Predictive Modeling of Alternative Splicing and Polyadenylation from Millions of Random Sequences

构成基因调控基础的人类基因组的比例使编码基因的比例相形见绌 蛋白质。然而，我们仍然不具备识别哪些基因变异危害基因的能力。 监管职能可能会增加罕见和常见人类疾病的风险。 了解非编码序列如何调节基因表达，以及能够预测 基因变异对基因调控的功能后果是该领域面临的最大挑战。这里, 我们建议将合成生物学、大规模并行功能分析和机器学习结合起来 这将极大地促进我们对人类基因组“调控密码”的理解。在具有挑战性的同时， 从大量经验数据中解开复杂代码的任务并非没有先例。例如, 在过去的十年里，致力于自然语言处理的计算机科学家取得了巨大的成就 进步，在很大程度上是由算法和计算改进以及 比从事这项工作的前几代科学家可用的训练数据集要大得多 区域。受大数据对机器学习中传统问题的革命性影响的启发，我们 建议使用多几个数量级的训练数据集来模拟基因调控现象 这些例子比人类基因组中自然存在的要多。我们预测，从海量数据中学习的模型 合成示例的数量将大大超过从少量自然样本学习的模型 举个例子。我们将通过开发全面、定量和可预测的方法来展示我们的方法 选择性剪接和选择性聚腺苷酸化的模型，这是两种广泛存在的调控机制 单个基因可以编码多个转录本和蛋白质。然而，我们预计这一基本的 范式-具体地说，是对极大的 在序列-功能关系的定量建模之后的合成序列的数量-可以是 以促进我们对不同形式的基因调控的理解。

The proportion of the human genome that underlies gene regulation dwarfs the proportion that encodes proteins. However, we remain poorly equipped for identifying which genetic variants compromise gene regulatory function in ways that may contribute to risk for both rare and common human diseases. Understanding how non-coding sequences regulate gene expression, as well as being able to predict the functional consequences of genetic variation for gene regulation, are paramount challenges for the field. Here, we propose to combine synthetic biology, massively parallel functional assays, and machine learning to profoundly advance our understanding of the `regulatory code' of the human genome. While challenging, the task of unravelling complex codes from large amounts of empirical data is not without precedent. For example, over the past decade, computer scientists working in natural language processing have made immense progress, driven in large part by a combination of algorithmic and computational improvements and enormously larger training datasets than were available to the previous generations of scientists working in this area. Inspired by the revolutionizing impact of “big data” for traditional problems in machine learning, we propose to model gene regulatory phenomena using training datasets with several orders of magnitude more examples than naturally exist in the human genome. We predict that the models learned from massive numbers of synthetic examples will strongly outperform models learned from the small number of natural examples. We will demonstrate our approach by developing comprehensive, quantitative, and predictive models for alternative splicing and alternative polyadenylation, two widespread regulatory mechanisms by which a single gene can code for multiple transcripts and proteins. However, we anticipate that this basic paradigm – specifically, the massively parallel measurement of the functional behavior of extremely large numbers of synthetic sequences followed by quantitative modeling of sequence-function relationships – can be generalized to advance our understanding of diverse forms of gene regulation.