A visible machine learning system to discover targeted treatment solutions in cancer

项目概要/摘要 了解遗传相互作用可以通过靶向治疗来设计针对个体癌症患者的治疗方案。 癌症基因组中特定的遗传脆弱性。例如，通过识别基因对， 当同时被击倒时（与单独的击倒相比），适应性缺陷，可以选择性杀死 通过抑制一种蛋白质的合成-致死配偶体而使其发生功能丧失突变的癌细胞。尽管 产生描绘肿瘤转录组、蛋白质组、代谢组、成像等的大规模数据， 关于不同的基因如何相互作用知之甚少，也不清楚如何设计 基于可用的组学数据的靶向治疗。为了应对这些挑战，建议进行研究。 将开发一个“可见的”机器学习框架，以系统地了解高阶遗传学。 研究癌症中的相互作用（即二基因和三基因相互作用）并设计靶向治疗。 所提出的框架的第一步是通过结合 可用的组学数据已经应用了多种方法来集成类似形式的数据，但是 缺乏有效的解决方案来集成质量和格式差异很大的数据。为了应付这一挑战， 岳琴开发了一种方法来推断一个分层的癌细胞图谱，在多个不同的细胞中捕获癌症通路。 通过融合免疫荧光（IF）成像数据和亲和纯化-质谱（AP- MS）。 在拟议研究的F99阶段，通过将深度神经网络的架构与 Yue将开发一种“可见”神经网络（VNN），可以预测癌细胞 来自遗传扰动（即敲除）和基因组背景（即突变）的适应性，同时提供 对基因型-表型预测至关重要的癌症途径的机制见解。 在K 00阶段的奖励，岳将开发基因工程方法，以实验 基于从VNN获得的机制见解，绘制癌细胞中的高阶遗传相互作用 在基因型-表型预测中。实验产生的数据可以直接启发靶向治疗 的设计.此外，新的数据可以整合到分层癌细胞图谱中，以提高准确性 和推断途径的分辨率，从而进一步提高VNN在基因型-表型中的“可见性”。 预测. F99阶段的计算集中训练与实验集中训练相结合 在K 00阶段，将为Yue领导自己的癌症生物学跨学科研究做好充分准备。此外，本发明还提供了一种方法， 个性化的培训计划，包括指导和教学，科学写作和口语 演讲将确保岳获得必要的技能，她未来的建立作为一个独立的 调查员

Project Summary/Abstract Understanding of genetic interactions can lead to therapeutic design for individual cancer patients by targeting the specific genetic vulnerability in the cancer genome. For example, by identifying gene pairs that pose severe fitness defects when knocked out simultaneously (compared to separate knockouts), one can selectively kill cancer cells that harbor loss-of-function mutation in one protein by inhibiting its synthetic-lethal partner. Despite generation of large-scale data delineating the tumor transcriptome, proteome, metabolome, imaging, and so on, little is known regarding how different genes interact with each other and it is unclear how one can design targeted treatments based on the ‘omics data available. To address these challenges, the proposed research will develop a “visible” machine learning framework to systematically understand the higher-order genetic interactions (i.e. di-genic and tri-genic interactions) in cancer and design targeted treatments. The first step for the proposed framework is to gain a holistic view of cancer pathways through combining the ‘omics data available. Multiple approaches have been applied to integrate data of similar forms, but there yet lacks an effective solution for integrating data of vastly different qualities and formats. To address this challenge, Yue Qin has developed a method to infer a hierarchical cancer cell map capturing cancer pathways at multi- scale resolution by fusing immunofluorescence (IF) imaging data and affinity purification-mass spectrometry (AP- MS). During the F99 phase of the proposed research, by tying the architecture of a deep neural network to the hierarchical cancer cell map, Yue will develop a “visible” neural network (VNN) that can predict the cancer cell fitness from genetic perturbation (i.e. knockouts) and genomic backgrounds (i.e. mutations) while providing mechanistic insights in cancer pathways critical for genotype-phenotype prediction. During the K00 phase of the award, Yue will develop genetic engineering approaches to experimentally map higher-order genetic interactions in cancer cells based on the mechanistic insights obtained from VNN during genotype-phenotype prediction. The data generated experimentally can directly inspire targeted treatment designs. In addition, the new data can be integrated into the hierarchical cancer cell map to improve accuracy and resolution of the inferred pathways, thus further improving the “visibility” of VNN in genotype-phenotype prediction. The combination of a computational focused training during F99 phase and experimental focused training during K00 phase will fully prepare Yue leading her own interdisciplinary research in cancer biology. In addition, the personalized training plan covering aspects including mentoring and teaching, scientific writing, and oral presentation will ensure Yue acquiring skills necessary for her future establishment as an independent investigator.