AI Outperforms Standard HRD Tests for Breast, Ovarian Cancers

A deep learning tool called DeepHRD, built on io9’s OncoGaze platform, predicted homologous recombination deficiency (HRD) with greater accuracy than FDA-approved standard molecular tests, achieving an area under the curve (AUC) of 0.81, according to a new study published in the Journal of Clinical Oncology.¹ The study focused on breast and ovarian cancers using multiple external cohorts, slide scanners, and tissue fixation variables to train and test the model.¹

“It was quite surprising that we outperformed [existing molecular tests], but when we looked at why we can outperform them, it’s [because] this model is going to pick up meaningful information, [such as] the presence of macrophages, the necrotic regions, aspects that we know biologically make sense, and that can be detected from the tumor microenvironment, which you usually cannot see from a molecular test,” Ludmil Alexandrov, PhD, told Targeted Therapies in Oncology in an interview.

Alexandrov is an associate professor in the Department of Cellular and Molecular Medicine and the Department of Bioengineering at the University of California, San Diego. He is also the cofounder and chief scientific officer of io9 and lead author of the study.

Methods and Results

In the study, DeepHRD used hematoxylin and eosin-stained slides from The Cancer Genome Atlas (TCGA) for the primary cohort. The artificial intelligence (AI) model’s performance was confirmed in 2 additional independent breast cancer cohorts, achieving AUCs of 0.76 and above in both cohorts.

“We trained our model using data from a set of patients sourced from public repositories and our own available data. For validation, we used data from entirely different sources that used varied staining and imaging protocols…. Our validation data includes sources from France [Georges-François Leclerc Cancer Center], Memorial Sloan Kettering [Cancer Center], and the Canada/UK Molecular Taxonomy of Breast Cancer International Consortium, each with distinct data generation methods,” Alexandrov explained.

DeepHRD identified 1.8 to 3.1 times more patients with HRD, who demonstrated improved overall survival (OS) in high-grade serous ovarian cancer and better progression-free survival (PFS) with platinum treatment in metastatic breast cancer, according to investigators. Through transfer learning, the model was applied to high-grade serous ovarian cancer, and DeepHRD-predicted HRD samples showed better OS following both first line (HR, 0.46; P = .030) and neoadjuvant platinum therapies (HR, 0.49; P = .015) in 2 cohorts.

In an external cohort of platinum-treated metastatic breast cancer, DeepHRD-predicted HRD samples had improved clinical outcomes, including a 3.7-fold increase in median PFS (14.4 vs 3.9 months) with an HR of 0.45 (P = .0047). No significant impact on outcomes was observed in patients receiving nonplatinum treatment, indicating that DeepHRD is a specific and predictive biomarker for platinum treatment.

Why Is HRD Important?

HRD, a condition in which cells struggle to repair DNA damage effectively, increases cancer risk and affects treatment responses.² Traditionally, cancer treatment relied on the clinicopathological characteristics of the tumor and the affected organ. However, with advancements in genomic analysis and targeted therapies, understanding the tumor’s genetic alterations is increasingly important for its classification and treatment.² Thus, identifying HRD in patients with various cancer types is crucial as it helps guide more targeted and effective therapies, particularly with platinum-based treatments.³

Trials have highlighted the value of HRD testing in identifying advanced ovarian cancer most likely to benefit from treatment with a PARP inhibitor.⁴ Ideally, universal HRD testing using genomic scar tests helps tailor maintenance treatment plans for patients after they respond to initial platinum-based chemotherapy. However, widespread implementation is hindered by limited access to HRD testing in many regions and the socioeconomic challenges associated with broad genomic testing.

Why Involve AI?

Standard tests require molecular profiling, which may not be universally available, whereas DeepHRD evaluates digital images, making it more accessible. “When we think about biomarkers, especially in cancer… usually one thinks about using a biological specimen, and there’s a number of problems with that: (a) you need to use biospecimens, which are precious, and there are too few of them, and (b) it takes time, and in the case of next generation sequencing, it is quite expensive,” Alexandrov explained.

The process of obtaining accurate cancer diagnostic tests is both frustrating and time-consuming, Alexandrov continued. Typically, it involves multiple steps: First, a biopsy is done, then analyzed by a pathologist, and often followed by additional tests such as next-generation sequencing or other precision oncology tests extending the process. This can lead to weeks or even months of waiting, and sometimes tests come back insufficient, requiring further biopsies and delays in treatment. According to study researchers, “currently, all approaches for detecting HRD in the clinic rely on molecular profiling leading to clinical workflow bottlenecks largely attributed to the availability of tissue.”¹

Due primarily to their expense, these tests are mostly available in the US and parts of Western Europe and are often inaccessible in other regions, Alexandrov explained. “Even within the US, it is done in very specific settings…for people who can afford to do them,” Alexandrov said. “This brings huge inequality in cancer diagnostics and in cancer care,” he added.

In contrast, once a biopsy is performed, the DeepHRD model uses a digital image of the biological specimen, which can be quickly sent and analyzed to provide test results almost immediately. DeepHRD analyzes the digitized patient biopsy to detect HRD by assessing the tissue at various magnifications. Based on this analysis, the model generates a final prediction score for the biopsy and suggests appropriate treatment options, such as PARP inhibitors or platinum-based chemotherapy (FIGURE)¹. T his streamlined process reduces wait times and costs, integrates pathologists more centrally into precision oncology, and minimizes reliance on shipping samples, resulting in a more efficient, economical, and digital workflow.

The Downfalls

The diversity in quality levels of medical record data collection can make implementing AI into clinical testing a challenge, according to a study on how AI is revolutionizing the clinical setting.⁵ Study authors explain that it would be ideal to have a single worldwide method for gathering and using data; however, researchers in the HRD study found a way to leverage this diversity in order to validate their findings by using different data sets with diverse forms of collection.¹

The typical challenges of inherent bias in AI models, due to the overrepresentation of White and Caucasian patients, persist. However, Alexandrov believes that the DeepHRD model could improve data collection across more ethnicities due to its accessibility and cost-effectiveness. That said, the DeepHRD model still needs to incorporate data from a broader range of populations, and its accessibility may help achieve this goal.

The Next Steps

“For our ovarian cohort, we had measurements of real-world [data],” Alexandrov said. In our breast cancer cohort, we could show that this tool is not only prognostic, but also a predictive biomarker because we had a controlled environment. In the ovarian cancer cohort, we were able to show it is a prognostic biomarker, but [we] need clinical trials to show whether it is a predictive prognostic biomarker, which we expect it to be,” he relayed.

Currently, biomarker approvals for HRD are done retroactively on clinical trial data. For example, AstraZeneca conducts clinical trials, and then companies such as Myriad retrospectively assess biomarkers on those trial cohorts, Alexandrov explained. “Our advantage is we do not need to use the samples…. We need [only] an image [of the sample], which is a standard that the companies have, and then we can run the test,” Alexandrov said. Study researchers plan to use the technology to validate treatments with different PARP inhibitors, following FDA precedents, and aim to develop it as a companion diagnostic for specific therapies.

“We are planning to build several other tests for detecting other clinically actionable biomarkers on this platform. We have now the results for ovarian and breast cancer and are able to do it for prostate and pancreatic cancers for other indications where it is clinically actionable. Then we’ll start looking at other biomarkers, KRAS, EGFR, and so on, by transferring them from a molecular platform to an AI platform for faster, better, [and] cheaper biomarker detection,” Alexandrov concluded.

REFERENCES:

1. Bergstrom EN, Abbasi A, Díaz-Gay M, et al. Deep learning artificial intelligence predicts homologous recombination deficiency and platinum response from histologic slides. J Clin Oncol. Published online July 31, 2024. doi:10.1200/jco.23.02641
2. Toh M, Ngeow J. Homologous recombination deficiency: cancer predispositions and treatment implications. Oncologist. 2021;26(9):e1526-e1537. doi:10.1002/onco.13829
3. Ray-Coquard I, Pautier P, Pignata S, et al. Olaparib plus bevacizumab as first-line maintenance in ovarian cancer. N Engl J Med. 2019;381(25):2416-2428. doi:10.1056/nejmoa1911361
4. Ngoi NYL, Tan DSP. The role of homologous recombination deficiency testing in ovarian cancer and its clinical implications: do we need it? ESMO Open. 2021;6(3):100144. doi:10.1016/j.esmoop.2021.100144
5. Chopra H, Annu, Shin DK, et al. Revolutionizing clinical trials: the role of AI in accelerating medical breakthroughs. Int J Surg. 2023;109(12):4211-422. doi:10.1097/JS9.0000000000000705