The introduction of next-generation sequencing (NGS) was one of the most significant advancements in molecular technology, as it made genomic level sequencing practical for numerous research and clinical applications. Today, NGS is commonly used to identify known and novel cancer mutations, detect familial cancer mutation carriers, and provide molecular rationale for appropriate targeted therapy.1 NGS is also a core research technology, used to identify novel driving mutations and associated oncogenic pathway alterations that are fertile ground for development of new tailored therapies.
Yet while there are numerous benefits of NGS, there are limitations. NGS requires relatively large amounts of tissue in order to have enough DNA or RNA to sequence. On limited biopsy specimens or specimens which are not especially rich for tumor cells, this can pose a challenge. NGS also requires considerable investment in devices and reagents, as well as bioinformatics support.
On the molecular biology side, sequencing indicates if a mutation is present within a tumor. The identified mutation is often associated with the expression of the mutation and thus the predicted phenotype—but this is not always the case. A mismatch between having a mutation and how, or even if, that mutation is used by the tumor can have significant treatment consequences. One example of a mismatch between identified, known oncogenic mutation by tumor sequencing and predicted therapeutic response is the failure of EGFR targeted agents to show benefit in EGFR mutant glioblastoma (GBM), despite high initial optimism from preliminary results and sequencing analysis identifying this as a common mutation in GBM.2,3 A technique indicating whether the tumor has the expected treatment-responsive phenotype paired with identification of a driving mutant oncogene activation, then, might be more predictive of treatment response than sequencing, or either technique alone.
Here at Paige, our version of a technique that could accomplish this is what we call digital biomarker assays, and we think of these as “advanced phenotypic analysis.” Our AI is trained on the complete morphology of a tumor in a hematoxylin and eosin (H&E)-stained slide, learning everything from how the tumor cells look and aggregate, to how the stroma looks near and far from the tumor, to the immune infiltrates and components of the tumor microenvironment. This morphology is shaped by a number of biologic processes, which means the morphology can tip us off to what biologic processes are active within that tumor, including how the tumor is using its genome transcriptionally and translationally, as well as any post-translational modifications that may be occurring. Cell signaling near and far from the tumor will also influence all those parts of the morphology, as will interactions with the immune system. In this way, the morphology on the H&E slide captures, in aggregate, all of the biological processes active on the tumor in a given slide.
Thus, AI can be applied to find deeper insights than are captured in a standard NGS test. The phenotype of the tumor, captured in the morphology our AI trains on, reflects not only if a mutation is present, but if it is active and driving the tumor. This allows our AI to be an important, functional, phenotypic complement to NGS testing, and in some settings, we can go beyond NGS entirely. After all, it is the phenotype we test, and the phenotype that the drug treats, not the genotype.
Therefore, if a biologic signal is specific and strong enough to be recognizable in morphology, our algorithms can find a phenotype that is predictive of outcome, rather than looking at another biomarker as a surrogate to outcome. This approach could encompass all the changes that would otherwise need multiple tests including NGS, RNAseq, reverse transcriptase polymerase chain reaction (RT-PCR), or immunohistochemistry (IHC) to get the same spectrum of tumor phenotypic assessment from genotype to RNA transcription to protein translation. To do any of NGS, RT-PCR, or IHC requires additional tissue beyond the one H&E slide needed for a digital biomarker; to do all of NGS, RT-PCR, and IHC on the same case would consume an enormous amount of precious biopsy material. In this way, digital biomarker assays can save time, money, and tissue, while providing equivalent or better clinically actionable information about the tumor.
To show the potential digital biomarker assays hold for clinical trials, let’s say we are developing a drug targeting a specific genetic mutation in a specific form of cancer. Normally, the pharmaceutical company would use traditional biomarker testing such as NGS to identify patients whose cancers contain these mutations and enroll these patients into their trial. However, by working with Paige AI, they see that digital biomarker assays using the scanned H&E slide can predict which cases of that cancer are likely to harbor that particular mutation. Further, they see that Paige can also identify cases where the mutation is driving activation of the oncogenic pathway. Thus, using the digital biomarker assay on the scanned H&E slide from the patient’s diagnostic biopsy or resection could allow them identify cases which are likely to have the targeted mutation present. Ideally, only these cases are then sent for NGS to confirm the presence of the targeted mutation, saving enormous amounts of money that would otherwise be needed to use NGS to screen all patients with this cancer for this mutation. Further, the patients whose cases do NOT harbor this mutation would ultimately save time and precious tissue by finding out this clinical study is not for them—letting them find an alternative study that might be better for their tumor or start a different treatment.
The combination of the digital biomarker assay, which recognizes the active phenotype, along with downstream diagnostic testing means that patients who enroll in the study are more likely to not only have the targeted mutation, but have the targeted mutation driving their tumor. This means more patients in the treatment arm are likely to respond to the therapy, since patients with a mutation that is not driving the tumor, or even a false positive identification of the mutation, are screened out by the digital biomarker and not enrolled in the study. Instead, they go to a treatment more suited to their tumor. On average, the treatment arm would contain more patients whose tumors are likely to respond to the therapy. The overall response rate for the treatment arm will be better, demonstrating that the new drug is effective for those patients whom the combination of the digital biomarker and NGS identify as likely responders.
This approach may also improve adoption of the drug in the clinic once it is approved. Not only does it show that the drug is effective in this population, but the ability to rapidly and cheaply screen for cases likely to harbor the targeted mutation means resource reluctant or restricted practice settings can save the expensive NGS needed to confirm the targeted mutation’s presence to those cases most likely to harbor the mutation. For NGS, the concern about the expense of the testing is well known, with the literature expressing concern about its cost-effectiveness4. An effective digital screening assay to triage for these expensive traditional biomarkers can shift the cost-effectiveness argument in favor of these combined digital biomarker-traditional biomarker complementary assays. More cases screened means more patients who could benefit from the new drug are found, improving access to the new medication, and alleviating the all-too-common payer anxiety about paying for medications that may not be effective in all of their patients.
Paige’s approach to biomarker AI is exciting thanks not only to the potential savings of time, treasure and tissue, but also the additional information it can provide to patients, physicians, and pharmaceutical researchers to better support treatment decisions in the clinic, clinical trial enrollment, and drug development.
We see a world where these same methods can support, or in certain settings even replace, NGS in diagnosis, offering benefits of time savings and accuracy directly to the patient. For now, we are working to develop these methods further, partnering with leading academic and industry research teams to deliver on the vast potential of digital biomarker assays.
1Guan YF, Li GR, Wang RJ, et al. Application of next-generation sequencing in clinical oncology to advance personalized treatment of cancer. Chin J Cancer. 2012;31(10):463-470. doi:10.5732/cjc.012.10216
2Raizer JJ, Abrey LE, Lassman AB, Chang SM, Lamborn KR, Kuhn JG, Yung WK, Gilbert MR, Aldape KA, Wen PY, Fine HA, Mehta M, Deangelis LM, Lieberman F, Cloughesy TF, Robins HI, Dancey J, Prados MD; North American Brain Tumor Consortium. A phase II trial of erlotinib in patients with recurrent malignant gliomas and nonprogressive glioblastoma multiforme postradiation therapy. Neuro Oncol. 2010 Jan;12(1):95-103. doi: 10.1093/neuonc/nop015. Epub 2009 Dec 14. PMID: 20150372; PMCID: PMC2940554.
3Karpel-Massler G, Westhoff MA, Kast RE, Wirtz CR, Halatsch ME. Erlotinib in glioblastoma: lost in translation? Anticancer Agents Med Chem. 2011 Oct;11(8):748-55. doi: 10.2174/187152011797378788. PMID: 21707495.
4Christofyllakis K, Bittenbring JT, Thurner L, Ahlgrimm M, Stilgenbauer S, Bewarder M, Kaddu-Mulindwa D. Cost-effectiveness of precision cancer medicine-current challenges in the use of next generation sequencing for comprehensive tumour genomic profiling and the role of clinical utility frameworks (Review). Mol Clin Oncol. 2022 Jan;16(1):21. doi: 10.3892/mco.2021.2453. Epub 2021 Nov 25. PMID: 34909199; PMCID: PMC8655747.