AI-Powered Uncovering of Mechanisms in Cancer Through Causal Discovery Analysis and Generative Modeling of Heterogeneous Data

项目总结 这份提案概述了一项为期五年的研究和职业发展计划，旨在建立计算机 了解癌症中扰动和体细胞变化的表型效应的框架。这个 申请在很大程度上是基于候选人在卡内基梅隆大学接受的广泛的博士培训- 著名的计算机科学系。这也是基于应聘者丰富的工作经验 作为布罗德研究所的一名副计算生物学家，以及他庞大的顶级医生网络 以及癌症领域的科学家。它还利用了他目前在Gad Getz博士手下的博士后职位 博德研究所，以及该研究所独特的资源、设施、合作和专业知识。 与一系列相关的教学和职业发展活动一起，这些研究将构成HIS的基础 在实现长期目标的指导下，过渡到独立的终身教职轨道科学家职位 将癌症作为一种疾病进行建模和理解。下一代测序的大规模可用性 癌症的数据为这种疾病中发生的躯体变化提供了前所未有的特征。 了解它们的组合表型效应仍然是一个悬而未决的问题，而且在体外具有强大的扰动作用。 已经设计了一些协议来实验地探测这些影响。然而，可能的搜索空间 筛查的扰动组合大得令人望而却步。这项工作的目标是提供原则性的 人工智能(AI)驱动的推断扰动和观察到的体细胞影响的方法 癌症的改变，这是理解其机制的关键一步。拟议的工作借鉴了最近的 在机器学习和因果发现技术领域的发展。特别是，有两个具体目标 将被评估：(目标1)从单细胞RNA-SEQ推断因果图(可选择将其与 全外显子组/全基因组测序)；(目标2)使用深度繁殖模型，以及配对的全基因组测序 外显子组/全基因组测序，以了解单细胞RNA序列中潜在的变异因素。这个 拟议的工作还包括验证这些计算目标的步骤。完成后，这项工作将 通过可用于以下方面的算法/资源推进该领域：(1)使用因果知识进行计算 选择要在实验室测试的目标组合；以及(2)通过计算推断体细胞DNA的影响 表达兴趣的改变，导致下游实验设计的改进。因此，把 总之，提出的目标是理解癌症机制的关键一步，并将导致 在有效发现治疗这种疾病的药物方面取得了重大进展。

PROJECT SUMMARY This proposal outlines a five-year research and career development program aimed at building computational frameworks for understanding the phenotypic effects of perturbations and somatic alterations in cancer. The application is heavily based on the candidate’s extensive PhD training in Carnegie Mellon University’s world- renowned Computer Science Department. It is also grounded in the candidate’s rich prior experience working as an Associate Computational Biologist at the Broad Institute, and his large network of top-level physicians and scientists in the cancer field. It also leverages his current postdoctoral appointment under Dr. Gad Getz at the Broad Institute, and the unique set of resources, facilities, collaborations and expertise in this institute. Along with a series of relevant didactics and career building activities, these studies will form the basis of his transition to an independent tenure track position as a scientist guided by the goal of enabling long-term modeling and understanding of cancer as a disease. The large-scale availability of next-generation sequencing data for cancer has offered an unprecedented characterization of somatic changes that happen in this disease. Understanding their combinatorial phenotypic effects is still an open problem, and powerful in vitro perturbation protocols have been designed to experimentally probe these effects. However, the search space for possible combinations of perturbations to screen is prohibitively large. The objective of this work is to provide principled Artificial Intelligence (AI)-driven methodology for inferring the effects of perturbations and observed somatic alterations in cancer, a crucial step in understanding the mechanisms. The proposed work draws on recent development in the technical fields of machine learning and causal discovery. In particular, two Specific Aims will be evaluated: (Aim 1) inferring causal graphs from single-cell RNA-seq (with the option of pairing it with whole-exome/whole-genome sequencing); (Aim 2) using a deep generative model, along with paired whole- exome/whole-genome sequencing, to learn latent underlying factors of variation in single-cell RNA-seq. The proposed work also includes steps to validate these computational aims. When completed, this work will advance the field via algorithms/resources that can be used to: (1) use causal knowledge to computationally select combinations of targets to test in the lab; and (2) computationally infer the effects of somatic DNA alterations of interest on expression, leading to improved downstream experiment design. Therefore, put together, the proposed aims are a crucial step in understanding mechanisms in cancer, and will lead to significant progress towards efficiently discovering drugs for this disease.