Leveraging functional profiling datasets with machine learning to uncover proteins and cellular processes important for ageing

在发达国家，老龄化是大多数人类疾病的最大风险因素，包括阿尔茨海默氏症和帕金森氏症等进展性疾病，癌症等发病率不同的疾病，以及心脏病发作和中风等灾难性系统故障。虽然对特定疾病过程的研究长期以来一直是研究的主要重点，但人们越来越认识到，将正常衰老过程本身作为问题的重要组成部分进行研究，并探索减缓或扭转其影响的方法的重要性。老龄化是一个多因素的过程，可以被视为时间蹂躏的必然特征。然而，最近的发现表明，通过简单的干预可以戏剧性地改变衰老。例如，单基因敲除可以延缓实验动物晚年的衰老和改善健康。不同生物体的衰老过程相似，影响这些过程的基因突变与人类寿命有关。然而，衰老研究的一个核心挑战仍然是梳理出决定寿命的生物因素和过程的完整和统一的图景。衰老是高度复杂的，受到不同蛋白质和过程的影响。现代生物分析可以同时测量数以千计的蛋白质或基因的性质和相互作用，但要理解如此庞大的数据集是具有挑战性的。被称为“机器学习”的计算数据分析方法的进步，为最大限度地利用大型生物数据集提供了令人兴奋的机会，从而增加了我们对衰老等复杂过程的理解。机器学习可以在人类无法处理的过于复杂的数据中发现隐藏的模式。计算机能力、算法和数据大小的进步使最近的机器学习体系结构(称为深度学习)能够准确地在不同类型的组合数据集中找到复杂的模式并进行分类。我们计划使用分裂酵母作为模式生物，结合多步机器学习，全面识别对衰老具有根本重要性的生物过程。值得注意的是，从酵母到人类，这些过程中的许多都是相似的，但在简单的酵母中更容易研究。酵母细胞在有限的营养条件下进入休眠、不分裂状态。这种休眠细胞提供了一个有用的系统来分析影响这种状态下寿命的蛋白质和过程。在之前的研究中，我们已经确定了116种蛋白质，当没有这些蛋白质时，酵母可以活得更长(长寿的基因敲除突变体)。因此，这些蛋白质与衰老有关，可以用来训练机器学习程序，以预测新的衰老蛋白质，这种方法被称为“关联内疚”。我们将把关于突变特征(表型)的大型系统数据与不同的现有数据结合起来，以增强机器学习预测器的能力。我们将在实验室中测试预测的老化蛋白质对酵母寿命的影响，并将这些信息反馈给计算机，以便它了解更多关于老化蛋白质的信息。然后，我们将使用计算机识别并在酵母中确认的新老化蛋白的突变来测量与所有其他突变的联系。这种“基因相互作用”数据提供了丰富的功能关系信息，这些信息将被用来探索其他可能更强大的深度学习方法，以预测与衰老有关的生物过程。然后，我们将用实验室实验来检验最具吸引力的预测。此外，我们将向感兴趣的科学家提供所有新的数据、方法和预测，以帮助他们的研究。我们预计，该项目将利用实验和机器学习的亲密循环，提供一个宝贵的平台，以更好地了解与衰老有关的所有生物因素，最终开发出延长人类健康寿命的干预措施。

Ageing is the largest risk factor for most human diseases in developed countries, including progressive diseases such as Alzheimer's and Parkinson's, diseases like cancer that show variable rates of onset, and catastrophic system failures such as heart-attack and stroke. While the study of specific disease processes has long been a major focus of research, there is a growing realization of the importance of studying the normal ageing process itself as an essential part of the problem, and of exploring ways to slow or reverse its effects. Ageing is a multi-factorial process that can be seen as an inevitable feature of the ravages of time. Recent discoveries, however, demonstrate that ageing can be modified in dramatic ways by simple interventions. For example, single gene knockouts can delay ageing and improve health late in the life of laboratory animals. The processes involved in ageing are similar in different organisms, and genetic mutations affecting these processes are associated with longevity in humans. A central challenge of ageing research, however, remains to tease out a complete and unified picture of the biological factors and processes determining lifespan. Ageing is highly complex and affected by diverse proteins and processes. Modern biological assays can simultaneously measure properties and interactions of thousands of proteins or genes, but it is challenging to make sense of such large datasets. Advances in computational data-analysis methods, called 'machine learning', provide exciting opportunities to get the most from large biological datasets and thus increase our understanding of complex processes like ageing. Machine Learning can find hidden patterns in data that is too complex for humans to process. Advances in computer power, algorithms and data sizes allow recent machine-learning architectures (known as 'deep learning') to accurately find and classify intricate patterns in combined datasets of different types. We plan to use fission yeast as a model organism, together with multi-step machine learning, to comprehensively identify biological processes with fundamental importance for ageing. Remarkably, many of these processes are similar from yeast to human, but are much easier to study in the simple yeast. Yeast cells enter a dormant, non-dividing state under limiting nutrients. Such dormant cells provide a useful system to analyse proteins and processes affecting the lifespan in this state. In previous studies, we have identified 116 proteins that, when absent, allow the yeast to live longer (long-lived knockout mutants). So these proteins are involved in ageing, and can be used to train machine-learning programs to predict new ageing proteins by a method known as 'guilt by association'. We will combine large systematic data on mutant features (phenotypes) with diverse existing data to empower the machine-learning predictor. We will test the predicted ageing proteins in the laboratory for lifespan effects in yeast, and feed this information back to the computer for it to learn more about ageing proteins. We will then use mutants of the new ageing proteins identified by the computer and confirmed in yeast to measure links with all other mutants. Such 'genetic-interaction' data provide rich information on functional relationships, which will be used to explore other, potentially more powerful deep-learning methods to predict the biological processes that are involved in ageing. We will then test the most attractive predictions with laboratory experiments. Moreover, we will make all the new data, methods and predictions available to interested scientists to help with their research. We anticipate that this project, using intimate cycles of experiments and machine-learning, will provide a valuable platform to better understand all the biological factors involved in ageing, to eventually develop interventions that extend healthy lifespan in humans.