Enhancing AI-readiness of multi-omics data for cancer pharmacogenomics

摘要/摘要 综合性药物基因组学资源的匮乏是发展的重大障碍。 儿童癌症的新疗法。我们的父R00项目试图通过创建和 多因素综合预测儿童肿瘤药物敏感性的深度学习模型验证 组学资料。然而，利用前沿的深度学习模型来分析多组学往往是 由于数据是高维和非结构化的，因此具有挑战性。在父项目下，我们已经评估了 将多组学数据转换为结构化格式的几种嵌入方法，使人工 智能(AI)应用。回应NOT-OD-23-082《支持的行政补充 改进NIH支持的数据的AI/ML就绪性的协作，我们建议补充父级 该项目通过进一步增强用于研究癌症药物基因组学的多组学数据的人工智能准备。我们的 假设是生物引导的非结构化多组学数据的图像嵌入增强了信息 通过深度学习捕获，实现对治疗反应的准确建模和预测。我们的目标是 实现三个相关但独立的目标：1)方法：开发更好的数据转换方法 AI-为癌症多组学做好准备，2)可访问性：使AI-Ready工具和数据更容易为 生物医学研究界，以及3)参与：促进 生物信息学、生物医学工程和生物医学社区。具体地说，目标1将评估一个 全面的一系列具有生物学意义的方法来转换非结构化的多组学数据，包括基因 突变和基因表达谱，到可通过卷积分析的类似图像的数据格式 模特们。我们的方法将嵌入基因的功能相似性，以确保可解释性。在目标2中，我们将 开发一种交互式Web服务器，提供对数据转换工具和人工智能就绪的癌症数据的轻松访问。 最后，在目标3中，我们将在生物信息学旗舰会议上组织一次社区参与活动，以 提高对当前人工智能准备方面差距的认识，并促进临床、基础、 以及计算科学家和实习生。拟议的补充方案汇集了一个协作团队 来自不同学科的专家组成，涵盖癌症生物信息学、基因组学和药物基因组学、人工智能 方法和社区参与活动。成功完成这项补充将有一个 对推进大型癌症数据的人工智能准备工作产生重大影响，与母公司的目标保持一致 R00项目。

Summary/Abstract The scarcity of comprehensive pharmacogenomics resources poses a significant obstacle to the development of new therapies for pediatric cancers. Our parent R00 project seeks to overcome this challenge by creating and validating a novel deep learning model for predicting drug sensitivity of pediatric tumors using integrative multi- omics profiles. However, the utilization of cutting-edge deep learning models to analyze multi-omics is often challenging since the data are high-dimensional and unstructured. Under the parent project, we have evaluated several embedding methods to transform multi-omics data into a structured format that enables artificial intelligence (AI) applications. In response to NOT-OD-23-082 “Administrative Supplements to Support Collaborations to Improve the AI/ML-Readiness of NIH-Supported Data,” we propose to supplement the parent project by further enhancing the AI-readiness of multi-omics data for studying cancer pharmacogenomics. Our hypothesis is that biology-guided image embedding of unstructured multi-omics data enhances the information captured by deep learning, enabling accurate modeling and prediction of treatment responses. We aim to achieve three relevant, but independent, goals: 1) methodology: to develop better data conversion methods for AI-readiness of cancer multi-omics, 2) accessibility: to make AI-ready tools and data more accessible to the biomedical research community, and 3) engagement: to promote collaboration on AI-readiness among the communities of bioinformatics, biomedical engineering, and biomedicine. Specifically, Aim 1 will evaluate a comprehensive array of biologically meaningful ways to transform unstructured multi-omics data, including gene mutation and gene expression profiles, to an image-like data format that can be analyzed by convolutional models. Our approach will embed functional similarities of genes to ensure interpretability. In Aim 2, we will develop an interactive web server that provides easy access to data conversion tools and AI-ready cancer data. Finally, in Aim 3, we will organize a community engagement event at a flagship conference of bioinformatics to enhance awareness of current gaps in AI-readiness and foster collaboration and diversity among clinical, basic, and computational scientists and trainees. The proposed supplement has brought together a collaborative team of experts from diverse disciplines, covering cancer bioinformatics, genomics and pharmacogenomics, AI methodology, and community engagement events. Successful completion of this supplement will have a significant impact on advancing the AI-readiness of large cancer data, aligning with the objectives of the parent R00 project.