Cross-platform structural variant discovery with deep learning

结构变异(SV)是人类基因组中遗传多样性和疾病的主要驱动因素 发现对于精确医学的进步和我们对人类遗传学的理解是必不可少的。由于 全基因组测序技术的革命性突破，我们现在可以通过 前所未有的规模和分辨率。然而，尽管在SV调用方法方面付出了巨大的努力并取得了巨大的进步， 一般的SV发现仍然没有解决。现有技术使用手工设计的特征和启发式方法 模型SV类，严重依赖开发人员的专业知识，无法扩展到SV类型的巨大多样性和 测序平台也不能充分利用原始测序数据中的所有信息。因此，这些 方法通常与特定测序技术的属性紧密结合，并以最佳方式运行 仅限于某些类型和大小的SV，使我们对许多其他类型的SV及其在疾病中的作用视而不见。深沉 神经网络具有从数据自动学习复杂抽象的能力，因此提供了一种 为普通SV发现开辟了一条很有前途的途径。深度学习最近改变了机器学习领域 并导致了科学和医学的显著进步。在这项提议中，我们的目标是利用深度 关于SV检测问题的学习。我们展示了如何有效地将SV检测作为一种深度学习 任务，并建议开发一个全面的框架，以调用和分型不同大小和 类型，包括复杂和亚克隆的SVS，给出了来自一系列测序平台的数据。特别是，我们 演示使用我们的短、链接和长阅读方法可以获得最先进的结果 数据集。为了确保我们的模型适用于不同的数据集，我们提案的一个重要目标 也是收集各种具有代表性的培训数据，并使用公开的- 可用的多平台数据集，可准确评估模型性能。我们的软件将由 考虑到可扩展性和可伸缩性，并将与预先训练的模型和调用集一起免费发布给 社区。

Structural variants (SV) are a major driver of the genetic diversity and disease in the human genome and their discovery is imperative to advances in precision medicine and our understanding of human genetics. Due to revolutionary breakthroughs in whole-genome sequencing technologies, we now have access to genomic data at an unprecedented scale and resolution. However, despite tremendous effort and progress in SV calling methodology, general SV discovery still remains unsolved. Existing techniques use hand-engineered features and heuristics to model SV classes, relying heavily on developer expertise, which cannot scale to the vast diversity of SV types and sequencing platforms nor fully harness all the information available in raw sequencing data. As a result, these methods are usually tightly coupled to the properties of a particular sequencing technology and operate optimally only on certain SV types and sizes, rendering us blind to many other classes of SVs and their role in disease. Deep neural networks have the ability to learn complex abstractions automatically from the data and hence offer a promising avenue for general SV discovery. Deep learning has recently transformed the field of machine learning and led to remarkable advances in science and medicine. In this proposal we aim to leverage the potential of deep learning for the problem of SV detection. We lay out how to efficiently formulate SV detection as a deep learning task, and propose the development of a comprehensive framework to call and genotype SVs of different size and type, including complex and subclonal SVs, given data from a range of sequencing platforms. In particular, we demonstrate that state-of-the-art results can be obtained using our approach for short, linked, and long read datasets. In order to ensure that our models generalize across different datasets, an important goal of our proposal is also to assemble diverse and representative training data and perform extensive evaluation using publicly- available multi-platform datasets to accurately assess model performance. Our software will be built with extensibility and scalability in mind, and will be released, along with pretrained models and callsets, freely to the community.