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

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Intra-tumor heterogeneity complicates understanding tumor evolution.
• Single-cell sequencing offers insights but generates noisy, sparse data.
• Extracting evolutionary features from single-cell genomic data is computationally challenging.

Purpose of the Study:

• To introduce Dhaka, a novel variational autoencoder method.
• To transform single-cell genomic data into a reduced-dimension feature space.
• To improve the differentiation of hidden tumor subpopulations and infer evolutionary trajectories.

Main Methods:

• Developed Dhaka, a variational autoencoder.
• Applied Dhaka to single-cell DNA sequencing (scDNA-Seq) copy number variation and single-cell RNA sequencing (scRNA-Seq) gene expression data.
• Tested on synthetic and six real single-cell cancer datasets.

Main Results:

• Dhaka effectively reduces dimensionality and differentiates tumor subpopulations.
• Identified marker genes for revealed subpopulations.
• Successfully inferred lineage and differentiation trajectories, outperforming existing methods.
• Applicable to various genomic data types.

Conclusions:

• Dhaka provides an efficient feature extraction and dimensionality reduction method for single-cell genomic data.
• The method aids in understanding tumor heterogeneity and evolution.
• Dhaka is publicly available with supporting information on GitHub.