SCAFG: Classifying Single Cell Types Based on an Adaptive Threshold Fusion Graph Convolution Network

Abstract

Single-cell RNA sequencing (scRNA-seq) technology has been a significant direction for single-cell research due to its high accuracy and specificity, as it enables unbiased high-throughput studies with minimal sample sizes. The continuous improvement of scRNA-seq technology has promoted parallel research on single-cell multi-omics. Instead of sequencing bulk cells, analyzing single cells inspires greater discovery power for detecting novel genes without prior knowledge of sequence information and with greater sensitivity when quantifying rare variants and transcripts. However, current analyses of scRNA-seq data are usually carried out with unsupervised methods, which cannot take advantage of the prior distribution and structural features of the data. To solve this problem, we propose the SCAFG (Classifying Single Cell Types Based on an Adaptive Threshold Fusion Graph Convolution Network), a semi-supervised single-cell classification model that adaptively fuses cell-to-cell correlation matrices under various thresholds according to the distribution of cells. We tested the performance of the SCAFG in identifying cell types on diverse real scRNA-seq data; then, we compared the SCAFG with other commonly used semi-supervised algorithms, and it was shown that the SCAFG can classify single-cell data with a higher accuracy.

Omics Data and Data Representations for Deep Learning-Based Predictive Modeling

“…Gene expression datasets are sometimes used in conjunction with gene interaction graphs to train the Graph Convolutional Networks (GCN). In [ 67 ], the researchers wanted to classify the cell types from their gene expression. They procured the gene expression profiles for various cells, and then used them to construct a cell similarity matrix by measuring the cosine similarity among the expression levels of the different cells.…”