Many pathologists have spent their entire careers examining slides of biopsies from cancer patients, honing in on their craft to spot a diagnosis. But these days, you can’t go to a precision medicine conference without hearing about how artificial intelligence (AI) threatens their work and the future of humans in pathology. Now, pathologists may be getting some of the power back into their own hands with the help of what has been considered their biggest enemy.
A new study in the Journal of Clinical Oncology demonstrates how AI-backed software can diagnose not only the cancer type and stage but also the implicated biomarker and the medicine they’re most likely to respond to by analyzing a digital slide. Researchers at the University of California, San Diego (UCSD) created a deep learning platform trained on histopathology slides to identify homologous recombination deficiency (HRD) in breast and ovarian cancer and, with it, whether the patients will respond to treatment. An AI model called DeepHRD can find breast and ovarian cancer patients who are likely to respond to poly(ADP-ribose) polymerase (PARP) inhibitors or platinum therapy, the two drug classes that have been shown to target HRD. Even more impressively, it can do this faster, more accurately, and for less money than standard genomic sequencing.
Greg Hamilton, CEO of io9, founded by senior authors Scott M. Lippman, MD, and Ludmil B. Alexandrov, PhD, and licensed DeepHRD from UCSD, said, “We can detect a genomic biomarker on a digital image of a pathology slide. [Alexandrov’s] team developed the concept that we can identify a biomarker while diagnosing cancer. Using the same pathology report, we can inform the oncologist that, yes, the patient is cancer-positive at this stage, and she is either HRD-positive or negative. Now, the oncologist and patient can begin developing a treatment plan immediately. Ultimately, that is a massive breakthrough in precision medicine.”
Cancer is a battle against time
Significant advances have been made in the development of effective precision oncology treatments. The biggest challenge is finding out which patients will benefit most from them. Many methods for identifying responsive patients are NGS tests that require pathologists to cut a piece of the tumor and send it to a third-party lab for sequencing. In the process, the patients and their providers, the oncologists, must wait weeks for these results.
“At the end of the day, cancer is a battle against time, and right now, we have extremely long lead times,” said Hamilton. “Sometimes, when you send a sample, you do not get a result because there is insufficient DNA. As a physician, you are still in the dark, and you ask yourself, ‘Will I sit here for three weeks waiting for my metastatic breast cancer patient to do nothing, or will I treat her?’”
And NGS solutions come with their own problems. Of note is that sequencing data can even obscure clinical interpretation and how to follow up with treatment. Hamilton said, “Sequencing is great, but what if I have a single copy of this mutation? How do I know if that’s clinically relevant to the cancer?”
This is especially important for HRD because it can be caused by differing germline forms, cell changes, and problems with how BRCA1, BRCA2, and other homologous recombination pathway genes, like PALB2 and RAD51, work. Additionally, specific patterns of genomic instability, gene expression, and somatic mutations are observed in cancers that harbor HRD.
Not surprisingly, all the approaches for detecting HRD in the clinic rely on molecular profiling. So far, the FDA has approved two HRD companion diagnostic (CDx) tests for patients with ovarian and breast cancers: Myriad myChoice CDx and FoundationOne CDx, which detect HRD by measuring genomic instability in conjunction with BRCA1/2 status. Multiple research and Clinical Laboratory Improvement Amendments (CLIA)-certified tests have been developed and used to detect HRD in breast and ovarian cancers.
Predicting genetic biomarkers from slides
To develop software that can tell the difference between a responder and a non-responder, Alexandrov’s team trained a deep-learning platform using pathology slides from breast cancer patients, the genomic data for those patients, and information about how well the patients responded to PARP inhibitors or platinum therapy. Alexandrov’s team trained the model in the context of breast cancer to detect HRD, which was validated on entirely independent data sets.
“We got a completely independent set from France, where we had treatment information, the digital slides, and data from a number of the other tests being performed,” said Alexandrov. “What we can see for breast cancer is that we capture significantly more patients, about two times more, than the molecular tests, and these patients have a better response [to treatment].”
This led Alexandrov and his team to ask whether their platform could be applied to other cancers. To test this, Alexandrov’s team then transferred the learned knowledge from breast cancer to ovarian cancer. After a similar training period, the platform was tested on another independent cohort from Memorial Sloan Kettering Cancer Center, where they run an FDA-approved targeted tumor sequencing test called MSK-IMPACT. Alexandrov’s team found that DeepHRD could detect three times as many patients as MSK-IMPACT responsive to ovarian cancer treatment, and there were significant differences in survival rates.
“On the companion diagnostic front, we believe this is pretty revolutionary,” Hamilton stated. “The concept of having biomarker information at the same time of diagnosis will benefit many people, not only in terms of speed and time to get them on the right frontline therapy but also for personal management. A woman who was recently diagnosed with triple-negative breast cancer must wait three weeks to find out how she will be treated. That has to be extremely painful.”
Bringing equity and accessibility to precision medicine
Hamilton said this work on DeepHRD has significant implications regarding the equity of precision medicine. After all, not many people or insurance companies like paying $3,000 for a test to determine what therapy to prescribe. But if the analysis of a digital image can provide the diagnosis and relevant biomarkers, that should eliminate financial.
“We can take an H&E slide image from anywhere in the world; it is a cloud-based solution,” Hamilton explained. “You do not have to ship samples to Utah, Myriad, or Boston. The foundation has eliminated all those barriers, and we now do H&E slides everywhere in the world to diagnose cancer, whether in Brazil, Japan, or France. “We all diagnose cancer using H&E slides.”
Another issue that io9 has been investigating is whether image or scanner quality matters. Ideally, OncoGaze would be so dependable that any scanner—including a smartphone—could use any digital image.
“Yes, we have used an iPhone through a microscope to take a picture and were able to use it,” Hamilton said. “We are technology agnostic, which is what we want to be because otherwise you are tied to a specific scanner and ordered the same issue that we see all over the world: if you want to do it in France, they are not going to go out and buy a five hundred thousand dollar scanner.”
Giving power back to the pathologist
The looming question has been whether pathologists will abandon their years of chiseling and fine-tuning diagnostic imaging procedures in exchange for software. There is a growing belief that software will eventually outperform humans in identifying microscopic objects. Indeed, there is a wealth of literature on pathological variability, cancer detection, and classification. As with anything else, software is more consistent and accurate once properly trained than humans. It is simply the reality of the world.
Hamilton stated that companies such as Paige and PathAI have paved the way for digital pathology in cancer diagnosis. However, they have both struggled because their solution involves putting people out of business. Many pathologists are not particularly interested in software, allowing them to come in and do their job. However, Hamilton insists that what io9 is doing is very different from what the digital pathology companies do.
“We are not attempting to diagnose the cancer—Paige, PathAI, and other companies are well on their way, and they can do it successfully,” said Hamilton. “What we are doing is very different: we can detect a biomarker, which is much more difficult. Then, we identify biomarkers already in clinical use, so we do not need to prove their clinical utility. We need to say that we can do it from a digital image, which is faster, cheaper, and, in the case of HRD, superior to sequence. Our business model is to go to pathologists and help them do something they cannot do right now.”
Instead of paying thousands of dollars for an NGS-based companion diagnostic test right off the bat, Hamilton thinks that with io9’s cloud-based software, the pathologist can identify the biomarker, include it in their path report, and apply for reimbursement. This is a new service they can provide to their oncologists and help them remain in and even move up in the clinical chain.
Hamilton also said, “We have a technology that can predict HRD from digital slides, and when you compare the response to treatment in breast and ovarian cancer, we can show that we capture a larger population group, and they respond better than if you were to use the molecular test for this patient. Just as sequencing has evolved, where we sequence every patient because we want that information for future target identification, we will also use digital pathology for every solid tumor trial in the future because we will want this information for future development as well. It’s going to become a standard.”
However, for DeepHDR to become the norm, researchers must test it on far greater numbers of patients. DeepHDR was only tested on 77 breast cancer patients and 141 ovarian cancer patients. Those numbers are too small for the FDA to seriously consider. But it’s a good start for what could be a paradigm-shifting methodology that many would welcome, including the pathologist who has resisted the move toward digital pathology.
