DREAMS: deep read-level error model for sequencing data applied to low-frequency variant calling and circulating tumor DNA detection.

Circulating tumor DNA detection using next-generation sequencing (NGS) data of plasma DNA is promising for cancer identification and characterization. However, the tumor signal in the blood is often low and difficult to distinguish from errors. We present DREAMS (Deep Read-level Modelling of Sequencing-errors) for estimating error rates of individual read positions. Using DREAMS, we develop statistical methods for variant calling (DREAMS-vc) and cancer detection (DREAMS-cc). For evaluation, we generate deep targeted NGS data of matching tumor and plasma DNA from 85 colorectal cancer patients. The DREAMS approach performs better than state-of-the-art methods for variant calling and cancer detection.

Error Characterization and Statistical Modeling Improves Circulating Tumor DNA Detection by Droplet Digital PCR.

BACKGROUND: Droplet digital PCR (ddPCR) is a widely used and sensitive application for circulating tumor DNA (ctDNA) detection. As ctDNA is often found in low abundance, methods to separate low-signal readouts from noise are necessary. We aimed to characterize the ddPCR-generated noise and, informed by this, create a sensitive and specific ctDNA caller. METHODS: We built 2 novel complimentary ctDNA calling methods: dynamic limit of blank and concentration and assay-specific tumor load estimator (CASTLE). Both methods are informed by empirically established assay-specific noise profiles. Here, we characterized noise for 70 mutation-detecting ddPCR assays by applying each assay to 95 nonmutated samples. Using these profiles, the performance of the 2 new methods was assessed in a total of 9447 negative/positive reference samples and in 1311 real-life plasma samples from colorectal cancer patients. Lastly, performances were compared to 7 literature-established calling methods. RESULTS: For many assays, noise increased proportionally with the DNA input amount. Assays targeting transition base changes were more error-prone than transversion-targeting assays. Both our calling methods successfully accounted for the additional noise in transition assays and showed consistently high performance regardless of DNA input amount. Calling methods that were not noise-informed performed less well than noise-informed methods. CASTLE was the only calling method providing a statistical estimate of the noise-corrected mutation level and call certainty. CONCLUSIONS: Accurate error modeling is necessary for sensitive and specific ctDNA detection by ddPCR. Accounting for DNA input amounts ensures specific detection regardless of the sample-specific DNA concentration. Our results demonstrate CASTLE as a powerful tool for ctDNA calling using ddPCR.