An Accurate Machine Learning Framework for Childhood Acute Myeloid Leukemia Subtype Identification by Integrating Bulk and Single-Cell Multi-Omics Data Within and Beyond the CCDI Ecosystem

摘要 作为一种致命的儿童血液系统恶性肿瘤，其特征是未成熟髓系克隆性扩张 作为先兆，急性髓系白血病(AML)通常会导致骨髓衰竭和造血功能受损。 AML有多个不同的亚型，以形态、分子和遗传改变为特征。 识别AML亚型可以促进下游风险分层和量身定制的治疗设计。虽然有各种各样的 常规方法，如形态分析、细胞遗传学分析、免疫表型分析或分子生物学 特征分析已被用于AML亚型识别，它们通常是昂贵的、耗时的、劳动密集型的， 有时是不准确的。最近的进展见证了下一代测序(NGS)的应用 用于识别AML亚型，但它们仅限于批量NGS数据或单个组学数据。伴随着大量的组学 儿童癌症数据倡议(CCDI)生态系统内外产生的数据，我们 假设单细胞和大量多组学数据的整合包括基因组学、转录组学和 表观遗传学数据将大大促进亚型特异性生物标记物的发现，并提高AML的准确性 子类型标识。在我们的父母奖(CA036727)下，在这个补充项目中，我们建议 为准确且经济高效的AML亚型开发一个集成的机器学习(ML)框架 通过结合CCDI生态系统内外的批量和单细胞多组学数据进行鉴定。 为了实现这一目标，我们计划实现两个具体目标。在目标1中，我们将建立一个知识转移ML 利用大规模批量和单细胞转录组数据识别AML亚型的模型。此外 识别注释良好的AML亚型，我们还将通过检测稀有细胞类型来探索新的AML亚型 从大规模的单细胞数据中，可以从这些数据中提取特定于簇的和特定于稀有细胞类型的基因签名 转移到批量转录数据，以提高AML亚型识别的性能。在目标2中， 我们将开发多核学习和多模式深度学习框架，以系统和 自动集成来自单细胞和批量多组学数据的与AML亚型相关的深层信息 (包括基因组学、转录组学、表观基因组学)，进一步促进AML亚型鉴定。我们的模型是 灵活地处理只有部分或不完整的多组学数据可用于新患者的病例。我们 我相信这项研究的成功完成将对改善下游儿童急性髓细胞白血病产生直接影响 风险分层，便于诊断和预后，并优化治疗选择。我们还预计， 我们在这项研究中提出的框架可以定制并可扩展到识别其他儿科疾病的亚型， 青春期和青壮年(Aya)癌症，特别是超罕见的肿瘤。

Abstract As a fatal childhood hematopoietic malignancy characterized by clonal expansion of immature myeloid precursors, acute myeloid leukemia (AML) usually leads to bone marrow failure and impaired hematopoiesis. AML has multiple distinct subtypes characterized by morphological, molecular, and genetic alterations. Identifying AML subtypes can facilitate downstream risk stratification and tailored treatment design. While various conventional methods like morphological analysis, cytogenetic analysis, immunophenotyping, or molecular profiling have been used for AML subtype identification, they are usually costly, time-consuming, labor-intensive, and sometimes inaccurate. Recent progress has witnessed the application of next generation sequencing (NGS) for identifying AML subtypes, but they are limited to bulk NGS data, or single omics data only. With tons of omics data being generated within and beyond the Childhood Cancer Data Initiative (CCDI) ecosystem, we hypothesize that integration of single-cell and bulk multi-omics data including genomics, transcriptomics, and epigenetics data will significantly facilitate subtype-specific biomarker discovery and boost the accuracy of AML subtype identification. Under our parent award (CA036727), in this supplemental project, we propose to develop an integrated machine learning (ML) framework for accurate and cost-effective AML subtype identification by combining bulk and single-cell multi-omics data within and beyond CCDI ecosystem. To achieve this, we plan to undertake two specific aims. In Aim 1, we will establish a knowledge-transfer ML model that leverages large-scale bulk and single-cell transcriptomics data for AML subtype identification. Besides identifying well-annotated AML subtypes, we will also explore novel AML subtypes by detecting rare cell types from large-scale single cell data, from which cluster-specific and rare-cell-type specific gene signatures can be transferred to the bulk transcriptomics data for improving performance of AML subtype identification. In Aim 2, we will develop a multi-kernel learning and a multi-modal deep learning framework to systematically and automatically integrate deep information related with AML subtypes from single-cell and bulk multi-omics data (including genomics, transcriptomics, epigenomics) to further boost AML subtype identification. Our model is flexible to tackle cases when only partial or incomplete multi-omics data are available for new patients. We believe successful completion of this study will have direct impacts on improving downstream childhood AML risk stratification, facilitating diagnosis and prognosis, and optimizing treatment selection. We also expect that our proposed framework in this study can be customized and extensible to identifying subtypes of other pediatric, adolescent, and young adult (AYA) cancers especially ultra-rare tumors.