Random forest of imputation trees (RITS) for sparse single cell genomics data

Abstract

A human body has billions of cells specialized with their own function and each cell carries genome in its nucleus. The activity of the genome is controlled by a multitude of molecular complexes called as epigenome. Previously scientists had a notion that human diseases are caused only due to changes in the DNA sequence or through the infectious agents present in the environment. However, recent studies have revealed that changes in the epigenome are also associated with disease. Our aims is to create an imputation method for noisy, sparse and highly unbalanced single cell epigenome data. This problem is challenging as there is no imputation method for imputing huge and unbalanced dataset of single cell epigenome. Moreover, its analysis holds a significant amount of importance in the biological domain for preventing and curing many critical diseases. Here we propose an imputation method called as RITs for imputing single cell epigenome profiles. We evaluated our proposed method through various possible techniques and compared its results with traditional imputation methods, although those imputation methods were made for imputing gene expression data. Our proposed method out-performs in every test and comes out as reliable imputation method even when we have huge unbalanced data. We tested our method on scATAC-seq dataset of cells from organs of the adult mouse to check the robustness and efficiency of this method. In all the conditions and tests, our imputation methods RITS remained at the top. The generality of RITs and it robustness for very noisy and sparse data-sets hints that it is the next generation imputation method for single cell profiles.