NEW YORK – Digital pathology firm Paige is hoping to fill an emerging need for assays that can detect individuals with low levels of HER2 expression that can be missed by existing technologies, but who still stand to benefit from treatment with a new generation of anti-HER2 therapies.
At the American Society of Clinical Oncology meeting in June, oncologists were wowed by new data showing for the first time that metastatic breast cancer patients with low HER2-expressing tumors, a group that is currently ineligible for HER2-targeted treatment, significantly benefited from AstraZeneca and Daiichi Sankyo's antibody-drug conjugate Enhertu (trastuzumab deruxtecan).
On the heels of this news Paige announced in late June that it had received CE-IVD and UK Conformity Assessed designations for a new, artificial intelligence-driven digital pathology assay called HER2Complete, intended to identify patients with breast cancers that express HER2, but at low enough levels that they might otherwise be missed by standard-of-care tests.
Jill Stefanelli, Paige's president and chief business officer, said that the regulatory nods will help Paige connect with partners in Europe for the clinical utility studies necessary to definitively link its test to patient treatment outcomes.
Because the firm's technology does not require actual samples — only their images — she added that this could be a relatively rapid process. "Once there's a digital picture of the FFPE slide, we can run our algorithm … so for all of the patients who have been treated on various [Enhertu] trials … if we can partner with those investigators and with the drug companies, we could potentially look at what we would call a retrospective prospective study," she said.
According to some oncologists at the ASCO meeting last month, the new data for Enhertu is strong enough that doctors treating women with advanced breast cancers should consider an immediate change in practice. The drug's sponsors also said they plan to waste no time in bringing the new data to regulators, receiving a confirmation from the European Medicines Agency late last month that their application to that body is now under review.
But clinicians face a challenge in that current test technologies don't offer a reliable or standardized way to identify HER2-low women in order to treat them with Enhertu, let alone maximize the population eligible for the drug.
In the trial reported at ASCO, patients were defined as HER2-low based on having IHC 1+ results, or IHC 2+ and HER2 gene amplification negativity by in situ hybridization. This group, along with those who are IHC 0, currently receives palliative single-agent chemotherapy.
Under this schema, pathologists could certainly identify some HER2-low women eligible for treatment, but evidence suggests many could be missed. Research has shown, for example, that labs vary widely in their results when performing IHC-based HER2 testing, especially when tasked with defining precise scores rather than classifying samples into a simple HER2-high-or-not binary.
Paige is one of several companies now employing machine learning and other artificial intelligence techniques to produce novel disease classifiers based on digital pathology images.
In contrast to traditional immunohistochemistry tests, the firm's HER2Complete is designed to assess protein and mRNA levels of HER2 based on analysis of digital images of hematoxylin and eosin (H&E) stained tissue sections, a pathological mainstay.
Not only does the approach stand to offer clinicians a robust, non-subjective alternative to IHC for catching the low-expressing women who were defined in the trial, but Stefanelli said it would also identify additional women who pathologists currently classify as IHC 0 who nevertheless are expressing the protein.
Paige boasts a lead position in the race to bring AI-based digital pathology to the clinic, having received the first US Food and Drug Administration authorization for a test of this kind in prostate cancer last year.
HER2 is a more recent target for the company, according to Stefanelli. "We decided to make it a challenge, so we [proposed] what we thought we could do with the data that we had access to, and we basically said to the team … if we give you 30 days, do you think that you can have a proof of concept … [and] they said yes."
Paige's development process is reflective of most AI methods, which involve using known samples, or in this case sample images, to train a computer program to then differentiate future unknown samples. But Stefanelli said Paige has several advantages over other groups exploring this space, including access to very large amounts of data and partnerships with medical experts that help it avoid machine learning pitfalls.
"It's a collaboration in terms of making sure we know what are the things that are impactful for the disease … and what are the confounding factors that need to be removed [so we can] foresee what could be a potential bias in the data," she said.
One thing Stefanelli said was a key realization for Paige, for example, was the fact that current antibodies used for IHC and other HER2 test technologies have been designed to detect what the field previously considered HER2 positivity.
"They're really not optimized at all for HER2-low," she said. "We understood that very early, and so we took a completely different approach where we leveraged both existing IHC, but also messenger RNA, because if you think about the biology of a protein, you need the RNA to be expressed."
According to Stefanelli, the firm's immediate next steps are further analytical validation to cement its algorithm's positive and negative predictive performance. The company will also be working to gain access to data from studies of Enhertu in which it can measure the assay's correspondence to actual patient outcomes.
Paige is not alone in this effort, and will likely face competition from others. Owkin, for example, has said it is working with the UK's National Health Service Foundation Trust to develop its own model for HER2-low detection. Researchers have also seen recent success with other technologies like mass spectrometry.
Regarding the challenge of bringing its classifier into the clinic, Stefanelli said Paige has a good understanding, from its earlier approval, of the level of validation that the FDA will require. "We're having conversations with other partners who have already gone digital or are interested in collaborating with us so that we can continue to generate more and more data," she said.
With its CE and UKCA marks, the company will also be able to reach new collaborators with different populations of patients, something Stefanelli said is a major goal for the company.
"There are so many regions where this [H&E-stained slide] is the standard sample for diagnosis of cancer," she said. "Many of those patients will not have the opportunity to receive an IHC or a PCR assay, and certainly not sequencing, so I think one of our goals is to really start looking at different regions in the world and seeing how generalizable, for lack of a better word, our algorithm is."
According to Stefanelli, Paige has reason to believe it's at least on the right track in terms of achieving validity of its classifier across diverse populations because of the samples it has already been trained on, which mainly came from Memorial Sloan Kettering.
"They're the largest referral cancer center in the world, and so, because of that, they receive … samples across 50 different countries," she said, adding that this is also an element of what Paige sees as its advantage over some of the other companies exploring AI digital pathology, which may have more limited access.
She also highlighted the potential of Paige's approach to accelerate development of novel therapies for truly HER2-negative patients by better differentiating these triple-negative tumors from HER2-low cases.
"There are clinical trials that require HER2 to be negative, and we have proved that IHC 0 doesn't [necessarily] mean negative … so I think there's also a huge opportunity to go into [failed] clinical trials … and see if we can potentially identify [for instance] a patient subtype who was actually responding to those drugs."
"Ultimately the potential here is to completely reclassify these breast cancer patients," she said.
As the firm moves forward with its plans for additional validity and utility studies, Paige has other breast cancer goals, including applying its technology to prognosis and recurrence prediction.
More broadly, Stefanelli said Paige is exploring a variety of other therapy response settings, hoping to be able to create classifiers that don't just predict the presence of existing molecular biomarkers like HER2, but directly differentiate responders from non-responders based on pathology slide images.
"I don't think we're going to displace other testing, but … our vision is that we want to try to interrogate as much information out of a sample at the time of diagnosis as possible so that we can really help with the decision-making process on how to best serve patients," she said.