Can an AI beat an in vivo model?

Last week, Axiom Bio came out of stealth mode with the goal of eliminating drug toxicity without using animals. At first I was impressed at how well they timed this with the FDA announcement about moving away from animal testing. But after I had the chance to talk to Axiom’s founders, Brandon White and Alex Beatson last week, I realized that there’s an even more interesting angle on what they’re doing. In particular, they have the potential to fundamentally change the role of early discovery within the drug development pipeline. Let me explain…

---

This newsletter explores news and trends in life science software. If there's something you think I should cover (even if it's self promotion) send me an email at scalingbiotech@substack.com. I can't promise I'll use every suggestion, but I'll try.

---

The thing that has always bothered me about the early discovery stage of the drug development pipeline is that it’s very easy to look at the numbers and decide there’s no point in trying to improve or optimize it. Which is not to say that people don’t try to improve it… just that that from a purely rational, ruthlessly economical perspective, it doesn’t necessarily make a lot of sense.

The problem is that it is impossible to tell with any certainty if a drug will succeed in clinical trials based on what is known at the end of early discovery. Roughly 85-90% of drugs fail in clinical trials. And that’s even after the translation/pre-clinical/in vivo work that happens after early discovery. At that point, there’s just no way to tell if your drug has more than a 10% chance of getting approved.

If you could optimize early discovery - make it faster or cheaper - then you could either get to that 10% chance sooner or get a larger selection of 10% chances in the same amount of time. But early discovery is already relatively fast and cheap compared to the rest of the pipeline, so faster and cheaper doesn’t get you much. And since you can’t tell the difference between one 10% chance and another, getting more of them doesn’t actually help.

So it’s easy to look at this and conclude that there’s no point in doing either. Which is probably why most of the investment goes into making the later, slow and expensive stages of the pipeline less slow and expensive.

But what if, instead, you could do something to identify drug candidates with a higher than 10% chance while still in early discovery?

Now, obviously, this would lead to huge savings by reducing the number of failed clinical trials. But it would also change the calculus around where to invest and optimize: If you could actually tell the difference between which candidates were more or less likely to succeed, then being able to cast a wider net and find more of them would actually be worth it. The investment that currently goes to reducing the cost of clinical trials would be better spent optimizing early discovery.

To be clear, this is my own convoluted thought process, not necessarily how Axiom’s founders got where they are. But they ended up in the same place: Looking at the reasons drugs were failing clinical trials. And while there’s obviously a wide array, they landed on a relatively simple one: About 20% of small molecule clinical failures are because of liver toxicity. And this is *after* the drugs go through animal models. It turns out that even animal models aren’t that good at predicting toxicity in humans.

So Axiom hasn’t just managed to build models that are as accurate as animal models. They claim that their models are significantly better, with 75% sensitivity at 95% specificity. There isn’t complete data on the accuracy of in vivo toxicity models because drug candidates that don’t pass animal models usually don’t get published. But it’s probably something like 50% sensitivity.

Axiom’s model is possible because of three things:

1: The Data

Publicly available datasets of clinical toxicity are pretty sparse. The FDA’s DILIrank has about a thousand compounds, but without dosage and PK. Other datasets that do have PK and dosage are much smaller - on the order of 100 compounds or less.

So Axiom started by creating its own internal dataset of clinical toxicity. Using LLMs to scrape the literature and other public sources of information, they were able to amass data on almost 10,000 (and growing) small molecules that have been tested for liver toxicity in humans. This gives them a source of truth for their models.

2: The Assay

Axiom’s model has both an in-vitro and an in-silico component. The in-vitro component is a high-throughput adipocyte assay with a high-content imaging (HCI) readout. This builds off of lower-throughput or less accurate assays that other labs have developed. (They’re also working on a multi-cell-type assay that should be even more accurate.)

They’re able to make the assay high throughput without sacrificing accuracy by using an image recognition model to identify more subtle phenotypes. Most of the other assays just use cell death, which turns out to be too blunt of a readout. By training a phenotype identification model on their dataset of 10K compounds, the model learns a more accurate correlation with clinical results.

3: The Models

There are two models that drive Axiom’s predictions: The image recognition model tied to the assay, and a structure model that can predict toxicity for compounds that haven’t been run through the assay, or even synthesized yet. This means the model can be used during the optimization stage, where synthesis is by far the most time consuming and expensive part of the process.

Note that models for identifying HCI phenotypes and for predicting properties of new small molecule structures have both been done before. While Axiom has used some novel tricks to make each model work in this context, the thing that has made this all possible is the dataset.

Of course, this is just for one type of toxicity. Axiom’s long term plan is to build similar models for other organs and other modalities. The more models they can build, the more they’ll be able to lower failure rates. Perhaps one day, clinical trials failing because of toxicity will be a thing of the past.

---

Thanks for reading Scaling Biotech! When I'm not writing this newsletter, I help Life Science SaaS companies build sales and marketing systems tuned to the specific needs of a complex industry.

If you're looking for ways to connect the dots between building great tools and getting them into the hands of users, click here to request a free gap analysis.