A predictive model of mRNA stability and translation for variant interpretation and mRNA therapeutics

MRNA的前导和尾随非翻译区(UTRs)，以及编码序列(CDS)， 通过调节翻译和信使核糖核酸的稳定性控制蛋白质的产生。然而，尽管我们已经确定 这些地区的大量监管特征，我们还远远无法预测，例如， 序列变异是否以及如何影响所制造的蛋白质水平。在这里，我们建议将 用机器学习高通量实验表征合成文库中的蛋白质表达 建立翻译和信使核糖核酸稳定性的预测模型，以满足迫切的需求。的最新进展 机器视觉、语音识别和其他计算机科学领域的驱动因素是 用来训练模型的海量数据集。机器学习方法也产生了显著的影响 在生物学中，但生物学数据集往往相对较小，限制了可以 学到了东西。例如，人类基因组中只有大约20,000个基因，这是一个限制性很小的集合 训练预测模型的例子，该模型捕捉到基因组“调控密码”的全部范围。在这 建议，我们的目标是通过训练蛋白质表达的预测模型来克服这种数据大小限制 来自数百万个合成结构的数据--一个数据集比 基因组中的基因。具体地说，我们将创建具有靶向变异的体外转录的mRNA文库 在UTRs和CDS中，并将通过执行高通量分析每个文库成员的蛋白质表达 多聚体图谱、核糖体图谱和信使核糖核酸稳定性分析。然后我们将使用神经网络 学习mRNA序列与蛋白质水平关系预测模型的方法 制作。我们将把我们的模型应用于三个具有实际意义的应用：第一，我们希望发现 新的生物学，例如识别调控序列元件以及它们之间的相互作用。第二, 我们将通过从头设计和序列的实验测试来验证我们的模型 比原始的数百万随机产生的成员中的任何一个产生更高的水平或蛋白质 文库或比目前生物技术中使用的内源非编码区序列更多的序列。这样的稳定和高度 翻译信使核糖核酸结构将对该领域或信使核糖核酸疗法具有特殊的价值。第三，我们将 预测UTRs遗传变异对蛋白质生产的功能影响，我们将验证 这些预测是实验性的。我们还远不能理解哪些基因变异会损害基因 以可能导致疾病的方式发挥调节功能，使这样一个全面和量化的 对变异的分析很有价值。

The leading and trailing untranslated regions (UTRs) of an mRNA, along with the coding sequence (CDS), control protein production by modulating translation and mRNA stability. However, although we have identified a vast number of regulatory features in these regions, we are still far from being able to predict, for example, whether and how a sequence variant affects the levels of protein being made. Here, we propose to combine high-throughput experimental characterization of protein expression in synthetic libraries with machine learning to create predictive models of translation and mRNA stability, addressing an urgent need. Recent progress in machine vision, voice recognition and other fields of computer science has been driven by the availability of enormous data sets on which to train models. Machine learning approaches have also had remarkable impact in biology, but biological data sets often are comparatively small, limiting the quality of models that can be learned. For example, there are only around 20,000 genes in the human genome, a restrictively small set of examples for training a predictive model that captures the full extent of the genome’s “regulatory code.” In this proposal, we aim to overcome this data size limitation by training predictive models of protein expression on data from millions of synthetic constructs -- a data set several orders of magnitude larger than the number of genes in the genome. Specifically, we will create libraries of in vitro transcribed mRNA with targeted variation in the UTRs and CDS and will assay protein expression of each library member by performing high-throughput polysome profiling, ribosome profiling, and mRNA stability assays. We will then use neural network approaches to learn predictive models of the relationship between mRNA sequence and levels of protein production. We will apply our models to three applications of practical importance: first, we expect to uncover novel biology, for example identifying regulatory sequence elements and interactions between them. Second, we will validate our models through the de novo design and experimental testing of sequences that result in higher levels or protein production than any of the millions of randomly generated members of the original library or than the endogenous UTR sequences currently used in biotechnology. Such stable and highly translating mRNA constructs would be of particular value for the field or mRNA therapeutics. Third, we will predict the functional consequences of genetic variation in UTRs on protein production and we will validate these predictions experimentally. We are far from understanding which genetic variants compromise gene regulatory function in ways that may contribute to disease, making such a comprehensive and quantitative analysis of variants valuable.