Interpretable Deep Learning Methods to Investigate Genetics and Epigenetics of Alzheimer's Disease at a Single-Cell Resolution

阿尔茨海默病和相关痴呆(ADRD)是一种复杂的多因素疾病，其特征是 老年人的渐进性记忆丧失、神志不清和认知能力受损。除了……之外 遗传变异，研究已经报道了某些表观遗传、网络和基因组组织 扰动及其复杂的相互作用有助于ADRD的进展，通知新的细胞 病因学。最近的单细胞革命，特别是多模式基因组图谱，使之成为可能 以最好的解决方案仔细审查ADRDS中的多尺度失调。然而，几乎没有人 已经开发了方法来解决由于高度缺失而具有挑战性的这一关键任务， 维度，以及单元格数据中复杂的特征交互作用。在这个项目中，我们将开发 可解释的深度学习方法和软件工具，以突出多尺度失调 导致ADRDS，包括遗传、表观遗传、网络结构和染色质结构改变 单元格分辨率。 与以前的工作不同，报告了一组一维(1D)功能顺式调节 只从一个基因组中提取元素(CRE)并将其应用于所有样本，我们的目标是构建个人、 紧凑的，以基因为中心的，特定细胞类型的脑调节组，来自sc-Multiome数据。具体来说， 我们将首先提出一种可扩展的多模式深度生成模型，以集成大规模、 具有单通道、多通道和混合通道的异类ADRD单单元数据。与现有的不同 方法，我们将包括一个不变表示学习方案来派生潜伏细胞 与混杂因素无关的表征(例如，年龄、性别、阅读深度和批次效应) 用于无偏见转录组和表观基因组重建(目标1)。然后，我们将超越一维 通过破译多尺度基因调控密码(AIM 2)进行基因组注释，包括细胞类型- 用于功能解释的特定染色质区隔、Cres及其靶基因以及 转录因子(Tf)调控网络(TRN)。最后，我们将开发可解释的深度学习 将多尺度失调与ADRD联系起来的模型和机械解释(目标3)。 这项建议是建立在张、元和 Gerstein实验室源于ENCODE和EPECENCODE项目，具有不同的专业知识 计算机科学、神经科学和基因组学。完成后，我们的提案将显著 通过为研究提供必要的工具，加速更广泛的科学界的研究 并对ADRD的多尺度风险因素进行优先排序。

Alzheimer's disease and related dementias (ADRDs) are complex multifactorial disorders characterized by progressive memory loss, confusion, and impaired cognitive abilities in older adults. In addition to genetic variants, studies have reported that certain epigenetic, network, and genome organizational perturbations, and their complex interplay, contribute to ADRD progression, informing new cellular etiologies. The recent single-cell revolution, especially multimodal genomic profiling, makes it possible to scrutinize multi-scale dysregulations in ADRDs at the finest possible resolution. However, few methods have been developed to address this critical yet challenging task due to the high missingness, dimensionality, and complex feature interactions in single-cell data. In this project, we will develop interpretable deep learning methods and software tools to highlight multi-scale dysregulations contributing to ADRDs, including genetic, epigenetic, network, and chromatin structural alterations at a single-cell resolution. Distinct from previous efforts reporting a set of one-dimensional (1D) functional cis-regulatory elements (CREs) from only one genome and applying it to all samples, we aim to construct personal, compact, gene-centric, and cell-type-specific brain regulome from sc-multiome data. Specifically, we will first propose a scalable multimodal deep generative model to integrate large-scale, heterogeneous ADRD single-cell data with single-, multi-, and hybrid modalities. Distinct to existing methods, we will include an invariant representation learning scheme to derive latent cell representations uncorrelated with confounding factors (e.g., age, gender, read depth, and batch effects) for bias-free transcriptome and epigenome reconstruction (Aim 1). Then, we will go beyond the 1D genome annotation by deciphering the multi-scale gene regulation code (Aim 2), including cell-type- specific chromatin compartmentation, CREs and their target genes for functional interpretation, and transcription factor (TF) regulatory networks (TRNs). Lastly, we will develop interpretable deep learning models to link multi-scale dysregulations to ADRD with mechanistic explanation (Aim 3). This proposal is built on an existing multi-year collaboration among the Zhang, Won, and Gerstein labs that originated from the ENCODE and PsychENCODE projects, with diverse expertise in computer science, neuroscience, and genomics. Upon completion, our proposal will significantly accelerate research in a broader scientific community by providing essential tools to investigate functional regions in the genome and prioritize multi-scale risk factors for ADRD.