practice? Let's head over to the Rowan platform to take a look at the result of some blood-brain barrier predictions and compare them to experimental results. So I've looked at the three compounds that we just discussed from the Schrödinger work. So here's one, here's two, here's three. And let's compare the predicted blood–brain barrier penetrants in each of these cases.
So here you can see this is part of our broader macroscopic pKa workflow that predicts the relative free energy of all these different microstates, as well as the logD. We're not really going to get into that right now. What matters from our point of view is this tab here, BBB, which stands for blood–brain barrier, as you might have guessed. And here we're predicting two things. So we predict the solvation energy in kcal/mol, which is exactly like we discussed earlier, how strongly this molecule will be solidated in water.
And then what we have here is the predicted likelihood that this molecule will be blood–brain barrier penetrant. And so this is sort of binarized at a Kp,uu of 0.3, which the Schrödinger authors allege is about right. If the Kp,uu is more than 0.3, it's probably going to get into the brain enough to be useful; and if it's less than that, it probably won't. This is done through logistic regression. And so what this value means is that we think this is a 41% chance of getting into the brain.
So, you know, this is okay. This is decent, but it's not perfect. Here, if we go over to compound two to compare, we can see that now we say we have a 58% chance of getting through the brain. So what's the change that we've made? So we can go here, we have this lactam ring, this five-membered ring that has a strong hydrogen-bond acceptor. And now here we've gotten rid of that. So we actually have a much weaker hydrogen-bond acceptor that's gonna bind to the water less tightly. So you see that the solvation energy has gone down from 18 to 16 kcal/mol. And as a result, the blood–brain barrier penetrance is predicted to increase.
Now, if we take even one more step, so adding an inductive electron-withdrawing group here in the form of this morpholine, we can see that now the solvation energy has decreased further and we expect even higher blood–brain-barrier penetrance, although it's a smaller difference. So you know that was 58, now this is 61, but the change is in the right direction.
And if we actually match this up to the literature results, again, we see that one to two is a big step up and then two to three is a little step up, but still meaningful. So these calculations are quite cheap, so we don't take 40 hours here. This one took nine minutes of CPU time. So this is, you know, this is fast enough to actually be done quickly and prospectively, so you know, before synthesis is needed. So you can use this to triage compounds and to quickly screen ideas and, you know, get quick and physics-based estimates of whether things will get into the brain or not.
One particular strength of this is that we're actually not fitting to any training data at all. So we evaluate against training data to come up with a logistic-regression model. But these, there's no sort of SMILES-based fingerprint matching here. So there's not really an ability for this model to overfit to data. So this should work just as well on unseen data as it does on data it's already seen before.
So if you're exploring new compound modalities for CNS therapeutics, you might want to check this out because this could save you a ton of time.