Evaluation of Computational Methods for Single Cell Multi-Omics Integration

Abstract

Humans evolve from a single-cell to 37 trillion cells. Next-generation sequencing advances help the sequencing of cells at the resolution of single-cell. Single-cell datasets are produced at various geographical locations, time-points, laboratories, and by distinct experimenters, using different sequencing machines and protocols. Integration of different samples can reveal more insights into the interplay of cells that define the state and function of tissues and organs. Several methods for integratiion of single-cell datasets have been developed. Choosing an appropriate integration method for the task at hand is a difficult decision to make. In this paper, a comparitive evaluation of five integration methods is presented using two batches of gene expression and chromatin accessibility data. The methods are evaluated for three tasks: integration and batch removal, cell-type clustering, and trajectory conservation using five evaluation metrics. For cell-type clustering, this study finds LMDS and MNN perform well, while preserving biology. Furthermore, trVAE outperforms all other models in batch effect removal and trajectory conservation (TC) tasks, where as totalVI does not perform well across all the tasks.