Overview of NGS Errors Working Group

Size: px

Start display at page:

Download "Overview of NGS Errors Working Group"

Jane Dean
5 years ago
Views:

1 Overview of s Working Group Or...My Declaration of War on the Bioinformatics Pipeline K. S. Dorman Department of Statistics and Genetics, Development & Cell Biology SAMSI - Beyond Bioinformatics May

2 Initial Proposed WG Goals Overarching Goal. Increase reproducibility by reducing noise from the pipeline. To learn about common pipeline components. To identify pipeline components most likely to affect downstream analyses. To write a review on statistical concerns in common pipeline components: What you need to know. Develop research projects to integrate more mathematical/statistical thinking in pipeline components.

3 WG Meetings - Pipeline Components Covered. Sequencing platforms. - Maria C. Rivera, Virginia Commonwealth U.. Base calling. - Xinping Cui, U. California, Riverside. Read quality filtering/trimming. - Gabriel Murillo, U. California, Riverside. Error correction. - Karin Dorman, Vahid Noroozi, Xin Yin, Iowa State U.. Alignment. - Adam B. Olshen, U. California, San Francisco Assembly - A. Severin, Iowa State U.. Bud Mishra, Courant Institute (TotalRe- Caller, Base-Calling, Alignment, Assembly, GWAS)

Sequencing Platforms Main platforms from: Illumina, Life Technologies, Pacific Biosciences, Oxford nanopore Isn t it dangerous to invest so much energy in a technology that will undoubtedly be soon

4 Sequencing Platforms Main platforms from: Illumina, Life Technologies, Pacific Biosciences, Oxford nanopore Isn t it dangerous to invest so much energy in a technology that will undoubtedly be soon replaced? Errors. Error rates remain high, highest for the newest technology. Trade-off. High throughput and errors or low throughput and precision. The key replication in these experiments is the reading and re-reading of short fragments.

5 Base Calling (Illumina) The overwhelming majority of calls are being made by Bustard, the unpublished Illumina base caller. Bustard is not the best method.

6 Error Removal Base calling. Really good base calling can prevent errors. The consequences of residual errors. De novo assembly. Errors lead to false kmers and increased complexity of and errors in resulting assembly. Alignment. Errors lead to misaligned or unaligned sequences. Read trimming. Identify low quality bases and drop them from the end(s) of reads. Error correction. Identify and correct low abundance kmers, but cannot work when coverage is not uniform (e.g. RNA-seq, metagenomics, single-cell, heterogeneous populations, like cancer). Alignment or Assembly. Smart aligners can tolerate some residual errors, and assemblers often do error correction.

7 Read Trimming Del Fabbro, C.; Scalabrin, S.; Morgante, M. & Giorgi, F. M. An extensive evaluation of read trimming effects on Illumina NGS data analysis. PLoS One, 2013, 8, e85024 What is the best trimming algorithm? [N]o generic answer can be given.... [It depends] on the dataset, downstream analysis, and user-decided parameter-dependent tradeoffs.... [but for] processing DNA-Seq data... trimming should be applied every time in order to improve quality and performance.

8 The Science of Pipelines - Concerns Few studies. There are few comprehensive comparisons of base-calling, trimming, error correction, or alignment pipeline components. The combinatorics is forbidding. Wrong outcome. In comparison studies, downstream performance may be implied via: Alignability, or correct genome localization if simulation data. Error rate conditional on alignment, or true error rate if simulation data. Complicated findings. Almost anything can affect downstream performance. The dataset, particularly its quality. The particular pipeline component used. The parameter settings of the pipeline components. The downstream objective itself.

9 The Usage of Pipelines - Concerns Hard to set up. If the manuals are good, setting parameters still takes extensive experience and empirical experimentation. There are tens of versions of each component, and generally not thoroughly compared.. Not generalizable. In desperation, we learn from practiced experts, but not all their experience extends to new situations, our situation. Poorly understood. Not only is our knowledge limited, it is largely empirical; there is little insight as to why one pipeline works better than another. Herd mentality. Since no one really understands them, pipelines are popularized as temporary fads, with little science driving their adoption or death.

10 The solution? Pipelines do not seem to be leaking completely at random, and Additional components have work to do only because signal is lost, unused, and ignored, especially at pipeline junctions.

11 The solution? Pipelines do not seem to be leaking completely at random, and Additional components have work to do only because signal is lost, unused, and ignored, especially at pipeline junctions. Death to the bioinformatics pipeline!

12 The solution? Pipelines do not seem to be leaking completely at random, and Additional components have work to do only because signal is lost, unused, and ignored, especially at pipeline junctions. Death to the bioinformatics pipeline! (At least everything between base calling and alignment/assembly.)

13 The solution? Pipelines do not seem to be leaking completely at random, and Additional components have work to do only because signal is lost, unused, and ignored, especially at pipeline junctions. Death to the bioinformatics pipeline! (At least everything between base calling and alignment/assembly.) Ashley Cacho, U. California Riverside: A comprehensive comparison of base calling methods. Xin Yin, : Combined base calling and error correction.

Hybrid and non hybrid error correction for long reads: LoRDEC and LoRMA

Hybrid and non hybrid error correction for long reads: LoRDEC and LoRMA Eric Rivals Computer Science Lab & Institute Computational Biology, CNRS & Univ. Montpellier November 7, 2016 Rivals (CNRS Univ.