之前我写过一篇利用3D-DNA流程基于Hi-C提升基因组组装，那个时候我做的项目并不多，也没有遇到坑，直到最近我用3D-DNA流程进行基因组组装，发现结果出乎意料

Fig1

在上图中，蓝色框表示组装出来的scaffold，其余则是3D-DNA认为的debris，也就是边角料。谁能想到一个基因组有那么多的边角料呢？

经 @上海欧易生物-鲍志贵的提醒，我检查了genome.0.hic, 也就是初步组装的结果，发现结果好的惊人。

Fig2

谁能想到，越纠错结果越错呢？现在的问题是，既然错误已经存在了，那么应该如何处理呢？

最容易的解决方法就是设置 -r 0, 也就是不纠错，直接用genome.0.hic的结果开始拆分染色体，后续通juicebox(https://github.com/aidenlab/Juicebox/wiki)手动处理潜在组装错误。这其实并没有解决问题，而是躲避了问题。于是我去阅读了3D-DNA 发表在science上文章的附录，和run-asm-pipeline.sh的源代码，功夫不负有心人，终于被我定位到了问题的关键所在。

run-asm-pipeline.sh的651-750是3D-DNA流程的核心算法: misjoin-correction. 其中下面的代码是用来找到潜在的mis-assembly, 用于后续纠错

bash ${pipeline}/edit/run-mismatch-detector.sh -p ${parallel} -c ${editor_saturation_centile} -w ${editor_coarse_resolution} -d ${editor_coarse_region} -k ${editor_coarse_stringency} -n ${editor_fine_resolution} ${genomeid}.${ROUND}.hic
bash ${pipeline}/edit/run-coverage-analyzer.sh -w ${editor_coarse_resolution} -t ${editor_repeat_coverage} ${genomeid}.${ROUND}.hic

输出的 mismatch_narrow.at.step.${ROUND}.bed和 repeats_wide.at.step.${ROUND}.bed 后续会被3D-DNA用于编辑原始的contig/scaffold。我分别统计了这两者的数量和长度， 其中mismatch长度为1，也就是没有发现组装错误，而重复区域则非常可怕

awk '{print $3-$2}' repeats_wide.at.step.0.bed | sort -k1,1nr | head
13775000
11450000
9200000
7075000
5325000
4350000
4225000
4225000
4175000
4100000

最长的重复区域居然长达13 Mb, 想想都不太可能啊（我统计了另一组数据，长度最长为50K）。显然这就是问题的所在，接下来就该解决问题了。

既然是重复区域出了问题，我翻阅了参数列表，找到了一个与其相关的参数，即--editor-repeat-coverage, 默认是2，我们可以运行run-coverage-analyzer.sh, 测试不同参数下(-t)下的重复序列长度情况，从而确定一个比较合适的大小。

bash /opt/biosoft/3d-dna/edit/run-coverage-analyzer.sh -w 25000 -t 3 genome.0.hic

此外，我还查阅了在Google Group上相关的讨论，发现这个问题是由于文库导致（Most likely it is due to coverage biases in your experiment. Load the tracks associated with repeats to help highlight. You can increase --editor-repeat-coverage to help mitigate.）

Highly likely repeats is main problem of this assembly. Significant part of Hi-C reads are filtered as below MAPQ but checking reads shows non-unique mapping at least one side of such reads, and it can be result of presence large fraction of repeats. As well as remapping reads to selected after JBAT chromosome-size scaffolds shows coverage anomalies looks like crosses with much more higher contact frequency than other genome, probably corresponding to condensed repeats. And can i ask your advise how should we deal with such repeats mark with repeatFinder and hardmask before reads mapping or something else? And we try to increase editor-repeat-coverage parameter.

参考资料:

• https://groups.google.com/g/3d-genomics/c/f6P_gJC-jMo/m/ADY8GBH2AAAJ