Predicting ADMET On Rowan
Predicting ADMET On Rowan
pros and cons of global ADME/toxicity prediction; the ADMET-AI model; some case studies; a cautionary note
When people talk about drug design, they often talk about finding a ligand to “hit” some protein target. This is oversimplified. Choosing a target and finding a lead compound with activity against it is indeed crucial, but there are a ton of other considerations to go from in vitro activity to successful in vivo outcomes:
The compound must be absorbed by the body successfully, both into the bloodstream and then into cells (for intracellular targets).
The compound must get to the right place in the body (distribution).
The compound shouldn’t be metabolized before it can act or interfere with important metabolic processes.
The compound should be excreted from the body at an appropriate rate, which impacts dosing schedules &c.
And the compound needs to not cause toxicity in patients, the most common cause of drug failure in clinical trials.
These considerations are often lumped together as ADMET, or ADME when toxicity is considered separately. Over the past few decades, there’s been a movement to incorporate ADMET considerations early in the drug-discovery process, which a recent editorial in J. Med. Chem. argued deserves the Nobel Prize in Medicine:
We believe that property-based drug design, used in concert with other drug design approaches and other drug discovery-related advances, has enabled the development of more novel and complex oral small molecule drugs… While drug design previously focused primarily on optimizing potency, [property-based drug design] introduced a more holistic approach based on the consideration of how fundamental molecular and physicochemical properties affect pharmaceutical, pharmacodynamic, pharmacokinetic, and safety properties.
Property-based drug design is tough, though. Many ADMET assays require animal studies and a non-negligible amount of synthetic material, making it tough to quickly guess how a given structural modification will modify e.g. cardiac toxicity. As such, there’s a ton of interest in predictive ADMET modeling that can be applied earlier and more often in the drug design process. If ChemDraw automatically flagged any toxic molecule right when you drew it, that would be great!
Do General ADMET Models Work?
This is a contentious question in the field, and in accordance with Betteridge’s Law of Headlines the answer is usually “no” or “not yet.” Pat Walters has a fantastic post about the limitations of existing datasets: basically, you can’t hope to teach an ML model to accurately predict “toxicity” from 1400 datapoints labelled with whether a drug was toxic or not. There’s other fantastic articles floating around out there: Abhishaike Mahajan has a post examining why predicting toxicology is so hard (and unlikely to have an “AlphaFold moment” soon), while Alex Rich and Ben Birnbaum from Inductive Bio have written about the shortcomings of conventional global performance evaluation:
With all this acknowledged… ADMET models still sort of work. A recent collaboration between Inductive and Nested shows that while ADMET models fine-tuned on program-specific data generally perform best, global pre-trained models often aren’t that much worse. And from a revealed preference angle, there’s clearly a ton of interest from industry in ADMET models, which implies that there must be at least some utility today. Even when more powerful models are possible, quick tools that are fast and give reasonable answers can still be very useful.
ADMET-AI Now On Rowan
All this to say: we’ve just launched an ADMET prediction workflow on Rowan. Our workflow currently uses ADMET-AI, a model recently published by Kyle Swanson and co-workers (paper, blog post, Github).
ADMET-AI uses a combination of ChemProp (a GNN for property prediction) and RDKit to achieve the highest overall performance on the Therapeutic Data Commons leaderboard. You can submit calculations to our ADMET workflow, and we’ll run them with ADMET-AI and automatically display calculations in a helpful and easy-to-parse format.
Why ADMET and why now? We’ve had users asking for features like this for some time, but we haven’t felt confident enough in any of the existing models or benchmarks to commit to taking steps in this direction. In our hands, ADMET-AI has performed well enough that we feel confident that it’ll be useful in the real world. Here’s three case studies showing what we mean:
1. Predicting hERG Toxicity
A 2008 J. Med. Chem. study investigated δ-selective opioid receptor agonists as potential painkillers. They found that some of the compounds under study inhibited hERG channels in HEK293 cells, which can cause long QT syndrome and life-threatening arrhythmia, and conducted a series of structural modifications to reduce hERG liability. Substituting the para position on one of the phenyl rings had a clear structure–activity relationship: while most substituents showed high hERG inhibition (IC50 < 1 μM), the carboxylic acid showed no hERG binding and the diethyl amide had a higher IC50 of 7.9 μM.
ADMET-AI successfully identified the carboxylic acid as the only compound with an IC50 > 40 μM, which is what the binary classification task in TDC aims to do. (We checked that these compounds weren’t in the training data.)
The other compounds are all correctly predicted as “dangerous,” but the model’s score doesn’t help us distinguish which are more or less dangerous, which is one of the limitations of classification-based ADMET tasks. Still, this is a pretty solid success.
One potential criticism: carboxylic acids are a well-known pharmacophore-based method for reducing hERG inhibition, and any seasoned medicinal chemist would suggest this modification (Di & Kerns’s book on ADMET suggests this in the hERG chapter). While this may be true, I think this actually demonstrates the value of a model like this. Conventional med chem intuition is difficult to acquire, and having a model at hand that can point out obvious changes like this can be very valuable!
2. Predicting CYP Inhibition
A 2000 Novartis paper on p38 MAP kinase inhibitors studied modifications to avoid CYP inhibition, a common source of drug–drug interaction-induced toxicity. They found that moving from a phenyl ring to a 4-hydroxypiperidine, along with some ancillary modifications, maintained activity while dramatically reducing inhibition of CYP3A4, CYP2C9, and CYP1A2. (Inhibition of CYP2D6 increased but was low.)
Our ADMET workflow successfully predicted the decreases for 3A4, 2C9, and 1A2 (left => right in the below image), but erroneously predicted that 2D6 inhibition would also decrease. Additionally, the magnitude of the inhibition isn’t predicted correctly: the left-hand compound inhibits 2C9 more than 2D6, but the opposite trend is predicted. (These compounds also aren’t in the TDC training dataset.)
3. Predicting Permeability
A Kymera paper published two weeks ago studied new CRBN binders as part of a program looking at IRAK4 degraders. As part of this study, they ended up methylating their initial CRBN binder to increase passive permeability, which worked. Our workflow was able to reproduce this same effect (see row “PAMPA”):
Once again, “removing a free N–H increases cell permeability” is hardly an earth-shattering insight, but the sane behavior of these models on real-world tasks gives us confidence that the ADMET-AI model can be useful for real scientists today.
ADMET and Rowan
We’re excited to bring ADMET-AI to Rowan and help our users get quick and easy ADMET insights with the molecules they already have in Rowan. To be intellectually honest and make sure our users don’t overestimate the confidence of these predictions, we’ve: (1) flagged potentially inaccurate predictions, (2) removed the TDC subsets we think are ill-advised like ClinTox, and (3) added a disclaimer to every ADMET workflow.
(Zero-shot ADMET prediction is, in the grand scheme of things, pretty crude. We encourage any users with important ADMET needs to explore more sophisticated solutions like Inductive’s platform.)
We hope that these results can be useful, and are excited to hear about what we can add or enhance based on real-world feedback. Moving forward, we’ll be looking for ways to add new useful ADMET models to our platform and improve the quality of these results.
Thanks to many people for helpful discussions about property-based drug design and ADMET modeling, including David Huang and Irving Ling.