Latest News

AI Predicts Cancer Treatment Response Directly From Tissue Slides

By Deborah Borfitz 

August 20, 2024 | New technology, powered by artificial intelligence (AI), aims to bring breast and ovarian cancer patients the expected value of genomic sequencing—a more precise diagnosis and personalized treatment strategy—potentially without ever having their DNA analyzed. At present, relatively few patients globally are having their tumors sequenced for genomic biomarkers, which serve to guide doctors toward optimal first-line treatment for their specific cancer, and those who are getting sequenced might spend thousands of dollars and wait weeks for results. 

So says Erik Bergstrom, Ph.D., a postdoctoral researcher in the lab of Ludmil Alexandrov, Ph.D., professor of bioengineering and professor of cellular and molecular medicine at the University of California San Diego (UCSD). Bergstrom is lead author on a recent study describing an AI protocol for rapid, low-cost detection of clinically actionable genomic alterations directly from routine digitalized tumor biopsy slides (Journal of Clinical Oncology, DOI: 10.1200/JCO.23.0264). 

The algorithm, called DeepHRD, was trained to read hematoxylin and eosin (H&E)–stained histopathological slides “like a pathologist” to identify homologous recombination deficiency (HRD), a known biomarker for breast and ovarian cancers that indicates a higher chance for a positive response to treatment with platinum chemotherapies and PARPi (poly-ADP ribose polymerase inhibitors), he explains. Just over 1,000 breast cancers and 450 ovarian cancers tumors from The Cancer Genome Atlas were used for the training exercise. 

DeepHRD was then compared with four standard HRD molecular tests using 349 breast and 141 ovarian cancers derived from four independent data sets. The robustness of the HRD prediction was demonstrated across multiple external cohorts, slide scanners, and tissue fixation variables, which are common confounders in this setting. Relative to molecular testing, the platform classified two to three times more patients with the genomic marker. 

The hope is that by making biomarker testing more accessible using routinely sampled data, more patients will start benefiting from precision therapy, says Bergstrom. DeepHRD can serve as a complement to traditional genomic testing to circumvent the often-slow turnaround time and prohibitive costs by offering relatively instantaneous predictions for the HRD biomarker upon obtaining a tumor biopsy.  

DeepHRD might also be a stand-in for genomic sequencing in resource-limited settings around the world, including certain regions and care facilities in the U.S. where testing equipment and accessibility to sequencing protocols are not readily available, he points out. Tissue biopsies, however, are routinely sampled for each patient undergoing a cancer diagnosis.  

One day, DeepHRD is expected to be routinely deployed on a standard computer via the cloud to enable oncologists anywhere in the world to almost instantly evaluate the HRD biomarker readout on whether platinum and PARPi treatments should be prescribed or guide them toward additional genomic testing. “Our argument is that for a large portion of these patients, we won’t need to do downstream molecular testing anymore,” Bergstrom says. 

To get the platform into clinical use, the DeepHRD technology developed at UCSD was licensed to io9 LLC in 2021. Alexandrov is one of the company founders and Bergstrom serves in a consulting capacity. 

Addressing the Bottlenecks

Other companies are attempting to bring AI tissue technology to the clinic for both diagnostic and prognostic evaluation of a cancer, says Bergstrom. But we wanted to go beyond this evaluation to determine whether we could detect a genomic biomarker from a tissue slide while also determining likely response to downstream treatment.  

The problems being solved here are the accessibility and financial bottlenecks in using genomic information at the point of care, says Bergstrom. DeepHRD explores a routine data point available in the cancer diagnosis setting to “capture complex, hidden patterns [within digital images] that the human eye isn’t capable of detecting.”  

As described by Bergstrom, DeepHRD is built off the backbone of standard deep learning architectures known as convolutional neural networks that get pieced together like Lego blocks. The algorithm first looks at a whole slide image holistically to find morphologies and structures in the tissue to arrive at a prediction, or probability, that a person has the HRD biomarker. The model then automatically zooms in on specific regions of interest in greater detail at the cellular level to make secondary predictions, and the scores get converted into a final composite prediction.  

DeepHRD was found to have a negligible failure rate, which is both remarkable but unsurprising given how AI “parses across the tissue identifying complex patterns of each tumor,” he continues. Genomic testing fails 20% to 30% of the time, necessitating re-testing and possibly a second biopsy, because of widespread disagreement about how to properly classify a patient with the deficiency.  

Path to the Clinic

Bergstrom says that the practical motivation behind the project is a desire to “help patients on a day-to-day basis” with what is possible today rather than waiting for genomic testing to live up to its promise in another decade or two. Next steps include deploying the algorithm on tissue samples from a larger cohort of breast and ovarian cancer patients to ensure it can reliably classify those with HRD. 

Approval of DeepHRD through the Software as a Medical Device pathway of the U.S. Food and Drug Administration is the goal, he says. That will require prospective clinical trials with new cancer patients where the algorithm provides recommendations as to whether they should be given platinum or PARP therapies to see if they have better overall outcomes.  

It could later be tested for its utility in classifying prostate and pancreatic cancers, where HRD has also been implicated. The chief holdup is lack of available data, he adds. 

For the latest study, retrospective clinical samples came from a collaborating hospital in France. Additional partnerships with hospitals and larger institutions are actively being sought to enable follow-up studies, says Bergstrom. 

The same technology could, in theory, be applied to most other genomic biomarkers and many forms of cancer, he adds. Importantly, the required datasets will need to cover both biomarker positive and biomarker negative patient groups along with their respective treatment information, which is hard to come by.  

Load more comments
comment-avatar