AI Drug Discovery with NeuroX1’s Max Dordevic

Why is the current approach to small molecule drug development? 

It will be helpful to use an example of what it more or less looked like when drug discovery first came about. Let’s look at aspirin for example, which is derived from salicylic acid found in willow bark. Salicylic acid is effective at reducing pain and inflammation, but it also causes significant stomach irritation, which is obviously not desirable. But by chemically modifying salicylic acid, specifically by adding an acetyl group, chemists created aspirin—a compound that retains the pain-relieving properties but with much less irritation to the stomach. So that’s classical medicinal chemistry: you take a molecule with a property you want and other properties you don’t want, and you modify its structure—by adding or removing a functional group, for instance—to achieve the desired effect. We’ve been doing this since the late 19th century. 

Starting in the 1980s, medicinal chemistry has started to evolve into chemoinformatics and high throughput screening. There are two variants of this: high throughput screening in the real world and high throughput screening computationally. In the early 2000s, chemo informaticists put together these compound libraries with millions or even billions of compounds and started to computationally guess and check so to speak. You can almost think of it like a janitor walking around with a big key ring of molecules and you have a locked door—your drug target—that you’re trying to open. So you just keep shoving different keys in until you find some that will unlock it so to speak. Of course, you’re also looking for specific characteristics of that “key” i.e. biological activities or effects.

With a big enough library, some fraction of those compounds are going to be hits. Once you’ve found the hits, you take them into a wet lab to see if they bind because the computation isn’t always accurate. So now you’re guessing and checking again, but this time in the real world. Some fractions of the compounds that you test in the wet lab are likely going to do what you thought they would do so the next step is to take those, run more experiments, drug some mice, etc., and keep narrowing down the candidates until you have a lead asset you want to take into human clinical development. 

How does NeuroX1’s approach to drug discovery differ? 

Well, what we did is say screw all of that guessing and checking—you can just let the target dictate your end goal by defining the desired properties and constraints. So rather than trying to shove a million different keys into a lock to see what works, now you’re molding a million keys designed for that lock to see which ones have the properties you want. You get a lot more hits and it increases exponentially as you keep iterating and seeing what properties you need. Ultimately that means you get to your lead molecule a lot sooner because all of the output is designed for that target. Being able to get to that lead molecule faster is a big deal because the chemical space is huge—there are more possible molecules for medicinal chemistry than grains of sand and you’re never going to be able to guess and check all of those. 

This is becoming particularly important because most, if not all, of the low-hanging fruit has been picked. In pharma, there’s something colloquially called Eroom’s law, which is just Moore’s law backward. So if Moore’s law is dictating that the number of transistors on a chip is doubling every 2 years or so, we have the opposite problem where every time a drug gets approved for a given disease it becomes more and more expensive to develop drugs for that disease. The reason for that is the incentive in pharma is, of course, to have the best-in-class molecule, biologic, etc. In other words, everybody wants their drug to be “Better than the Beatles” but unfortunately most drugs only provide modest improvements over existing drugs, which then requires larger clinical trials to demonstrate efficacy. So if it cost $50 million to get a drug approved in the 1960s today it would probably be closer to $2.5 billion.

Neurodegenerative disorders offer tons of opportunity in that there’s a lot of white space, but the mechanisms of these diseases aren’t well understood. There isn’t a great way to diagnose Alzheimer’s, for instance, especially in the early stages with the kind of accuracy somebody not totally immersed in this space would expect. For example, because we don’t know what causes it, though there are plenty of theories, identifying a target is difficult because even though we can mold a great key, we don’t know if we’re trying to unlock the correct lock, let alone unlock the right door. That’s the case with a lot of neurodegenerative disorders like MS and ALS. It becomes this weird cat-and-mouse game between trying to ID the target and then developing the drug for that target. So, we’re trying to solve that on both sides. 

Are there other companies taking a similar approach to neurodegenerative or other diseases? 

There are a lot of cool companies in our space doing a lot of really great work and you kind of need them all because the odds are so low that a given lead-candidate molecule is going to get approved. So, you need to have a pretty robust pipeline across companies. There’s arguably the most opportunity in neurodegenerative disorders for taking this approach but there’s a reason there’s so much opportunity—this is all the hard stuff. Building something and sending it to space is easy by comparison. I say that sort of tongue-in-cheek, but it’s true. Going to space uses known physics and it’s mostly an engineering problem. But with drugs, you’re dealing with physics, chemistry, and biology—which I like to say are similar sciences but at different zoom levels—and, even for experienced drug hunters and large pharma companies, true understanding of how all interact isn’t something any single person can predict with 100% accuracy. It’s a part of the equation that everybody has to deal with. What causes Alzheimer’s? Why did a particular drug fail its MS trial? We don’t know, but that’s the fun part of working in this space. It’s hard but you’re working on something worthwhile. 

Why is NeuroX1 focused so strongly on Alzheimer’s? 

Alzheimer’s is one of the last few diseases that we don’t have a good therapy for. Lilly just got a drug approved, but all it does is sort of slow the disease progression because we don’t really know what causes Alzheimer’s. It’s like if your sink is broken and it’s draining water into your apartment. What these drugs do is target amyloid–a plaque that builds up in the brain—which is like getting a bunch of mops to stop the leak in your apartment rather than plugging the burst pipe. If you don’t stop the leak, the apartment will still flood eventually, it just might take longer. So that’s sort of where we’re at with Alzheimer’s. We’re not targeting the right thing, but we know we’re targeting something that’s relevant, which is why these drugs slow the rate of change. To me, it’s an interesting question of whether these drugs that slow Alzheimer’s should be used because, while they do represent significant progress, they ultimately are a stop-gap solution until curative therapies come to market. My grandma just died of Alzheimer’s and it’s one of the reasons I wanted to start this company. She had been diagnosed years ago but just died a few months ago. The thing about Alzheimer’s is there’s a certain point where you’re not really lucid anymore, which makes the question of life extension challenging. And it’s something a lot of people are going to have to deal with because the lifetime risk is roughly 20% for women and 10% for men.  

How do you do target discovery using your platform? 

So, it’s easier to go after a known or well-validated target when you’re doing drug discovery because there’s a lot of data behind it. Once you have your platform built out you can start doing stuff on the target side, which is a lot harder because you must validate the target. It’s just a much more difficult process than running a binding assay and seeing if your molecule attaches to your target as you’d expect. You have to go out and get the right kind of cells for your assays and run tons of wet lab experimentation. So right now, while we’re building that aspect of it out, we’re focusing on known targets and making the platform better on the chemistry side. One way we’re doing that is by focusing on rare diseases where the target is well known but there’s not much financial incentive to go after it. So that allows us to prove our platform with these rare diseases while also focusing on big indications like MS. 

NeuroX1 uses a physics-based AI model for drug discovery—how does that differ from, say, a transformer model? 

So a transformer model is what ChatGPT or another type of LLM would use, and it is useful for basic research in life sciences, like finding relevant papers about a target or getting a good understanding of the existing and future pipelines in the disease area you’re looking into. They’re basically a great basic science and market research tool. In the case of tools, molecular interactions. But these are very very new models that are not well validated, in the case of diffusion-based models like AlphaFold which can predict hundreds of millions of different protein interactions without having to go get an x-ray crystal structure, if you’re building out a program, you are going to go get the crystal structure for the target you want to drug. So instead we use physics-based AI models to generate and simulate how novel small molecules interact with biological targets, allowing us to tailor treatments specifically for neurological conditions in a way that minimizes trial and error by applying well-established scientific principles to guide drug design and optimization. In essence, they’re less error-prone, and flavors of physics-based drug discovery have been around for decades. 

What’s the blue sky vision for NeuroX1? 

Tackling Eroom’s law head-on and buck the current trend of increasing drug development costs by leveraging physics-based AI to streamline the process of drug discovery and development. We’re not just looking for small improvements here – we’re aiming for big breakthroughs that could reshape how we bring new neurotherapeutics to market, both in terms of economics and timelines. We’re working on both sides of the equation – target identification and drug development. It’s a bit of a cat-and-mouse game, especially with neurodegenerative disorders where we’re still trying to understand the underlying mechanisms. But that’s what makes this work exciting. We’re dealing with some of the hardest problems in medicine, but the potential impact is enormous.