Detection of early-stage cancer using similarity-based classification of cell-free DNA fragmentation patterns

Current cancer screening efforts face the dual problems of low patient adherence and a lack of available tests for many cancers, which motivates the development of less invasive multi-cancer early detection strategies. Such strategies will likely involve the use of liquid biopsies, which can analyze biomarkers found in blood for the presence of cancer. We introduce Blackjack, a new detection approach that focuses on the fragmentation patterns of plasma-derived cell-free DNA (cfDNA). Blackjack characterizes a person's cfDNA fragmentation sites with a position probability matrix (PPM) that encodes information about the sequence context of the sites. For classification, Blackjack computes two features that represent the similarity of an individual's PPM to those of known healthy and cancer cohorts. With these two features, Blackjack then uses logistic regression to classify a new sample as being from a healthy individual or a person with cancer. In cross-validation experiments using two external whole genome sequencing datasets, Blackjack showed high performance, achieving a minimum area under the receiver operating characteristic curve (AUC) of 0.993 and minimum sensitivity (Sn) of 91.8% (at 98% specificity). This high performance was also present for stage I (AUC 0.991, Sn 89.8%) and stage II cancers (AUC 0.993, Sn 90.2%), and across eight different cancers (minimum AUC 0.985, minimum Sn 80.0%). We demonstrate that Blackjack's accuracy is stable at low simulated sequencing depths (~0.24x), and that tissue of origin prediction using Blackjack's similarity measures is high (minimum 82% accuracy of the top two predicted tissues). Our findings show that cfDNA fragmentation site contexts have valuable signals for early cancer detection, and that developing diagnostics with machine learning can benefit from a focus on having a small number of informative features.