The Oncologist Running Quantum Algorithms Between Clinic Hours

Cracking Cancer’s Data Code

Alexander Pearson, MD, PHD

Oncologist and Computational Scientist, University of Chicago

Pearson runs quantum-classical hybrid algorithms in search of cancer biomarkers across genomic, imaging, and pathology data at UC’s Research Computing Center. He also will be part of the SC26 Regional Tech Hub, a multimedia forum making its conference debut in Chicago. Pearson will contribute to the Hub’s Human Impact/Translational Biomedical Research area, which will focus on how advanced computing serves society by transforming research, medicine, and healthcare.

Hello Alex. You run a research center and a clinical program in the same week. How did you end up wearing both hats?

My first love was mathematics. I took a statistics class in high school and just thought, “Oh my gosh, there’s so much we can do by processing information in unique ways.” I was an undergraduate biometry and statistics major, and then applied into MD-PhD programs with a PhD in statistics.

In those days, the utility of statistics was viewed more narrowly. Many programs thought statistics departments might be helpful for things like analyzing crops or building clinical trials, and the information age in clinical medicine was really just at its beginning. But I was aware that new data modes like genomics were becoming more routine, and that they would require unique ways of processing. So I did my statistics PhD work in machine learning on methods that could process multiple genes’ expression information together.

Just as I was transitioning into independent practice, deep convolutional neural networks were becoming available on consumer-level hardware. I got involved in the early days of computer vision for digital pathology. We suggested there was information contained in the patterns that cells grew under, in routine diagnostic slides, that corresponded with molecular states and, therefore, with the treatments patients should get. Instead of thousands of dollars of testing, we thought we might get to the same result with very inexpensive tools and the assistance of computation.

That base of computational medicine set us up to explore quantum computing. We participated in a Wellcome Leap challenge called Q4Bio, made it through three rounds to the finals, and developed hybrid quantum-classical models to process complex biomedical data. We were actually able to load some cancer data onto quantum hardware in order to build truly quantum-optimized subsets of genetic features associated with clinical outcomes.

At its core, what problem are you actually trying to solve?

When I as a clinician can make a decision with more information, those decisions – and the outcomes that come with them – tend to go better than when I make decisions in an information void. If there are specific biological drivers of a patient’s cancer that we can target, we always want to do that. But biomarkers don’t currently exist in a huge proportion of cancers, and at the same time, we’re struggling with a reality in which the complexity and interaction of the information is rapidly increasing. We have imaging, like pathology and CT scans; molecular information, like mutations and gene expression; and downstream genomic molecular processing, like methylation events. It is very hard for classical methodologies, even cutting-edge ones, to sift through that vast, complexly interacting space in a way that truly represents the underlying cancer drivers I should aim my treatment at.

So did the sheer volume of data push classical computing to its limit?

It’s maybe not the total quantity of data so much as the complexity of it. It’s not necessarily the size of the n. It’s the quantity of the p, the parameter space. The number of different modes of data keeps increasing really quickly, and the parameter space within each mode increases too. We might sequence the transcriptional information of tens of thousands of genes, have information on mutations across thousands of genes, and tens or hundreds of types of mutations within those genes. And we have imaging-based features that might be thousands upon thousands of dimensions long. So it becomes very, very challenging to find features in that complex space that can be reliably associated with clinically relevant outcomes.

What is the quantum subroutine in your pipeline actually doing?

In the ideation phase with our partners, we brainstormed a lot of potential applications, and the one that offered the most discrete, defensible, near-term potential was a subroutine for combinatorial optimization. Meaning, if we could do the “n choose m” combinatorial problem, that would allow us to look over this large space of potential features and select a subset that contained the maximum amount of unique information, which we could then combine, potentially with other features, into a model associated with an outcome. So it’s that combinatorial optimization that was the subregime we were most interested in.

What does a session on the supercomputer look like for your team?

Over the course of the challenge, we didn’t get access to actual quantum hardware until the very end. Prior to that, we used our allocations for the NERSC Perlmutter assets for quantum simulations. We used GPU acceleration to run the simulation of what optimization potential existed across a range of different-complexity experiments, in terms of the number of selected features we were curating in that first-pass, quantum-classical-inspired subroutine.

Your group published one of the early surveys of quantum computing applications in oncology. What did it reveal about where the field actually is?

Our group was one of the first to review the potential of quantum computing specific to cancer applications. The purpose of the paper was to disseminate information as early as possible about the unique potential of quantum computers to solve some challenges intrinsic to the way we think about precision oncology. Part of it was describing what we mean by the quantum physical properties that lead to a quantum computer, how that differs from a classical setup, and then thinking about the applications: drug discovery, computational subset selection for optimization, and potentially housing quantum information derived from quantum sensors directly onto quantum computers in order to maximize the efficiency of information transfer. Those are all topics we think will be relevant to the oncology research and clinical care community in the coming decades.

This year’s SC theme is “HPC Unites.” Your collaboration spans across quantum physicists, computer scientists, oncologists, and genomic data scientists. How difficult was it to get everyone on the same page? Surely there were moments those groups were genuinely not speaking the same language?

That point about language is so important to doing science. Each discipline has its own way of approaching information transfer. For me, as a statistician by training, I’m a data generalist. That has allowed me to collaborate with a lot of different groups, from computer science to molecular engineering to clinical domains to statistics.

And we got really lucky here. When we teamed our group with industry partners at Infleqtion and academic partners here at [UC] Chicago, the genomics lead for the project, Dr. Samantha Riesenfeld, had a background in quantum computing. Those connections across disciplines were transformative for the success of our project. From my perspective, there was still a lot to learn, to understand things like quantum theory notation and the advantages and disadvantages of quantum hardware setups. I was teaching clinical context to computer scientists, and the computer scientists were teaching me the specifics of specific algorithms. One of the things that made our group really successful was the interest in learning from other domains, in order to bridge the gap between what are pretty disparate islands of very intense knowledge acquisition.

Your patients have head and neck cancer. In realistic terms, how close are you to a workflow where what a supercomputer finds about a tumor directly informs a treatment decision?

I think we’re very close, in a variety of different domains. We’re trying to push toward this broader goal of computationally enhanced clinical treatment planning. Our group, and others, have used supercomputing-enabled processing of clinical data, like CT [computed tomography] scan images and digital pathology, combined with research data like genomics, in order to pick which types of treatments would be more or less likely to help a given patient.

We’ve also been using computationally intensive AI tools to make our regular clinical care more efficient. I’m thinking of methods like AI-enabled digital scribing that allows me to talk fluently with a patient in the room, face to face, instead of facing away and typing on a keyboard, and have a real conversation about what their priorities are.

In clinical oncology broadly, we’ve already seen computationally enhanced companion diagnostics, and we’ve now seen tools built on supercomputing infrastructure gain FDA [U.S. Food and Drug Administration] clearance to effectively perform superhuman tasks, meaning inference from a cheap substrate to accomplish what was an expensive genomic test – and that happened just in the last few weeks. These advances show how much information can truly be extracted from existing data infrastructures with the assistance of high-performance computing, effectively getting more out of the same amount of information. As the collection of information gets done more consistently, I think we’re going to see the opportunities get even greater.

There is a lot of talk about an AI bubble. Listening to you, it does not sound like you see one.

I’m a pragmatist about this. My lab builds and fine-tunes foundation models, but when I think about my career in clinical medicine, medicine is a discipline that is all about ‘incremental, incremental, incremental’ gains. That’s kind of the opposite of the bubble mentality, right? 

The way AI is going to permeate clinical medicine, backed by high-performance computing, is by building incrementally more complex and higher-performing models, in a way where even very conservative physicians feel that at every stage, we’re moving in the right direction rather than ceding all of medicine to an AI assistant. We’ll see these incremental benefits happen very quickly, but not that there’s going to be one model to rule all of medicine that arises spontaneously in the next six months.

What do you want the HPC community to understand about what clinical researchers need from national computing infrastructure? 

On national computing infrastructure, there’s a huge opportunity. One thing that’s happening rapidly is that the barriers to entry for utilizing cutting-edge computational models are dropping quickly. The amount of expertise required for code and deployment is falling really fast, so the complexity of the tools an expert clinical researcher could use is expanding really quickly.

If there’s one thing I can say, it’s that having governance structures that allow for the processing of large-scale human health data, and then an interface with infrastructure where code could be run with the assistance of some sort of low-code or chat-based system, would do a lot to speed the development of AI models for routine clinical use. 

At our own institution, just in the last six months, because of the utility of agentic software-development tools, we’ve seen substantial use increase on the local high-performance computing infrastructure. I’d imagine the same thing is happening nationwide. 

Making sure we have national resources to meet the needs of national priorities like human health would be an answer to the call.

Join the Conversation at SC26