Dhaka: variational autoencoder for unmasking tumor heterogeneity from single cell genomic data

Abstract

Intra-tumor heterogeneity is one of the key confounding factors in deciphering tumor evolution. Malignant cells exhibit variations in their gene expression, copy numbers, and mutation even when originating from a single progenitor cell. Single cell sequencing of tumor cells has recently emerged as a viable option for unmasking the underlying tumor heterogeneity. However extracting features from single cell genomic data in order to infer their evolutionary trajectory remains computationally challenging due to the extremely noisy and sparse nature of the data. Here we describe 'Dhaka', a variational autoencoder method which transforms single cell genomic data to a reduced dimension feature space that is more efficient in differentiating between (hidden) tumor subpopulations. Our method is general and can be applied to several different types of genomic data including copy number variation from scDNA-Seq and gene expression from scRNA-Seq experiments. We tested the method on synthetic and 6 single cell cancer datasets where the number of cells range from 250 to 6000 for each sample. Analysis of the resulting feature space revealed subpopulations of cells and their marker genes. The features are also able to infer the lineage and/or differentiation trajectory between cells greatly improving upon prior methods suggested for feature extraction and dimensionality reduction of such data.

Searching large-scale scRNA-seq databases via unbiased cell embedding with Cell BLAST

“…The adversarial component in Cell BLAST enables a free form of variational posterior distribution, which could be learned from data directly. In contrast, the canonical variational autoencoder model used by scVI and several other tools 11,[25][26][27] enforces Gaussian distribution (with diagonal-covariance matrices) for the variational posterior. Since the variational posterior is sufficiently accurate and efficient to sample from, the Cell BLAST model offers a unique opportunity of utilizing the posterior distribution for an improved, manifold-aware cell-to-cell similarity metric (NPD, see Supplementary Figs.…”

“…UMAP 51 and scPhere 35 were performed using Python package umaplearn (v0.3.8) and scPhere (v0.1.0), respectively. ZIFA 52 (v0.1), Dhaka 26 (v0.1), DCA 7 (v0.2.2), scVI 11 (v0.2.3), scScope 53 (v0.1.5), and SAUCIE 54 source code were downloaded from their Github repositories. For ZIFA and scScope, we removed hard-coded random seeds and added options for manually setting them.…”

Enhancing scientific discoveries in molecular biology with deep generative models

“…Applications of DGMs to scRNA‐seq data emerged as a useful way to embed and analyze cells in a low‐dimensional space that summarizes their transcriptomes. Here, the distances between cells in the embedding space can be used to identify phenotypically coherent groups of cells, reflecting either discrete cell types (e.g., T cells, B cells), hierarchies of types (e.g., subtypes of T cells), or variation along some continuum (e.g., progression along the cell cycle) (Ding et al , ; Lopez et al , 2018a; Wang & Gu, ; Amodio et al , ; Eraslan et al , 2019b; Rashid et al , ; Grønbech et al , ). For example, scvis (Ding et al , ) employs a VAE to learn a biologically meaningful two‐dimensional representation of single cells from oligodendroglioma samples.…”

Prediction of Potential miRNA–Disease Associations Through a Novel Unsupervised Deep Learning Framework with Variational Autoencoder

“…Through combing high-throughput cell line assays of drug-induced transcriptomic perturbation effects, the drug response variational autoencoder (Dr.VAE) has been developed to improve drug response prediction [41]. Rashid et al have well applied VAE to unmask tumor heterogeneity from single cell genomic data [42]. Further, Tezcan et al have succeeded in reconstructing magnetic resonance (MR) images from under-sampled measurements with VAE [43].…”