SNPmanifold : detecting single-cell SNV clonality and lineages using binomial variational autoencoder

Abstract

Recently, single-cell lineage tracing experiments using random passenger DNA mutations as genetic markers become popular, because they are non-invasive and can be applied to many biological samples without biochemical intervention. However, SNV clone assignment and lineage inference in this type of data remain a challenge due to hierarchical mutation structure and many missing signals. Existing SNV analysis methods were mainly designed for simpler mutation patterns with low covariance in different SNPs, so they do not work well in lineage tracing data with high covariance in different SNPs due to hierarchical evolutionary history. Applying them to lineage tracing data easily results in abnormal convergence and inaccurate conclusions. To solve this problem, I developed SNPmanifold, a Python package that learns an embedding manifold with a better cell-cell distance metric using a shallow binomial variational autoencoder. Based on this better cell-cell distance metric, SNPmanifold can avoid numerical issues brought by high covariance in different SNPs and assign SNV clones and lineages more accurately based on neighboring cells. Also, visualization of hierarchical mutation structure becomes clear under UMAP with this better cell-cell distance metric. In this thesis, I demonstrated that SNPmanifold can effectively identify a large number of multiplexed donors of origin (k = 18) with nuclear SNPs that all existing unsupervised methods fail, and lineages of mitochondrial somatic clones with meaningful biological interpretations such as agreements with cell types, CNV clones, and TRB clonotypes. In short, SNPmanifold is a better SNV analysis method for high-covariance single-cell lineage tracing data that I developed in this thesis.