Quantum computers, AI, and Drug Discovery: Q&A With Nadia Harhen, GM of AQBioSim at Sandbox AQ

Nadia Harhen is the general manager for the molecular simulation optimization group at Sandbox AQ. She recently spoke with Medical Device & Technology about the work she and her team are doing using new technologies like quantum simulation and AI to improve the drug development process.

(MDT:) Can you explain Sandbox AQ’s work with drug development?

Nadia Harhen: Let me just touch on the AI side really quickly. About two weeks ago, the pharmaceutical industry had a monumental moment where the first drug that was designed purely by AI went into phase II of clinical trials. That’s fantastic and we’re excited, especially because it took about three years to get it to this point. We all know the failure rate in phase II and III is quite high, and something like a third of tested drugs will fall out of the ability progress past phase II.

With AI-based tools, it remains to be seen what will happen. That’s based on a few things. First, it requires a lot of training data in order to actually come up with the predictions. What we focus on at SandboxAQ are undruggable targets that don’t usually have a lot of data, or in situations where binding affinity is not enough of a signal to proceed. In these situations, you must bind to something and make a series of chemical reactions happen. Typical AI tools and data available don’t necessarily address those types of problems.

There’s a large space of highly impactful problems in neuro degeneration and cancers, the hard diseases that are often associated with end of life. Rare diseases also fall into this category in some cases. The industry did have a drug get approved a few weeks ago for cognitive decline in Alzheimer's patients, but that’s still not addressing the problem at its core.

What we’re seeing is that the mechanism of reaction is not well understood, or the therapeutic hypothesis is wrong to begin with. When it moves through clinical trials, it fails. You have that assumption about what should be binding to what incorrect. We’re trying to really focus on these problems. When we use AI, we’re focused on edge cases, not the mainstream like many others and casual relationships.

Those types of approaches, like understanding the therapeutic hypothesis and what is happening in the body. We’re doing work with a Noble Lauriat, and he has been studying prion proteins for years. These are things that are associated with dementia, Parkinson’s, and other neurodegenerative diseases leading to cognitive decline. When we showed him the simulation, which is a physics-based simulation, of what is happening when these prions bind, he was blown away. He had never seen this before in any of the models or studies.

That’s the sort of thing that the tool sets allow you to do: see what is going on at the cellular level. You can adjust for thermodynamic properties and other cellular activity that you can’t simply see in an assay that is just lighting up green when something binds.

Normally, when you study a disease, you find a molecule to bind to it. You study it with animals and then you get into human testing. When you fail clinical trials, you’re looking at the compound and maybe it has a lot of the properties you’re looking for, but it failed in a way that stops the trial. One of things our module does is it allows you to optimize on that compound so you can get right in that Goldilocks zone of effective and safe.

You can move them throughout the process, so you don’t have to wait until the end.Not only are you compressing the timeline, but you’re moving through the most optimal version of those compounds. You’re more likely to be successful in clinical trials because you’re doing things computationally instead of in a serial manner.

(MDT:) What is AI’s relationship to the work being done by the doctors and researchers?

Nadia Harhen: What we’re doing is physics based simulations computationally, and then we’re using the output of those simulations to train AI models. The reason we’re doing it that is that the physics simulations are accurate. When you train your AI model on things that are done in reality, they’re very likely to be predictive in an accurate way. This works better than starting with data that is not casual or trained on ground truth, which is what a lot of AI models do.

(MDT:) What future applications do you see for these technologies (AI and quantum computers) in the pharma industry?

Nadia Harhen: There are companies out there that are building quantum computers for drug discovery. We do not recommend that any company develop a specialized computer for this purpose because the hardware and software have got to be treated so specifically for that use case that it becomes hard to capitalize on that investment in a way that is future proof.

There are companies that are doing it, however, especially for harder to treat situations. In terms of future applications, a big area that we see that could bear fruit is in optimization, especially in what we call the traveling salesman problem. That refers to companies are looking at different computations and permutations for something like a manufacturing process. There’s probably a broader application there, where if you do it once, then you can use that case again and again in different industries.

In terms of drug discovery, we’re not seeing that there’s massive applicability. Companies are doing it, so clearly, they’re seeing something, but they may also just be doing this because they don’t want to miss the boat.