Comparative Analysis of Fusion Techniques for Integrating Single-cell Multiomics Datasets

Abstract

In this study, the performance of different fusion techniques for integrating single-cell multi-omics datasets is compared. Integrated measurements enable the unveiling of complex interaction networks between cells and heterogeneous structures that remain unseen in a single omics type. The applied fusion technique for integration alters the scope and characteristics of the integrated measurements. Variational auto encoder (VAE) models incorporating early, intermediate, and late fusion techniques were compared in this study in order to compare the effect of different fusion techniques. CITE-seq datasets containing proteomic and transcriptomic measurements from single-cells are utilized in experiments. Both computational and biological performance metrics were employed for comparing the integrated measurements. According to the experimental results, the fusion models improved the silhouette score of the raw data between 10% and 23%. The late fusion model with the highest performance in computational metric graph connectivity improved the performance of the raw data by 4%. The results demonstrate that the late fusion technique outperforms its competitors for integrating single-cell multi-omics datasets.