Integrative analysis of single-cell multi-omics data with interpretable deep learning methods

单细胞多组学测序技术(如RNA-seq和ATAC-seq)的最新进展提高了在单细胞水平上量化复杂生物系统的转录组和表观基因组的敏感性，为检测细胞异质性和理解疾病发病机制提供了新的见解。新兴的多组学技术使用不同的平台和模式在多个组织中的不同细胞类型中生成丰富的异质数据集。对多组学数据集的综合分析是非常需要的，对于揭示单细胞潜在的生物学过程和疾病发病机制至关重要。然而，嘈杂的单细胞数据的海洋给生物学家和临床医生带来了前所未有的挑战，需要新颖的大规模机器学习算法来应对这些挑战。拟议的多学科项目的主要目标是连接遗传生物学和计算机科学领域，以开发一流的、大规模和可解释的机器学习工具，以整合不同技术平台和模式的异质单细胞组学数据集。我们将使用细胞到细胞、细胞到基因和基因到基因图(即相互作用网络)形式的图形卷积网络来模拟单细胞多组学数据。图形卷积网络(GCN)是一种新型的端到端深度神经网络，它学习图形上的卷积高阶结构，并自动学习节点(即细胞或基因)的表示，从而支持下行应用，如定义新的细胞类型和检测驱动基因。最近在稀疏图上应用卷积算子的建立揭示了深度学习算法分析单细胞数据的潜在方向。我们建议扩展这种方法，以允许联合细胞类型检测和差异基因选择。我们还将在我们的方法中纳入解释性人工智能，使生物学家和临床医生能够了解网络的整合过程。我们建议在私有数据集和公共联盟的单细胞多组学数据集(如Human Cell ATLAS)上验证我们的算法。考虑到如此多的细胞跨越如此多的异质技术平台和模式，我们提出的基于深度神经网络的方法将展示其可扩展性、更高的准确性和与许多现有计算工具相比的可解释性。这项研究项目具有革命性的单细胞多组学数据集的综合分析的潜力，帮助生物学家和临床医生更好地了解单细胞的生物过程和疾病发病机制，最终使我们更接近造福于个别患者的精准医学。

Recent advances in single-cell multi-omics sequencing (e.g., RNA-seq and ATAC-seq) technologies have increased the sensitivity in quantifying both transcriptome and epigenome landscapes of complex biological systems at the single-cell level, providing novel insights into detecting cell heterogeneity and understanding disease pathogenesis. The emerging multi-omics technologies generate copious heterogenous datasets across different cell types in multiple tissues using various platforms and modalities. Integrative analysis of multi-omics datasets is in great demand and is vital to unveil the underlying biological processes and disease pathogenesis of single cells. However, the ocean of noisy single-cell data poses unprecedented challenges to biologists and clinicians, calling for novel and large-scale machine learning algorithms to address these challenges. The main goal of the proposed multi-disciplinary project is to bridge the fields of genetic biology and computer science to develop a first-in-class, large-scale and interpretable machine learning tool to integrate heterogeneous single-cell omics datasets across different technological platforms and modalities. We will model single-cell multi-omics data using graph convolutional networks in the form of cell-to-cell, cell-to-gene, and gene-to-gene graphs (i.e. interaction networks). Graph convolutional network (GCN) is a new end-to-end deep neural network that learns convolutional high-order structures on graphs and automatically learn representations of nodes (i.e., cells or genes), therefore enabling down-streaming applications such as defining novel cell types and detecting driver genes. Recent establishment of applying convolutional operators on sparse graphs sheds light on potential directions of deep learning algorithms to analyze single-cell data. We propose to extend this method to allow for joint cell-type detection and differential gene selection. We will also incorporate explanatory artificial intelligence in our method that will enable biologists and clinicians to understand the integration processes of the networks. We propose to validate our algorithms on a private dataset as well as public consortium single-cell multi-omics datasets such as Human Cell ATLAS.   Given such a large number of cells across so many heterogeneous technological platforms and modalities, our proposed method based on deep neural networks will demonstrate its scalability, higher accuracy and interpretability over many existing computational tools. This research project has the potential of revolutionizing the integrative analysis of single-cell multi-omics datasets, helping the biologists and clinicians better understand the biological process and disease pathogenesis of single cells, and ultimately moving us closer to precision medicine that benefits individual patients.