scMinerva: an Unsupervised Graph Learning Framework with Label-efficient Fine-tuning for Single-cell Multi-omics Integrated Analysis

Abstract

The development of single-cell multi-omics technologies profiles DNA, mRNA, and proteins at a single-cell resolution. To meet the demand, we present scMinerva for single-cell multi-omics integration utilizing graph convolutional networks and a new random walk strategy, which outperforms existing methods on various datasets. Our method is especially robust on high-noise more-omics data and is lightweight concerning speed and memory. scMinerva can effectively perform downstream tasks, such as biomarker detection and cell differentiation analysis. We extensively interpret the robustness of scMinerva by analyzing components’ occurrence frequency in walks during training at omics level, cell-type level, and single-cell level.