ľֱ

MedpageToday

Ludmil Alexandrov, PhD, on Use of AI to Predict HRD and Response Directly From Routine Biopsy Slides

– 'Bypasses the need for traditional, time-consuming, and expensive genomic sequencing,' researchers say


This Reading Room is a collaboration between ľֱ® and:

Medpage Today

A novel artificial intelligence (AI) protocol called DeepHRD was able to predict homologous recombination deficiency (HRD) in ovarian and breast cancers directly from routine histopathological slides. Compared with molecular testing, DeepHRD classified 1.8 to 3.1 times more patients with HRD, predicting overall survival (OS) in high-grade serous ovarian cancer and platinum-specific progression-free survival in metastatic breast cancer.

In the past 5 years, there has been an explosion of AI approaches to predict genomic changes from digital images of histopathologic slides that may lead to clinical action. As detailed in a study in the , researchers at the University of California San Diego introduced DeepHRD, which allows rapid, low-cost detection of clinically actionable genomic alterations, specifically HRD, directly from tumor biopsy slides.

In the following interview, Ludmil B. Alexandrov, PhD, of the Moores Cancer Center, elaborates on the findings and discusses the pros and cons of using AI to predict HRD.

What does this study add to the literature?

Alexandrov: This method bypasses the need for traditional, time-consuming, and expensive genomic sequencing, marking a significant advancement in the use of AI in precision oncology. The study showcases that AI can accurately and instantaneously identify HRD biomarkers, potentially revolutionizing the way precision oncology is practiced, especially in resource-constrained settings.

We trained DeepHRD using 1,008 primary breast and 459 ovarian cancers from The Cancer Genome Atlas. DeepHRD was compared with four standard HRD molecular tests using 349 breast and 141 ovarian cancers from multiple independent data sets, including platinum-treated clinical cohorts.

Through transfer learning to high-grade serous ovarian cancer, DeepHRD predicted that HRD samples had better OS after first-line and neoadjuvant platinum therapy in two cohorts.

In addition:

  • The AI test had a negligible failure rate, compared with a 20-30% failure rate of current genomic tests, eliminating the need for re-testing or invasive re-biopsy
  • The approach significantly reduced the time and cost associated with identifying actionable genomic biomarkers, thereby accelerating the initiation of targeted therapies for patients with HRD-positive breast and ovarian cancers
  • The study emphasizes the potential of this AI-driven method to address health disparities by making precision oncology accessible and equitable, particularly in resource-constrained regions

What are the disadvantages of detecting HRD in the clinic via molecular profiling?

Alexandrov: The primary disadvantages of detecting HRD through traditional molecular profiling include:

  • High costs: Molecular profiling is expensive, which can be a barrier, especially in resource-limited settings. In contrast, an AI test like DeepHRD has virtual zero cost for performing as it requires only a digital image of a routinely generated histopathological slide
  • Time delays: The process can take weeks, leading to life-threatening delays in initiating treatment. In contrast, an AI test like DeepHRD is instantaneous, making prediction without the need for additional tests

How does DeepHRD compare with other AI approaches for predicting genomic changes?

Alexandrov: DeepHRD is unique compared with other AI approaches because, to the best of our knowledge, no other AI tests currently exist that allow for clinically actionable predictions directly from routine biopsy slides.

While other AI methods may focus on predicting genomic changes, they often require complex and expensive inputs, such as additional staining, which are not always readily accessible or timely. Moreover, other AI approaches have not demonstrated clinical actionability, limiting their immediate utility in guiding treatment decisions.

In contrast, DeepHRD operates on easily obtainable histopathological slides, allowing for rapid, accurate, and actionable predictions with minimal delay. This makes DeepHRD not only faster and more cost-effective, but also uniquely positioned to be immediately useful in clinical settings, providing oncologists with the critical information needed to initiate targeted therapies promptly.

In the future, for what other cancers might DeepHRD be potentially applied?

Alexandrov: While DeepHRD is currently focused on breast and ovarian cancers, there is significant potential for its application in other cancer types. As an immediate next step, we are extending its use to detect HRD in pancreatic and prostate cancers. Following this, our goal is to expand the technology to identify other critical biomarkers, such as KRAS and BRAF mutations, across a broader range of cancers.

This extension will further enhance the applicability of our AI approach in precision oncology, making it a versatile tool in cancer diagnostics and treatment planning.

What is your main message for practicing oncologists?

Alexandrov: Practicing oncologists should recognize that AI approaches, such as DeepHRD, represent a significant advancement in precision oncology, aimed at enhancing clinical care and patient outcomes. This technology addresses the challenges associated with traditional molecular profiling, including high costs, delays, and frequent failure rates.

Oncologists should consider leveraging this innovation to improve patient outcomes by accelerating access to targeted therapies, particularly in settings where genomic testing may not be feasible.

The adoption of AI technologies like DeepHRD has the potential to reduce disparities in cancer care and ensure that more patients receive timely and precise treatments.

Read the study here.

Alexandrov disclosed employment with bioTheranostics, a leadership role with io9 LLC, and consulting/advising with Genome Insight; in addition, he noted institutional patents for several related patent application.

Primary Source

Journal of Clinical Oncology

Source Reference:

ASCO Publications Corner

ASCO Publications Corner