Harbinger Health Presents Data Demonstrating Novel Capabilities of its Computational Platform to Enhance Performance and Clinical Informativeness in Blood-Based Early Cancer Detection

Data underscore Harbinger’s data science capabilities that power early-stage multi-cancer detection technology and advance the platform’s ability to provide multiple dimensions in clinical informativeness

Cambridge, Mass., November 14, 2023 – Harbinger Health (“Harbinger,” “the Company”), a biotechnology company pioneering new technologies to fundamentally change cancer screening and detection, today announced the presentation of four abstracts at clinical and scientific meetings for cancer research and data science. Three abstracts were presented at the American Association for Cancer Research (AACR) Special Conference in Cancer Research: Liquid Biopsy: From Discovery to Clinical Implementation in San Diego, California from November 13-16, 2024, and an additional abstract was presented at the Cold Spring Harbor Laboratory Biological Data Science meeting in Cold Spring Harbor, New York from November 13-16, 2024.

“The results presented across these four studies underscore the power of our computational platform to tackle some of the most significant computational challenges in early cancer screening, including data limitations in critical cohorts, signal complexity, and model accuracy,” said Kieran Chacko, Executive Vice President of Data Science and Strategy at Harbinger Health. “Our platform’s capabilities are enabling us to generate precise, disease-specific insights that could provide doctors and patients with critical information on early-stage disease, tissue of origin, and disease severity – helping deliver performance and clinical informativeness to blood-based cancer detection.”

Chacko continued, “Our computational platform is pioneering new frontiers in cancer detection by integrating advanced machine learning capabilities with foundational cancer biology. We developed transfer learning methods to map high-signal cancer tissue data to low-signal cfDNA, achieving highly accurate tissue of origin predictions. Additionally, we built a biology-guided deep neural network designed to interrogate and learn cancer methylation patterns at the single fragment level, significantly improving early-stage sensitivity. Finally, by using data augmentation approaches to generate high-fidelity cfDNA training datasets, we are overcoming challenges like limited data and signal heterogeneity, enabling our platform to isolate tumor signals with unmatched precision and opening pathways for broader applications across the cancer care continuum.”

The following abstracts were presented at AACR Special Conference in Cancer Research: Liquid Biopsy: From Discovery to Clinical Implementation:

A large-scale data simulation platform isolates tumor signal from cell-free DNA (cfDNA) and improves tissue of origin prediction accuracy (B065)

  • Harbinger developed capabilities to generate diverse, large-scale, high-fidelity simulated cell-free DNA (cfDNA) datasets for robust and accurate machine learning (ML) model development.
  • Specifically, the platform technology addresses challenges like limited data availability, heterogeneous low-signal cfDNA, and confounding technical and biological biases.
  • Supplementing tissue of origin (TOO) models with a curated simulated dataset resulting in a 10%-point increase in balanced accuracy across 10 groups of cancer types.
  • A key advantage of the platform is its ability to enhance ML models by matching confounding variables within training data, such as age or target coverage.
  • The platform is adaptable to other diseases and biofluids, enabling potential expansion to additional diagnostic and screening applications.

Transfer learning for accurate tissue of origin classification from cfDNA methylation (PR019, B055)

  • Harbinger developed a transfer learning method that leverages methylation signatures learned from cancer tissue biopsy samples to predict the TOO of clinical cancer cfDNA samples.
  • This method overcomes major challenges of developing robust and generalizable TOO models, including insufficient cancer type-specific training data and technical confounders in low-signal cfDNA.
  • The method utilizes the high methylation signal in cancer tissue biopsies for feature extraction and adapts it to plasma cfDNA data for accurate TOO classification.
  • The proposed model achieved 89% balanced accuracy in classifying the TOO across 10 groups of cancer types, and outperformed models trained solely on plasma cfDNA data.
  • This approach demonstrates the potential for scalable, automated feature extraction from signal-rich tissue biopsy data for TOO and could potentially extend capabilities to rarer cancer types and further disease sub-stratification.

Liquid biopsy-based detection of triple negative breast cancer using DNA methylation biomarkers (A053)

  • Harbinger investigated whether a blood-based test could provide a non-invasive, sensitive method for detecting early triple-negative breast cancer (TNBC), aiming to overcome the constraints of conventional imaging techniques.
  • Specific methylation biomarkers distinguishing TNBC were identified, and a feed-forward neural network classifier was developed that applied transfer learning and these biomarkers to learn methylation signatures in cancer tissue biopsy samples and map them to cfDNA.
  • Harbinger’s platform accurately distinguished TNBC from non-TNBC using a binary classifier, achieving 84% accuracy in cfDNA samples.
  • These results demonstrate the platform’s ability to identify high-risk, aggressive disease with a blood-based methylation test, offering insights into cancer severity and potentially reducing over-treatment and over-diagnosis.

Additionally, Harbinger presented the following abstract at the Cold Spring Harbor Biological Data Science Meeting:

Extracting tumor signal from cell-free DNA using deep neural networks improves early cancer detection

  • Harbinger designed a novel deep-learning framework to detect cancer methylation signal at the single-DNA-fragment resolution to improve early-stage sensitivity where the signal is weakest.
  • This capability was enabled by Harbinger’s simulated data platform, through which billions of simulated samples were generated for model training.
  • Classifiers trained on feature vectors extracted using this framework improved overall sensitivity by 12.8% points at 98% specificity, compared to a standard benchmark classifier trained on summary methylation statistics.
  • Improvements to early-stage sensitivity were observed across multiple cancer types, and classifier scores showed lower correlation with biological (e.g., age) and technical variables (e.g., bisulfite conversion efficiency).
  • These results, combined with the model’s training efficiency, highlight the transformative potential of models developed for fragment-level resolution data in early-stage cancer detection.

About Harbinger Health

Harbinger Health is pioneering the detection of early cancer and enabling foundationally new approaches to cancer screening, diagnosis and management. The company combines advances in artificial intelligence with proprietary insights into the biology of the beginnings of cancer to identify cancer before it is visible or symptomatic with the aim of developing low-cost, multi-cancer blood tests. Harbinger envisions a future where, instead of keeping cancer from spreading, it could be kept from forming, making a cancer diagnosis a routine health problem to be addressed rather than a life-altering event to be feared with profound implications for people, healthcare systems and societies. Harbinger was founded by Flagship Pioneering after three years of foundational research in its Labs unit and launched in 2020. Learn more about Harbinger by visiting www.harbinger-health.com or following us on Twitter and LinkedIn.

Media Contact

press@harbinger-health.com