SCDF: A Novel Single-Cell Classification Method Based on Dimension-Reduced Data Fusion

Abstract

Single-cell RNA sequencing (scRNA-seq) technology has recently brought a new insight into identifying and characterizing novel cell types and gene expression patterns. The study of single-cell RNA sequencing is irreplaceable for the exploration of biology and also faces significant challenges. This has led to the emergence of abundant methods and tools that aim to automatically process specific problems associated with scRNA-seq data. However, a prominent feature of scRNA-seq data is noisy and of high dimensionality due to the large-scale expression datasets with hundreds of thousands of genes. So, the raw single-cell datasets are generally processed for dimensionality reduction before being classified for analysis. There are many methods for dimensionality reduction, such as Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbor Embedding (t-SNE). Each method has different effects on the results of single-cell classification. But, choosing an appropriate method for gene expression data to perform dimensionality reduction in scRNA-seq classification is an unsolved problem. This paper integrated seven dimensionality reduction methods to process gene expression profiles based on data fusion and ensemble learning. Moreover, we determined the optimal number of low-dimensional components for each dimensionality reduction method and each dataset before integration. Different from existing dimensionality reduction techniques, the proposed method implements data fusion and ensemble learning schemes that utilize massive weak learners for accurate classification. Our analysis demonstrates that classification using different data representations may capture a more complete data relationship, and integrating various methods to reduce the dimension of the data promotes better performance for single-cell classification.