Evolution-Aware Deep Reinforcement Learning for Single-Cell DNA Copy Number Calling

Abstract

Recent high-throughput single cell DNA sequencing technologies enable one to detect copy number aberrations (CNAs) on thousands of individual cells within a tumor. Several new methods enable haplotype-specific copy number calling on such data. However, these algorithms do not utilize evolutionary constraints, leading to the inference of spurious CNAs due to measurement noise that are inconsistent with a realistic evolutionary model. To address this gap, we introduce DeepCopy, an evolution-aware deep reinforcement learning algorithm for haplotype-specific copy number calling on single-cell DNA sequencing data. On simulated data, DeepCopy's predictions better fit the ground truth than existing methods. As shown on 10x Genomics Single Cell CNV sequencing of several breast cancers and DLP+ sequencing of an ovarian cancer, DeepCopy's joint estimation of copy number profiles and evolutionary model parameters results in larger, more realistic clones of cells with identical copy number profiles that lead to more parsimonious CNA phylogenies than existing methods, while retaining agreement with truncal single-nucleotide variants. Finally, on a breast cancer patient sequenced with both ACT and 10x technologies, we demonstrate the consistency of DeepCopy's predictions across sequencing technologies.