By Joe Oakley, MD, medical director of biomarker development, Paige
Just as much of medicine has begun to undergo a digital transformation in the past decade, so has pathology. Now, biomarkers, which are indications of particular features of cancer and are essential to diagnosis and treatment, are beginning their own digital revolution.
This is an exciting advancement, as digital biomarker assays have the potential to work with existing chemistry-based tests or even fill in gaps where chemistry-based tests are either imperfect solutions or don’t exist to give pathologists more information at the point of diagnosis.
This information can be passed on to the oncologist and patient to then guide downstream care. As a result, pathologists, oncologists, and patients will be able to better understand cancers and how they might respond to certain treatments, all to improve the way cancer is treated.
What is a Digital Biomarker Assay?
Today, when we hear the phrase ‘digital biomarker assay’, what most frequently comes to mind are the many assays that are available for standardizing immunohistochemistry (IHC) testing with computer assisted image analysis. However, these are not what we think of as complete digital biomarkers, as the IHC has still done the work identifying the biomarker; the algorithm simply functions to standardize the interpretation of the IHC.
Instead, a true digital biomarker assay is one where the algorithm itself is able to identify the biomarker, such as within a hematoxylin and eosin (H&E)-stained slide, either to complement the information that can be identified with existing testing, or to uncover information that existing testing methodologies cannot find.
What purpose do biomarkers serve in the diagnostic process?
Such information has become critical in diagnosing many cancer types. Traditionally, in making a diagnosis, pathologists look at the morphology and phenotype of a tumor sample to decide if it appears to be malignant or not. Increasingly though, pathologists need ancillary techniques to better understand the cancer they are looking at. For example, in invasive breast cancers, pathologists must look at how estrogen receptors (ER), progesterone receptors (PR), HER2 and potentially Ki-67 are being expressed in the tumor—at a minimum.
This is because the gene alterations that are present, or the presence or level of expression of certain proteins like ER, PR, and HER2 in breast cancer, will have prognostic and therapeutic consequences for how the patient’s cancer is going to be treated. Yet these are not features that can be seen with the naked eye by the pathologist, and thus chemistry-based biomarker assays have become critical.
However, chemistry-based testing requires additional sections of the patient’s tumor tissue, and if there is not enough tissue available, some of the tests may not be able to be done successfully. This is where AI comes in and can transform the results.
Biomarker tests are really looking for “phenotypic” signals in the tumor. The phenotype is how the cancer is using the genes that mutated to cause the cancer. Just as these genes may cause fluctuations in ER, PR and HER2 protein levels, they can change the way the cancer cells look—sometimes in ways too subtle for the human eye, but in a way that AI may detect. Thus, digital biomarker assays can complement chemistry-based biomarker tests by finding the same or better phenotypic signals digitally on existing H&E slides, such that the pathologist can easily interpret the results at the time of diagnosis.
The potential of these assays, as with all pathology AI, is not that they will replace pathologists, but that they can offer more comprehensive information to pathologists based on H&E alone to complement existing practices and tests in new and helpful ways.
How do we train digital biomarker assays?
To make digital biomarker assays a reality, we have to carefully train the algorithms to be able to uncover signals from H&E effectively. This starts with a clinical digital pathology AI module, which has been trained on tens of thousands of slides of a particular form of cancer, as well as its various mimics, to recognize what that cancer looks like. Then we layer in data on what phenotypic features we are hoping to teach the AI to identify. For example, on an invasive breast cancer slide, we would incorporate the HER2 score associated with a slide, so that the AI can learn how to recognize phenotypes that are associated with various levels of HER2 expression.
Once we have trained the model to find these features, we can add other labels for other biomarker targets, or even different chemistry-based assays looking at the same biomarker, like IHC score and in situ hybridization (ISH) or next-generation sequencing (NGS), to get a more comprehensive picture. We can even train with clinical data, such as “did this patient respond to the drug they were treated with or not,” independent of any chemistry-based test, to get an algorithm that can predict clinical response directly, so long as a phenotype for that can be found by the AI on H&E alone.
Digital biomarkers complement existing tests and fill in gaps
This means we can develop different kinds of digital biomarker assays depending on the clinical problem we are trying to solve. First, we can create what we call “replicative digital biomarkers,” which use H&E to complement existing chemistry-based tests. For example, an AI algorithm might recognize features of bladder cancers on H&E that suggest a strong likelihood that a gene called FGFR has been altered in that tumor.
Armed with this knowledge from the digital biomarker assay, the patient and their treatment team can send that tumor on to chemistry-based genomic testing to confirm which FGFR alteration was present, and then select the right therapy. Patients with bladder cancers who do not have that FGFR alteration signal detected by the AI algorithm, on the other hand, would not need to sacrifice the time, tissue, or money to test for an FGFR alteration that isn’t in their tumor—they can instead go on to a different kind of therapy altogether, one that would be more appropriate for their cancer.
Another form of digital biomarker assay we can create is what we call “novel digital biomarkers,” which create a test for that which cannot be found using current chemistry-based methods or for which current chemistry-based tests are significantly limited.
Novel biomarkers pose even more exciting benefits. They can be used to help pathologists more confidently identify biomarkers where there are gaps in current testing methods and can allow them to identify subsets of patients that may potentially respond better to available treatments from the moment the cancer is diagnosed on H&E.
HER2Complete: A case study
The potential for novel biomarker assays to transform the treatment of HER2 expressing breast cancers is a great example. Historically, about 15% of breast cancer patients were believed to be HER2-positive.1 Recent studies revealed however that there is a subset of patients who do have HER2 expression within what was previously considered the HER2-negative group, now called HER2-low.2
This is important because new drugs have come to the market that are able to treat this HER2-low group, such that an additional 51% of breast cancer patients could now benefit from HER2 targeting treatments.1,2 The problem is that current IHC testing was designed for the previous generation of HER2 treatment, and the cutoff between what is truly HER2-negative and what might now be considered HER2-low is subjective and hard to define with IHC.
Moreover, there may be patients in the IHC-0 group that actually do have enough HER2 to potentially benefit from the new drugs. The challenge though is that at a score of IHC 0, IHC is not sensitive enough to identify them with confidence. So, we are developing Paige HER2Complete*, the first and only H&E based digital biomarker capable of detecting morphological phenotypes consistent with HER2 expression in IHC negative or equivocal-IHC1+ samples. Our hypothesis is that with support from future clinical trials, this algorithm could enable pathologists, oncologists and patients to better classify HER2 expression and select the appropriate HER2 based therapy from the options currently available.
Looking ahead: What digital biomarkers could mean for the future
Clinical trials are a very important part of helping new digital biomarkers translate into actual clinical practice and outcomes. Developing evidence to not only validate the algorithms, but to set the stage for regulatory clearance, ensures they are safe and effective ways of supporting pathologists and oncologists, and ultimately helping patients.
Not only that, but with digital assays being less labor intensive and costly than chemistry-based testing such as NGS, IHC or FISH, these tests can become more accessible to underserved areas where specialists are rare or do not exist, expanding patient care and creating a landscape in which every patient, in all geographies, receives the best possible care.
1 Tarantino, P. et al., (2020) J Clin Oncol. 10;38(17):1951-1962.
2 Modi S, Jacot W, Yamashita T, et al; DESTINY-Breast04 Trial Investigators.
Trastuzumab deruxtecan in previously treated HER2-low advanced breast cancer. N Engl J Med. 2022;387(1):9-20.
*In the European Union and U.K., HER2Complete is CE-IVD marked for clinical use with Leica Aperio AT2 Scanner. In the U.S. and where research use is permitted, HER2Complete is limited to Research Use Only and not for use in diagnostic procedures.
Joe Oakley, M.D. is medical director of biomarker development at Paige,the first company to receive FDA approval for an AI product in digital pathology. He is a pathologist with board certification in anatomic, clinical and molecular genetic pathology. His work history ranges from academic clinical practice to the pharmaceutical and information technology industries.