Deconvolution methods for quantification of copy number variations in liquid biopsy sequencing

Abstract

Copy number variations, prevalent in cancers, are genomic alterations that result in losses or gains of entire genomic regions. Such alterations can be evaluated using cheap low-pass whole genome sequencing using liquid biopsies. These methods are promising for tracking the evolution of the cancer in real-time, due to their low cost and noninvasive nature which enable frequent sampling. The DNA sequenced from liquid biopsies is a mixture of cancer-specific DNA and DNA from healthy cells, the latter without these alterations. Therefore liquid biopsies can, for example, help monitor the proportion of an emerging cancer subtype in the tumor. Lakatos et al. [10] introduced methods for estimating the cancer proportion and the proportion of the most dominant cancer subtype in the sample, termed purity estimation and subclonal tracking. However, our ability to track the cancer evolution is hindered by the low signal in such samples, due to the contamination of healthy DNA, and measurement noise. We thus aim to develop methods for denoising and deconvolution of the underlying copy number profile of the tumor, to enhance the signal in liquid biopsy sequencing measurements. In this work, we evaluate two frameworks for deconvoluting such samples: a denoising autoencoder and Bayesian change point detection. We compare these methods to rolling median-based segmentation, using the mean squared error of the reconstructed copy number profile and the F1-score. We demonstrate that both deconvolution methods work better than the rolling median in low-purity and noisy regions. We then implement our methods for purity estimation and subclonal tracking, based on the methods by Lakatos et al. and using the denoised data obtained from the previous step. In general, we find that Bayesian change point detection outperforms the other methods, is suitable for denoising liquid biopsy samples, and can be used for subclonal tracking. Using our full updated pipeline, we can improve the estimation of purity and subclonal ratio values, especially in low-purity and low-quality samples.