banner of LMP department's DP&AI initiative

Giubellino’s team deploys AI in a hunt for tumor-killing cell biomarkers

|
William Hoffman
8 minutes
Patients will benefit as pathology enters a new era

Digital pathology and artificial intelligence imageWith the risk of wildfire outbreaks on the upswing due to changing regional weather and climate patterns, artificial intelligence (AI) analysis of mapped heat and smoke data is being deployed as an early detection tool.  Tissue microenvironments in the human body also serve as potential hotspots -- for the emergence of disease and resistance to treatment.  Could AI be deployed to predict the likelihood of cancer treatment success based on the interaction of cells in those hotspots?

Alessio Giubellino photo

LMP associate professor Alessio Giubellino and his research team aim to answer that question with their digital pathology project “Development and cross-validation of digital pathology assessment of tumor infiltrating lymphocytes (TILs) as a laboratory developed test (LDT) for clinical implementation.”  If they succeed, it will be because they have used digital pathology tools including AI to identify biomarkers in specimens from tumor tissue microenvironments that predict whether a patient will have a lasting response to cancer immunotherapy and incorporated those biomarkers into a laboratory test.  Knowing in advance who is and who is not likely to benefit from the new immunotherapies is what precision and personalized medicine is all about. 

“Our goal is to design and validate a biomarker-based test for solid tissue tumors that is useful in the clinic,” Giubellino said.   By “useful” he means a test that predicts whether or not immunotherapeutic drugs called immune checkpoint inhibitors (ICIs) will be effective, leading to a durable patient response. 

Giubellino was a research fellow at the National Institutes of Health when the U.S. Food and Drug Administration (FDA) approved the first in a growing class of ICI drugs in 2011.  The approval was for treating deadly metastatic melanoma, Giubellino’s research interest as a dermatopathologist.  Some patients did experience a durable response.  Many did not.  Why not?  And would it be possible to know in advance who would be likely to respond so that the side effects of ICI treatment could be avoided?

In 2012, the year after the FDA approved the use of the ICI monoclonal antibody drug ipilimumab (Yervoy®) to treat metastatic melanoma, University of Toronto computer scientists published a report on a convolutional neural network they had developed called AlexNet.  It’s often described as a “defining moment” for the use of deep neural nets for image recognition – for what we call today machine learning and artificial intelligence. One of the authors of the report was Geoffrey Hinton.  The “godfather of AI,” Hinton was awarded the 2024 Nobel Prize in Physics. 

Giubellino’s team is employing cutting-edge AI tracking tools just now available for biomedical research to find out how cutting-edge ICI immunotherapies work and in whom.  They are bringing digital pathology and AI to bear on the microenvironmental “hotspots” of eight tumors that take the lives of hundreds of thousands of Americans every year:  melanoma, colorectal cancer, bladder cancer, cervical cancer, breast cancer, head and neck cancer, glioblastoma (the most common brain cancer), and non-small cell lung cancer (by far the most common form of lung cancer).  The answers they seek reside in small tumor tissue compartments where cancer cells come under attack by arguably the most protective surveillance system on Earth:  The human immune system.

Lymphocyctes infiltrating tumor precincts

Key to success for Giubellino and his colleagues is to harness remarkable advances in computer vision and vision transformers for image recognition to gauge the number and spatial distribution of tumor infiltrating lymphocytes (TILs) in solid tumor tissue.  TILs are immune cells that reside in tumors or infiltrate tumors from the blood system.   They do recognize and kill cancer cells, but their effectiveness is highly varied, often haphazard, and almost always incomplete, as reflected in the response of many patients to ICI drug treatment.

For four decades, researchers have been working on ways to spur TILs into launching a comprehensive, sustained assault by removing them from tumor tissue, expanding and potentiating them in the laboratory, and reinfusing them into the patient.  Last year, the FDA granted an accelerated approval to a TILs therapy for patients with metastatic melanoma who experienced progressive disease following ICI drug treatment.  “TILs hold great promise for major improvements in our ability to cure cancer,” said the National Cancer Institute’s Steven Rosenberg who pioneered “adoptive cell therapies” in the 1980s, insisting that the cell therapy approach offers curative potential for solid tumors. But can the secrets of TILs be teased out with powerful new molecular and computational tools?  Giubellino thinks they can.

“Since ICI therapy is directed to TILs – that’s its direct target, to release the brakes that cancer cells put on TILs -- it makes sense to look at TILs for biomarker potential for response to therapy,” Giubellino said.  A much better picture of tumor dynamics in the “hotspots” is needed.  “Determining the density in numbers and the spatial distribution of TILs in the tumor microenvironment are the main goals of our project.”   Determining the functional state of the tumor infiltrating lymphocytes through gene expression profiling, though “super interesting,” would be a “downstream experiment,” the next step.

Giubellino and his team focus initially on what he calls “regions of interest” – the tumor microenvironment, the cells within, and elements and structures within cells, most notably the cell nucleus.  “The reason we are focusing on the nucleus in our research is because it is the most reliable and identifiable object within the cells.  Unlike, say, the cell membrane, which is more difficult to define, the nucleus is clear, it’s defined.  So we can identify it and extract different parameters from the nucleus.” 

Besides holding the cell’s code for its reproduction, it just may be that the nucleus also harbors distinguishing features that algorithms can mine to identify, say, cell type, perhaps whether a cell is a tumor infiltrating lymphocyte and how it interacts with its neighbors.   “We are really aggressive in using the new tools,” Giubellino said.  The software tools they are using have names like DeepImageJ, ResNet50, and Hover-Net. The open-source pathology software QuPath is “a great resource” for bioimage analysis, particularly in its capacity to export images in a format that can be read by other software, he said “Down the pipeline of image analysis, we can use these images and get different information from different software.”   

Giubellino thinks the soon-to-be-implemented HALO AI will be a critical addition.  Deep learning networks enable HALO AI to train a classifier or machine learning algorithm “to quantify tissue classes, to find rare events or cells in tissues, or to categorize cell populations into specific phenotypes,” according to promotional literature.

TILs enumeration and spatial distribution can be done at a relatively low cost using digital technologies due to the unprecedented power and fidelity of what Hinton and his colleagues in the field of machine learning and AI have brought to biomedical research.  “For us, it’s kind of a black box,” Giubellino said.  “An initiative like this one will help us to enter inside the black box and understand how things work, how we can use these tools, because when we understand how to use these tools we can implement them in the clinical workflow.”

The imaging approach gives investigators the best chance of near-term implementation in a laboratory diagnostic test, he said.

Time for a TILs LDT

The FDA defines laboratory developed tests, or LDTs, as “in vitro diagnostic products (IVDs) that are intended for clinical use and designed, manufactured, and used within a single clinical laboratory that is certified under the Clinical Laboratory Improvement Amendments of 1988 (CLIA) and meets the regulatory requirements under CLIA to perform high complexity testing.”  Last spring, a federal judge quashed the FDA’s plan to regulate LDTs and the FDA declined to appeal the judgment.  The biomedical research community generally takes the position that LDTs should be given wide latitude. 

“Those tests are developed locally, because it’s difficult to develop them in a broader way,” Giubellino said.  “There are LDTs that we are currently using, but there are none for TILs that I know of.  For implementation of TILs enumeration and spatial distribution within our effort, we need to have a test that is certified, reliable, and reproducible. We will need to look to many samples to validate our approach.” 

It would be important to link the LDT developed for TILs to Current Procedural Terminology (CPT) codes so that their use can be paid for, Giubellino said.   “Hopefully in the future these digital pathology CPT codes will be monetizable. For that, the test it will have to be validated and certified.  I think an LDT is the way to go.”

Successful validation of their imaging approach to understanding TILs dynamics in the tumor microenvironment and identifying biomarkers for an LDT will require a larger number of samples.  That means collaborating with researchers in other groups in other institutions.  “The idea is to leverage many collaborations that every one of us has with other institutions and hopefully get images of those tumors, which will strengthen our validation and implementation efforts,” Giubellino said, adding that he has ongoing collaborations with MD Anderson in Texas, Ohio State University, and New York University. 

The ultimate goal – it’s not something that’s going to happen tomorrow – is to put this information into the pathology report, he said.  “The idea is that the pathology report of the future will have some prognostic, predictive information that is actionable we can give to the clinician.  We are trying to give to the clinician more tools to manage their patients and improve patient care.”

“Development and cross-validation of digital pathology assessment of tumor-infiltrating lymphocytes (TILs) as a laboratory developed test (LDT) for clinical implementation.” 

PI: Alessio Giubellino, MD, PhD

Collaborators:

  • Shijia Zhu, PhD (Computational research faculty)
  • Rick Jansen, PhD (Biostatistician – MCC Biostatistics Core)
  • Brian Bagley, PhD (Digital Imaging Specialist)
  • Vidhyalakshmi Ramesh, MS (CTSI)
  • Andrew Nelson, MD PhD
  • Khalid Amin, MD
  • Diana Oramas-Mogrovejo, MD
  • Emil Racila, MD
  • Kevin Turner, DO
  • Liam Chen, MD, PhD
  • Mahmoud Khalifa, MD, PhD
  • Molly Klein, MD
  • Paari Murugan, MD
  • Ahmed Jamshed, MBBS (Resident, 3rd year)

  • Dina El-Rayes, MBBS (Resident, 4th year)

Stories in our DP & AI series: