Deep Learning-based Method For SCRNA-SEQ Analysis And Downstream Task

Living cells constitute complex systems of interaction between large numbers of biological entities.Genes,transcripts,proteins,metabolites and species perform cellular processes in temporally and spatially controlled ways.Typically,these biological interactions and their parameters cannot be measured directly only from experiments.Since 2013,when single-cell sequencing was recognized as the most promising sequencing method to answer complex biological research queries,the sequencing technologies have advanced,resulting in a significant increase in single-cell data.However,there is a three-fold rise in the computational challenge to treat these massive and complex datasets.Typically,to design a.modeling approach and then the applicability of these algorithms is determined by the use case,particularly the size of the problem.In general,all methods are challenged by the noise in biological measurements,scarce data relative.to model complexity,and nonlinear and combinatorial behavior.However,the required computationally expensive numerical algorithms,must be implemented in computationally efficient ways.By combining Artificial Intelligence(AI)methodologies such as Deep Learning(DL)with biological data obtained from single-cell RNA sequencing(scRNA-seq)technologies,it is possible to reveal the underlying structure of cells and their sub-populations predicted more accurately and effectively.However,this persuasive technology is delicate when concerning biological factors and technical noise.For a variety of reasons,the single-cell data analysis can be computationally intensive,such as imputing missing values in the data,denoising it,accounting for the zero-inflated(ZI)nature of the data and,most challenging of all,reducing dimensionality.Deep learning(DL)is recognized as a competitive choice for single-cell analysis,in addition to traditional Machine Learning(ML)approaches.Deep generative models can possibly learn the underlying construction from omics data,such as pathways or gene systems.We give an introduction and an overview of such analysis techniques,specifically illustrating their usage with single-cell expression data.In the thesis,a unified mathematical formulation is presented of DL models;Variational Auto-encoder(VAE),Autoencoder(AE),and Generative Adversarial Network(GAN),comparing their training steps and strategies and their loss functions,also we relate the loss functions of these model types to particular objects of the single-cell RNA-seq data processing steps.Although,Deep generative models assist to learn the structure of high-dimensional multi-omics and single-omics data by efficiently encapsulating nonlinear dependences between gene pairs;they are often hard to interpret due to their Neural Network(NN)based structure.The thesis’ descriptive style will aid in the selection of appropriate algorithms for each level of the analytic pipeline.The comprehensive survey of recently established models and techniques will serve as a useful information portal for understanding the use of DL for scRNA-seq analysis,as well as inspire its application to a wider range of new challenging tasks emerging in the multi-omics and single-cell sequencing fields.Key Points:1.Single-cell RNA sequencing technology generates millions of transcriptome profiles,allowing biological study for hidden expression functional structures or cell types,more exact prediction of their effects or responses to treatment,or the utilization of subpopulations to cover research gaps.2.This research examines current DL based analysis algorithms for single-cell data in terms of the challenges they tackle and their part in the analysis pipeline.3.Each technique is examined with a mathematical explanation of the three DL models,as well as a discussion of the model’s special attributes.4.Each model in the survey is also presented with a thorough overview of the evaluation measures,comparison algorithms,and datasets.5.Finally,we demonstrated the applicability of the strategies outlined in a real application using single-cell expression data.6.We present a Deep Learning-based pipeline for single-cell RNA-seq downstream analysis using an Auto-encoder for dimensionality reduction.