Interpretable Machine Learning Approaches Applied to Omics Datasets

近几十年来，测序技术的进步掀起了一场革命，导致了基因数据的爆炸性增长，推动人类基因组学进入了大数据时代。最近，新的生物技术使我们能够在分子水平上获得每个个体的信息，例如代谢物(抗氧化剂、维生素等小分子)的浓度或RNA转录本和蛋白质的定量。这些所谓的“组学”数据集，结合出生时建立的遗传学，有望揭示人类之间差异的分子原因，如身高、体重或患疾病的风险。与此同时，人工智能的最新进展导致了从大数据集做出预测的强大方法的开发，涉及从自动驾驶到自然语言处理的多个领域。然而，关于个体组学签名的机器学习技术正在落后，因为仍有几个挑战需要解决。首先，这些“黑箱”方法在不提供解释的情况下产生预测，使专家无法获得必要的证据来验证结果。其次，目前的方法倾向于“背诵”数据集，而不是从数据集中提取一般知识。事实上，尽管我们的数据集很大，但它们包含的参与者比为每个参与者测量的变量要少得多，这降低了推广的能力。最后，必须考虑生物和技术差异的来源，从一个数据集到另一个数据集通常是不一致的，以便在现实世界中获得可靠的预测。我的研究计划提供了具体的解决方案，使这些方法适用于组学数据，针对特定的生物学问题。其中一项研究旨在通过个体的遗传学来预测RNA转录物和代谢物水平的变化。另一个问题是根据个人的组学数据预测罹患复杂疾病的风险。对于这些具体应用，我们将使用来自几个生物库的组学数据，包括本地(蒙特利尔心脏研究所生物库)、国家(CanPath队列)和国际(英国生物库)队列。我们将开发这些方法，同时确保获得可解释和合理的结果，这些结果具有很好的普遍性，并且不受噪声源的影响。我们的项目是跨学科的，为学生提供了丰富的培训机会。我们的结果将有可能改善机器学习在分子生物学研究中的适当使用，为研究人员和加拿大行业提供可由人类解释的强大的组学数据分析工具。

In recent decades, advances in sequencing technologies have set off a revolution resulting in an explosion of genetic data, propelling human genomics into the era of big data. Even more recently, novel biotechnologies allow us to obtain information at the molecular level for each individual, such as the concentrations of metabolites (small molecules such as antioxidants, vitamins) or the quantification of RNA transcripts and proteins. These so-called `omics' data sets, combined with genetics established at birth, hold the promise to reveal the molecular causes of the differences between humans, for traits such as height, weight or risk of developing disease. In parallel, recent advances in artificial intelligence have led to the development of powerful methods for making predictions from large data sets, in several areas from autonomous driving to natural language processing. However, machine learning technologies on individual omics signatures are lagging behind, because several challenges still need to be addressed. First, these "black box" methods produce predictions without providing interpretations, preventing experts from getting the evidence necessary to validate the results. Second, current methods tend to learn the datasets "by heart" instead of extracting general knowledge from them. Indeed, although our datasets are large, they contain far fewer participants than variables measured for each participant, which reduces the ability to generalize. Finally, the sources of biological and technical variation, generally inconsistent from one dataset to another, must be considered to obtain reliable predictions in the real world. My research program offers concrete solutions to make these methodologies applicable to omics data, for specific biological problems. One of them aims to predict, from an individual's genetics, changes in the levels of RNA transcripts and metabolites. Another problem targets the prediction of the risk of developing a complex disease based on an individual's omics data. For these concrete applications, we will use omics data from several biobanks, including local (Montreal Heart Institute Biobank), national (CanPath cohort) and international (UK Biobank) cohorts. We will develop these methodologies while ensuring to obtain interpretable and plausible results, which generalize well, and are independent of the noise sources. Our program is interdisciplinary and offers a rich training opportunity for students. Our results will have the potential to improve the appropriate use of machine learning in molecular biology research, providing researchers and Canadian industry with robust tools for omics data analysis that can be interpreted by humans.