DISMIR: a deep learning-based cancer-detection method by integrating DNA sequence and methylation information of individual cell-free DNA reads

Abstract

Detecting cancer signals in cell-free DNA (cfDNA) high-throughput sequencing data is emerging as a novel non-invasive cancer detection method. Due to the high cost of sequencing, it is crucial to make robust and precise prediction with low-depth cfDNA sequencing data. Here we propose a novel approach named DISMIR, which can provide ultrasensitive and robust cancer detection by integrating DNA sequence and methylation information in plasma cfDNA whole genome bisulfite sequencing (WGBS) data. DISMIR introduces a new feature termed as “switching region” to define cancer-specific differentially methylated regions, which can enrich the cancer-related signal at read-resolution. DISMIR applies a deep learning model to predict the source of every single read based on its DNA sequence and methylation state, and then predicts the risk that the plasma donor is suffering from cancer. DISMIR exhibited high accuracy and robustness on hepatocellular carcinoma detection by plasma cfDNA WGBS data even at ultra-low sequencing depths. Analysis showed that DISMIR tends to be insensitive to alterations of single CpG sites’ methylation states, which suggests DISMIR could resist to technical noise of WGBS. All these results showed DISMIR with the potential to be a precise and robust method for low-cost early cancer detection.

Cost-effective methylome sequencing of cell-free DNA for accurately detecting and locating cancer

“…The α-value of an individual sequencing read captures the pervasive nature of methylation, therefore allowing us to "purify" tumor DNA reads from background reads for enhancing the signal-to-noise ratio. A recent study further used the α-value concept for identifying liver cancer methylation markers 28 . Here we developed a general framework for using read-level α-values to robustly identify markers from impure tissue samples or even cfDNA plasma samples.…”

Cell type deconvolution of methylated cell-free DNA at the resolution of individual reads

“…After all, the method for finding markers was optimized for CpG count data and then translated almost exactly to read average data. Read averages may, however, require a completely different approach for finding markers, such as the switching reads defined by Li et al (12). An adequate set of differential regions not only improves model performance but also allows for targeted sequencing of these regions only, for example using RRBS, and can thus reduce the sequencing cost (20).…”

“…A read average is calculated by dividing the number of methylated CpG sites by the total number of CpG sites on a read, where only reads with three or more CpG sites are used. This heuristic is adopted from previous methods (11,12). The read average is then rounded to the closest value i. E.g.…”

“…A read-based approach has been adopted in multiple other tumor fraction estimation methods, such as DISMIR (12) and © The Author(s) This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/2.0/uk/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited. EpiClass (13).…”

“…A read-based approach has been adopted in multiple other tumor fraction estimation methods, such as DISMIR (12) and EpiClass (13). Even though the effectiveness of this approach has been shown for tumor fraction estimation, it has not yet been used in the related task of cell type deconvolution.…”