New Frontiers

Fighting cancer with AI

Mention artificial intelligence, and most people will immediately think of the publicly accessible tools whose ability to generate passable term papers, uncanny artwork, and runnable computer code has helped ignite everything from internet memes to union strikes. Behind the broad discussion of AI’s role in society, though, scientists and clinicians have been quietly exploring its potential to revolutionize cancer care and research.

It’s not a theoretical exercise. The US FDA has already approved several AI-based tools for clinical use, especially in cancer screening, diagnosis, and monitoring. At the Herbert Irving Comprehensive Cancer Center (HICCC), investigators are already developing and testing the next generation of cancer-focused AI tools, in a research pipeline extending from basic science to current clinical practice.

New Frontiers

Picture the Future

The hottest area for clinical AI use is imaging. “Radiology is leading the way on the translation of AI tools into the clinic, and here at Columbia, we’re trying to capitalize on that momentum to really make an impact,” says Despina Kontos, PhD.
Kontos led the development of new digital breast tomosynthesis biomarkers for breast cancer risk assessment. Digital breast tomosynthesis uses AI to construct a 3D mammogram from multiple images of the breast, providing much greater detail than a standard mammogram. Now, as the head of the new Center for Innovation in Imaging Biomarkers and Integrated Diagnostics, she hopes to grow Columbia into a major hub for moving AI tools into the clinic.

To do that, she’s tackling some of the biggest challenges facing wider adoption of the technology. For example, as breast cancer screening guidelines have changed to call for more frequent mammograms and additional imaging techniques, screening costs have skyrocketed, and false-positive results have become more common. Kontos wants to develop risk assessment algorithms that can optimize each patient’s screening schedule, improving care while also decreasing costs.

However, developing and validating those algorithms won’t be easy. “We have the ideas, in principle, and the population data, but it takes forever to figure out how to organize the data, load them on a computing environment, store them, analyze them, and share them between institutions,” says Kontos, adding “that’s what stops us from making the breakthrough contributions that can lead to change in patient care.”

With her background in computer science, Kontos is well placed to address that. One idea she’s pursuing is a federated learning model, using data from multiple institutions. In this approach, researchers bring the algorithm to the data, instead of moving massive amounts of data to the algorithm. Each institution trains an experimental algorithm locally on their own data, then the team aggregates the results into a single model. Kontos has already demonstrated the feasibility of that strategy for neural networks that can estimate breast density, and she hopes to extend it to other areas as well.

Despina Kontos, PhD

  • Herbert and Florence Irving Professor of Radiological Sciences (in Radiology and the Herbert Irving Comprehensive Cancer Center)
  • Professor of Biomedical Informatics and Biomedical Engineering
  • Vice-Chair of AI and Data Science Research, Department of Radiology
  • Director, Center for Innovation in Imaging Biomarkers and Integrated Diagnostics (CIMBID)
  • Chief Research Information Officer (CRIO), Columbia University Irving Medical Center (CUIMC)
  • Director of Biomarker Imaging, New York-Presbyterian Hospital (NYP)
  • Member, Cancer Population Sciences Program, HICCC

Ready, aim, treat

In addition to diagnosing cancer, AI is making waves in radiation therapy. These days, that often means precisely aiming the beam of a linear accelerator into a patient’s body, delivering high doses of radiation to a tumor while sparing surrounding tissues. Unfortunately, many tumors present moving targets.

“With patients who have cancer, the first thing we do after consultation is a planning session, which we call a CT simulation,” says David Horowitz, associate professor of radiation oncology at Vagelos College of Physicians & Surgeons (VP&S) and member of the HICCC. That produces a 3D model of the tumor and surrounding tissue, based on a computed tomography (CT) scan.

After the initial scan, it typically takes the clinical team several days to develop a personalized 3D radiation plan, which is then used to guide an entire series of radiation treatments. “Patients can be treated for weeks…sometimes up to two months with this original plan,” says Horowitz. In the meantime, the tumor has likely shrunk, and flexible internal organs have probably moved, causing the treatment to miss its target and strike healthy tissue instead. The problem is especially severe for highly flexible organs, such as those in the digestive tract.

Using a cutting-edge AI-powered imaging system called EthosTM, Horowitz’s team can scan a patient and construct a detailed 3D model of their cancer in a matter of minutes. That enables them to proceed directly from imaging to treatment, and to image the tumor before each round of radiation, giving them a real-time view of the tumor that leads to more precise treatments. Columbia is one of the leading centers nationwide for this type of treatment, which is already proving superior to traditional methods.

That approach is especially promising against notoriously treatment-resistant pancreatic cancer. “If you can achieve very high doses of conformal radiation, you can potentially ablate the local pancreatic tumor, but you have to be mindful of the normal organs that surround the cancer,” says Lisa Kachnic, MD, FASTRO, chair of radiation oncology and associate director for radiation oncology and science at the Herbert Irving Comprehensive Cancer Center (HICCC) .

She adds that “the beauty of the EthosTM is that we can see on a day-to-day basis where those organs are, and the AI algorithm makes sure they’re accounted for properly, so we’re able to really give the highest and safest dose to those pancreatic cancers.”

Photo: Adaptive radiotherapy can provide an enhanced targeted dose directly to the tumor, (top left image), as well as sparing the surrounding tissue (bottom left image) in comparison with conventional radiotherapy (top and bottom right images).
New Frontiers

Biology’s next top model

Other HICCC members are laying the basic science groundwork for the cancer therapies of the future. Mohammed AlQuraishi, PhD, is one of those taking the long view. He wants to build a high-fidelity digital model of a human or animal cell that can simulate any measurable biological process, including cancer development, by 2050.
“We use artificial intelligence to model molecules, protein structure, protein dynamics, interactions of proteins with nucleic acids, drugs, antibodies, and so on,” says AlQuraishi. The idea is to combine models of individual molecular activities to recapitulate the behavior of a complex cell, which would allow researchers to find or test potential therapies in a computer model, vastly accelerating drug discovery.

AlQuraishi’s team is developing open-source algorithms to predict the 3D structures of proteins based only on their amino acid sequences. That’s been a longstanding problem in biology, but in recent years an effort funded by tech giant Google’s DeepMind subsidiary has made major strides in solving it.

“DeepMind’s AlphaFold changed the field considerably, and the latest version, AlphaFold 3, is quite capable,” says AlQuraishi. It’s a powerful tool for studying the molecular mechanisms at play in cancer cells, and how specific drugs can affect them. DeepMind’s engineers won the 2024 Nobel Prize in chemistry for their work in developing AlphaFold 3.

DeepMind made AlphaFold 3 open source for academics in November 2024, a move that AlQuraishi applauds. “The availability of the AlphaFold 3 source code is an important milestone, particularly for academic labs that can now access state-of-the-art protein structure prediction tools without financial barriers. It will also allow for direct assessment and comparison of AlphaFold 3 against other methods, which was not possible before.” he says.

Building on this foundation, AlQuraishi’s lab is advancing OpenFold, an open-source alternative designed to address current limitations and expand accessibility, including to industry. AlQuraishi expects OpenFold3 to be the most performant structure prediction system in the field when it is released.

This new iteration of OpenFold will allow drug companies to test a protein target and obtain a prediction of what the ‘pocket’—where the drug could bind—looks like. This is typically a very laborious and costly process.

“The intent is to make this widely available, including in industry, which is not possible today. I would hope that really moves the needle in drug discovery, and that is something that is going to happen in the next few months,” says AlQuraishi.

Mohammed AlQuraishi, PhD

  • Assistant Professor of Systems Biology (in Computer Science), Vagelos College of Physicians and Surgeons
  • Member, Precision Oncology and Systems Biology Program, HICCC
Back to Report
Contents