Classification of RNA biological cells to their cycle stages, using a stacked autoencoder neural network
Single cell RNA-sequencing concentrates on extracting genetic information from that cell including expression profile, in which millions of reads are generated to indicate whether some gene is expressed in a cell and to what extent. This data is used in research and as input data to health systems. The evolution of each cell consists of three phases: G1, S, G2.
We use the cell read count of 38293 genes in each of the 288 cells (96 samples of each cycle phase). • We remove all genes with read count of 0 in all cells. • Remove all genes that are expressed (read count greater than 0) only in less than 20% of the cell samples. Choosing greater percentage of samples for considering the expressiveness of a gene (for example, 60% as was initially intended) can eliminate genes that are related to one specific cell cycle phase, and are most meaningful for classification. • Normalize the data by applying logarithm on all values of read counts.
We build a network consisting of three fully connected dense hidden layers of 3000, 1000 and 30 neurons, respectively. After fitting the auto-encoder on the designed model, we froze the weights of the trained layers and replaced the last reconstruction layer. We measure the accuracy of classificaion.
The GUI enables to upload the data files, consisting of cell read counts, and present the evaluation results of the classification. The GUI is implemented using the Python TKinter package.
