Department of Microbiology & Immunology, University of Michigan, Ann Arbor, Michigan, United States of America.
PLoS One. 2011;6(12):e27310. doi: 10.1371/journal.pone.0027310. Epub 2011 Dec 14.
The advent of next generation sequencing has coincided with a growth in interest in using these approaches to better understand the role of the structure and function of the microbial communities in human, animal, and environmental health. Yet, use of next generation sequencing to perform 16S rRNA gene sequence surveys has resulted in considerable controversy surrounding the effects of sequencing errors on downstream analyses. We analyzed 2.7×10(6) reads distributed among 90 identical mock community samples, which were collections of genomic DNA from 21 different species with known 16S rRNA gene sequences; we observed an average error rate of 0.0060. To improve this error rate, we evaluated numerous methods of identifying bad sequence reads, identifying regions within reads of poor quality, and correcting base calls and were able to reduce the overall error rate to 0.0002. Implementation of the PyroNoise algorithm provided the best combination of error rate, sequence length, and number of sequences. Perhaps more problematic than sequencing errors was the presence of chimeras generated during PCR. Because we knew the true sequences within the mock community and the chimeras they could form, we identified 8% of the raw sequence reads as chimeric. After quality filtering the raw sequences and using the Uchime chimera detection program, the overall chimera rate decreased to 1%. The chimeras that could not be detected were largely responsible for the identification of spurious operational taxonomic units (OTUs) and genus-level phylotypes. The number of spurious OTUs and phylotypes increased with sequencing effort indicating that comparison of communities should be made using an equal number of sequences. Finally, we applied our improved quality-filtering pipeline to several benchmarking studies and observed that even with our stringent data curation pipeline, biases in the data generation pipeline and batch effects were observed that could potentially confound the interpretation of microbial community data.
下一代测序的出现恰逢人们对利用这些方法来更好地了解微生物群落的结构和功能在人类、动物和环境健康中的作用产生了浓厚的兴趣。然而，使用下一代测序来进行 16S rRNA 基因序列调查导致了围绕测序错误对下游分析影响的相当大的争议。我们分析了分布在 90 个相同模拟群落样本中的 2.7×10(6)个读取，这些样本是来自 21 个具有已知 16S rRNA 基因序列的物种的基因组 DNA 的集合；我们观察到平均错误率为 0.0060。为了提高这个错误率，我们评估了许多识别不良序列读取、识别读取中质量较差区域以及纠正碱基调用的方法，并且能够将总错误率降低到 0.0002。实施 PyroNoise 算法提供了错误率、序列长度和序列数量的最佳组合。比测序错误更成问题的是在 PCR 过程中产生的嵌合体。因为我们知道模拟群落中的真实序列和它们可以形成的嵌合体，所以我们将 8%的原始序列读取识别为嵌合体。在对原始序列进行质量过滤并使用 Uchime 嵌合体检测程序后，总体嵌合体率下降到 1%。无法检测到的嵌合体主要负责识别虚假的操作分类单元 (OTU) 和属水平的分类群。随着测序工作量的增加，虚假 OTU 和分类群的数量增加，表明应该使用相同数量的序列来比较群落。最后，我们将改进的质量过滤管道应用于几个基准研究，并观察到即使使用我们严格的数据管理管道，也会观察到数据生成管道和批次效应中的偏差，这些偏差可能会混淆对微生物群落数据的解释。
