Machine learning analyses of single-cell multi-modal data for understanding cell-type functional genomics and gene regulation

点击翻译按钮获取中文摘要

Project Summary Understanding cell-type-specific gene functions, expression dynamics, and regulatory mechanisms in complex brains is still challenging. To this end, the increasing amount of single- cell multi-modal data in the BRAIN Initiative allows a better understanding of molecular and cellular mechanisms that occur in various cellular phenotypes such as electrophysiology, transcriptomics, and morphology. Many computational methods have thus been applied to integrate such multi-modal data for discovering genes, functions, and cross-modal cell types. However, many of these methods output descriptive results such as differentially expressed genes of various cell types, barely providing functional and regulatory mechanistic insights. The multi-modal data from different studies potentially give rise to inconsistency and bias and lack interpretability for understanding mechanisms. It is crucial to integrate and analyze single cell multi-modal data using coherent, biologically interpretable methods to address these problems. Thus, the objective of this project is to perform machine learning analyses to integrate single- cell multi-modal data in the BRAIN Initiative for predicting the gene functions and gene regulatory networks for cellular phenotypes and improving phenotype prediction. Our machine learning analyses in this project can further serve the BRAIN Initiative project to enable multi- modal data integration and discover functional biomarkers (e.g., genes, regulatory elements, pathways) for various cell types and cellular phenotypes. These cell-type biomarkers will provide an increased understanding of complex brain mechanisms that potentially lead to novel, testable, mechanistic, and translational biological hypotheses. We will have three aims to accomplish this project. In Aim 1, we aim to apply manifold learning analysis to align single-cell multi-modalities and reveal cell trajectories with continuous phenotypic changes such as gene expression and electrophysiology. In Aim 2, we aim to predict cell-type gene regulatory networks for multi-modal characteristics. In Aim 3, we will apply the deep learning analysis to improve cellular phenotype prediction from multi-modal data and prioritize cell-type gene regulatory mechanisms for phenotypes. Finally, all of our analyses will be open source and publicly available as general bioinformatics tools.