Machine learning methods for genome reconstruction in metagenomics

宏基因组学已经成为微生物群落分析的标准方法。环境或临床样本的测序结果是大量的短序列片段，可以对物种组成和代谢潜力进行全面分析。测序技术的进步大大提高了测序深度，因此即使对于最复杂的群落，也可以组装长连续区域（contigs）的微生物基因组。最近的研究表明，通过将这些contigs分组到基因组特异性箱中，可以重建几乎完整的基因组。宏基因组分组是一个计算上具有挑战性的问题，尽管目前用于基因组恢复的分组工具都基于类似的聚类技术，但结果的质量取决于用户规格，并且在不同的社区和工具之间差异很大。为了充分利用宏基因组分类在基因组重建中的潜力，我们的目标是开发一个机器学习框架，该框架将能够自动优化结果的准确性和可重复性。为了实现这一目标，我们将把最先进的机器学习模型与计数数据的统计模型和基于仿真的整体模型质量控制结合起来。在与Walnut Creek的联合基因组研究所的合作中，我们将开发一种基于蛋白质结构域含量和与已知基因组比对来评估恢复基因组的完整性和污染的新方法。与将质量检查应用于最终箱的元基因组箱的常见工作流程相反，我们将把质量控制作为基于机器学习的箱优化过程的一个组成部分。在与Göttingen基因组学和应用微生物学部门的持续合作中，我们还将研究第三代测序技术的长度读数可以用于测量和提高分集精度。与森林植物学和树木生理学部门的Göttingen合作伙伴一起，我们的目标是将分类方法扩展到真菌群落，这将为生态学研究开辟新的可能性。

Metagenomics has become a standard approach for the analysis of microbial communities. The sequencing of environmental or clinical samples results in large collections of short sequence fragments that enable a comprehensive analysis of the species composition and the metabolic potential. The progress in sequencing technologies has given rise to a significantly increased sequencing depth so that it has become possible to assemble long contiguous regions (contigs) of microbial genomes even for the most complex communities. Recent studies have shown that it is possible to reconstruct nearly complete genomes by grouping these contigs into genome-specific bins. Metagenome binning is a computationally challenging problem, and although current binning tools for genome recovery are all based on similar clustering techniques, the quality of results depends on user specifications and largely varies across different communities and tools. To fully exploit the potential of metagenome binning for genome reconstruction, we aim to develop a machine learning framework that will enable an automatic optimization of the accuracy and the reproducibility of results. To achieve this goal we will integrate state-of-the-art machine learning models together with statistical models for count data and a simulation-based control of the overall model quality. In a collaboration with the Joint Genome Institute in Walnut Creek we will develop a new approach to assess the completeness and contamination of recovered genomes based on the protein domain content and on alignments with known genomes. In contrast to the common workflow of metagenome binning which applies quality checks just to the final bins we will include quality control as an integral part of the machine learning-based bin optimization process. In an ongoing collaboration with the Department of Genomic and Applied Microbiology in Göttingen we will also investigate how long reads from third generation sequencing technologies can be utilized to measure and improve binning accuracy. Together with our Göttingen collaboration partner from the Department of Forest Botany and Tree Physiology we aim to extend the binning approach to fungal communities which will open up new possibilities for research in ecology.