Abstract
Background: Cancer diagnosis, treatment and management are possible through the usage of multidisciplinary platforms. Herein, we present a unique Artificial Intelligent (AI)-enabled liquid biopsy platform that can provide sensitive and specific signals for the purpose of multi-cancer early detection (MCED), diagnosis, and therapy selection to minimal residual disease (MRD) and monitoring.
Methods: We utilized the Caris database composed of genetic data from over 350,000 tissue Whole Exome Sequencing (WES)/Whole Trasnscriptome Sequencing (WTS) of solid malignancies to train deep learning neural networks aimed at uncovering the molecular drivers of cancer. Additionally, WES/WTS sequencing was performed on 4,276 samples to develop a set of cell free DNA (cfDNA) specific features. One hundred and sixty-six of these samples had paired tissue sequenced WES/WTS to facilitate liquid biopsy variant validation. These features were organized into ‘pillars’ and integrated into the Assure Blood-based Cancer Detection AI (ABCDai) for early detection (binary classifier with 20-fold cross-validation) and tissue-of origin determination (multiclass classifier with 20-fold cross-validation). Subsequently, these models were evaluated using survival analysis on MRD and Monitoring samples.
Results: The Assure workflow detected CHIP mutations in 27% of samples, many of which were in clinically actionable genes (Figure 1). By leveraging CHIP subtraction, the detection of driver mutations from blood collected within 30 days of matched tumor tissue showed high concordance, with a Positive Percent Agreement (PPA)of 93.8% and a Positive predictive value (PPV) of 96.8%. Assure Plasma-derived features were independently significant for MCED (examples in Figure 3). When combined with tissue-identified features through ABCDai, we achieved an overall sensitivity of 87.4% at 99.5% specificity (Figure 4). ABCDai applied to plasma extracted before (MRD) and after (Monitoring) therapy showed significant stratification (p<0.05) for patient Disease-Free Survival (DFS), even when controlling for common clinicopathological variables (Figure 5). Tissue of origin is focused on those tumor types that are most common and thus provide the most utility (breast, CRC, gastric, H&N, NSCLC, ovarian, pancreatic, and prostate cancers). Among these cancers, we were able to determine the tissue of origin for 100% of the positive calls with an accuracy within the top three predictions of 84%.
Download Publication