III: Medium: Knowledge-Guided Meta Learning for Multi-Omics Survival Analysis

癌症是世界上最致命的恶性疾病之一。生存分析预测事件发生的时间（即癌症的死亡时间），对癌症生物学研究和疾病治疗至关重要。利用肿瘤分子信息根据疾病分期或严重程度对患者进行分层，对于为每位患者量身定制治疗方案至关重要。识别驱动不良生存结果预测的分子特征可以为癌症的机制研究提供见解，并导致新的治疗靶点的发现。高通量基因组学技术的最新进展导致大量高维组学数据可用，这使得应用机器学习方法来建立预测模型，预测肿瘤分期或严重程度，以及确定驱动肿瘤发生和转移的关键分子因素，从而可能导致新的治疗靶点成为可能。虽然深度学习已经应用于癌症生存分析，但它在每个队列中具有高维特征的样本数量相对较少的生物数据（即众所周知的大p，小n问题）上并不是很有效。癌症基因组图谱（TCGA）就是这样一个例子，它具有跨越许多癌症类型的多层次高维临床和分子图谱的特征。为了使机器学习有效地处理少量的训练数据，例如TCGA中的每种癌症类型，我们提出了一套先进的机器学习方法来解决癌症生存分析中众所周知的大p，小n问题。本项目开发了一种新的以知识为导向的元学习框架，创新地将生物学知识与元学习相结合，用于多组学生存分析。该框架解决了多个领域（如癌症）但每个领域样本数量少的情况下带来的一系列技术挑战，并利用少次学习的原则，通过训练一组不同的癌症来实现对新癌症样本的快速适应过程。提出的知识引导的嵌入学习框架制定了一个新的基本结构，可以将生物学知识整合到潜在患者表征的构建中，从而利用人类管理的知识库中存在的丰富信息，如基因本体和可在不同癌症之间共享的调控网络。提出的生存分析元学习方法能够预测未知癌症患者的生存概率。为了进一步加强生存预测的元学习，我们开发了一个名为“具有元知识的上下文感知学习”的元学习框架，该框架明确地将经过整理的领域知识纳入元学习框架，并意识到预测上下文，包括预测难度和输入患者样本的异质性。提出的知识引导元学习范式在适应新任务或新领域方面明显比传统迁移学习更强大。该项目将带来一种新颖、先进的方法来分析各种癌症数据集，每种癌症都有一个小样本量。因此，本项目对于推进少镜头学习的理论和实践具有重要的潜力，具有很强的社会意义。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Cancers are among the world's most deadly malignant diseases. Survival analysis, which predicts the time to an event (i.e., death in the case of cancer), is of crucial importance for cancer biology research and disease treatment. Stratifying patients according to disease stage or severity using molecular information from the tumor is critical for tailoring treatments for each patient. Identifying molecular features that drive the prediction of poor survival outcomes could provide insights into mechanistic studies of cancer and lead to discoveries of new therapeutic targets. Recent advances in high-throughput genomics technologies have led to a massive amount of high-dimensional omics data being available which makes it possible to apply machine learning approaches to build predictive models that predict tumor stage or severity as well as identifying the key molecular factors driving tumorigenesis and metastasis which could lead to new therapeutic targets. While deep learning has been applied to cancer survival analysis, it does not work very well on biological data with a relatively small number of samples with high-dimensional features in each cohort (i.e., the well-known big p, small n problem). The Cancer Genome Atlas (TCGA) is such an example, which characterizes multi-level high-dimensional clinical and molecular profiles spanning many cancer types. To make machine learning effectively work on a small amount of training data such as each cancer type in TCGA, we propose a set of advanced machine learning approaches to tackle the well-known big p, small n problem for cancer survival analysis. This project develops a new knowledge-guided meta-learning framework which innovatively integrates biological knowledge with meta-learning for multi-omics survival analysis. This framework tackles a series of technical challenges posed by the scenario of many domains (e.g., cancer) but small numbers of samples in each domain and utilizes the principle of few-shot learning to implement a fast adaptation process to new cancer samples by training of a set of different cancers. The proposed knowledge-guided embedding-learning framework formulates a new fundamental structure that enables the integration of biological knowledge into the construction of a latent patient representation, which exploits the rich information present in human-curated knowledge bases such as gene ontology and regulatory networks that can be shared across different cancers. The proposed meta-learning for survival analysis approach enables the prediction of survival probability for patients from unseen cancers. To further enhance meta-learning for survival predictions, a meta-learning framework called Context-Aware Learning with Meta-Knowledge is developed, which explicitly incorporates curated domain knowledge into the meta-learning framework, and is aware of the prediction context, including the prediction difficulty and heterogeneity of the input patient samples. The proposed paradigm of knowledge-guided meta-learning is significantly more powerful than traditional transfer learning in adapting to new tasks or domains. The project will lead to a novel, advanced approach to analyzing a variety of cancer datasets with each cancer having a small sample size. Thus, this project has significant potential to advance the theory and practice of few-shot learning, with strong social implications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.