Novel blood-based assay for detection of early stage multi-cancer.

Abstract

Listen

Translate article

e15035 Background: Early cancer detection has the potential to significantly improve patient outcomes and reach the Cancer Moonshot goals of reducing the death rate from cancer by at least 50 percent over the next 25 years. Harbinger Health is pioneering early cancer detection with a blood-based test that combines recent genomic and epigenomic discoveries of early cancer and biology-informed artificial intelligence. Unlike prior approaches that are purely statistical to identify informative biomarkers, Harbinger’s approach is informed by insights into specific biological events early during tumorigenesis and is therefore optimized for detecting cancer in patients with very low levels of circulating tumor DNA. Methods: Here, we present data generated using the Harbinger Health assay from 1,046 subjects, 621 with newly diagnosed, treatment-naive cancer (15 different cancer types) and 425 individuals with no history, diagnosis, or cancer symptoms. Using this sample cohort, we developed and applied a rigorous framework for training, calibration and predicting likelihood of cancer using a multi-layered logistic regression-based machine learning algorithm, generating final outputs of binary classification (cancer yes/no). Our framework involved multiple iterations of 10-fold cross-validation of the full dataset, and we report solely on samples that appear in the held-out test set in each iteration. Additionally, we developed a similar framework to predict tissue of origin (TOO) for cancer samples. Results: The overall sensitivity of cancer detection was 82% (95% confidence interval (CI): 72.7-91.0%) at 95% specificity. Notably, the sensitivity was 74% (95% CI: 54.8-92.7%) for stage 1 and 84% (95% CI: 65.3-100%) for stage 2. Furthermore, we were able to correctly predict cancer in 95% of patient samples with at least 0.037% tumor fraction. The overall sensitivity for high incident cancers were: breast (73%), prostate (82%), lung (85%), and colorectal (96%) at 95% specificity and the overall accuracy of TOO prediction was 86% when the tumor fraction was greater than 0.1% in the top 3 most prevalent cancer types (breast, colorectal and lung). Importantly, we noted that technical variability was introduced when performing assay optimization strategies, yet we observed comparable performance when assessing subsets of stably processed samples, indicating that our performance is robust. Conclusions: Overall, these results demonstrate that the Harbinger Health platform, with its biology-informed approach, has the potential for highly sensitive multi-cancer diagnostic accuracy and specifically those with early-stage cancer. The platform is now being validated in a 10,000-subject prospective clinical trial (CORE-HH/NCT05435066).