An Explainable Unified AI Strategy for Efficient and Robust Integrative Analysis of Multi-omics Data from Highly Heterogeneous Multiple Studies

健康的百岁老人携带保护性变种，这种变种可以抵消与年龄相关的疾病风险变种，前者大多是罕见的。因此，需要通过综合多组学数据分析来发现与异常长寿(EL)相关的标记，以提高检测能力。然而，现有的多组学数据综合分析方法没有对同一模式中和研究之间的标记之间的关系进行建模，从而混淆了来自不同研究的多组学数据所提供的相关信息的有效利用。我们提出了一种统一的人工智能策略，该策略对标记、模态和研究之间的关系进行建模，并通过图神经网络(GNN)学习公共空间中数据的非线性低维表示。我们通过在单个GNN中实现研究表示和表型预测精度之间的相似性最大化来实现深度整合。该提案有三个具体目标：1)开发一个可解释的统一人工智能战略和软件，用于高效和稳健地综合分析来自高度异质多项研究的多组学数据。2)将AIM 1中开发的方法应用于长寿家庭研究(LLFS)和EL联盟提供的综合长寿Omics(ILO)数据，以确定与EL相关的途径和生物标记物。3)将目标1中开发的方法应用于人类组学数据和EL联盟提供的100个不同寿命物种的组学数据，以确定保守的和物种特定的EL相关途径和标记。这项工作的结果将导致一个公开可用的综合组学数据分析软件，该软件不仅能够识别与长寿相关的健壮路径和生物标记物，而且还将适用于任何具有类似组学数据分析需求的复杂疾病研究。我们的工作将对确定改善人类健康的治疗干预措施做出重大贡献。

Healthy centenarians carry protective variants that counteract age-related disease risk variants, the former of which are mostly rare. Therefore, markers associated with exceptional longevity (EL) need be discovered through integrative multi-omics data analysis to improve detection power. However, existing integrative analysis method for multi-omics data do not model the relationships among markers in a modality and among studies, muddying the efficient use of pertinent information provided by multi-omics data from heterogeneous studies. We propose a unified AI strategy that models the relationships among markers, modalities, and studies, and learns nonlinear low-dimensional representations of data in a common space via graph neural networks (GNN). We achieve deep integration by enforcing the maximization of similarities between study representations and the phenotype prediction accuracy in a single GNN. The proposal has three specific aims: 1) Develop an explainable unified AI strategy and software for efficient and robust integrative analysis of multi-omics data from highly heterogeneous multiple studies. 2) Apply the methods developed in Aim 1 to Long-Life Family Study (LLFS) and Integrative Longevity Omics (ILO) data provided by the EL consortium to identify EL-associated pathways and biomarkers. 3) Apply the methods developed in Aim 1 to omics data from human and 100 species of diverse lifespan provided by the EL consortium to identify conserved and species-specific EL-associated pathways and markers. The outcome of this work will result in a publicly available integrative omics data analysis software which not only is able to identify robust longevity-associated pathways and biomarkers, but will also be applicable to any complex disease study with similar omics data analysis demands. Our work will contribute significantly to identify therapeutic interventions for improving human health.