Denoising the Denoisers: An independent evaluation of microbiome sequence error-correction methods
- Published
- Accepted
- Subject Areas
- Genomics, Microbiology, Data Science
- Keywords
- Microbiome, Denoising Tools, Comparison, DADA2, Deblur, UNOISE3, Mock community
- Copyright
- © 2018 Nearing et al.
- Licence
- This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, reproduction and adaptation in any medium and for any purpose provided that it is properly attributed. For attribution, the original author(s), title, publication source (PeerJ Preprints) and either DOI or URL of the article must be cited.
- Cite this article
- 2018. Denoising the Denoisers: An independent evaluation of microbiome sequence error-correction methods. PeerJ Preprints 6:e26566v1 https://doi.org/10.7287/peerj.preprints.26566v1
Abstract
High-depth sequencing of universal marker genes such as the 16S rRNA gene are a common strategy to profile microbial communities. Traditionally, sequence reads are clustered into operational taxonomic units (OTUs) at a defined identity threshold to avoid sequencing errors generating spurious taxonomic units. However, there have been numerous bioinformatic methods recently released that attempt to correct sequencing errors to determine real biological sequences at single nucleotide resolution by generating amplicon sequence variants (ASVs). As the microbiome field moves from OTUs to higher resolution ASVs, there is a need for an in-depth and unbiased comparison of these novel “denoising” methods. In this study, we conduct a thorough comparison of three of the most widely-used denoising methods on mock, soil, and host-associated communities. We tested three different methods - DADA2, UNOISE3, and Deblur - on four mock communities and found that, although they produced similar microbial compositions based on relative abundance, the methods identified vastly different numbers of ASVs. Our analysis of a soil dataset also showed that the three methods were consistent in their per-sample compositions, resulting in only minor differences based on weighted UniFrac distances. However, DADA2 tended to find more ASVs than the other two methods when analyzing both the real soil data and two other host-associated datasets, suggesting that it could be better at finding rare organisms. The three tested methods were significantly different in their run times, with UNOISE3 running greater than 1200 and 15 times faster than DADA2 and Deblur, respectively. Our results indicate that the choice of denoising method will depend on a researcher’s individual importance for identifying rare ASVs, the availability of computational resources, and their willingness to support open-source or closed-source software.
Author Comment
This is a preprint submission to PeerJ Preprints.
Supplemental Information
Figure 1: Total number of ASVs identified by each denoising method for four different mock communities
Amplicon sequence variants (ASVs) were compared to a database of full-length amplicon sequences for jSMB5ust the microbes supposedly in the community (“Expected”) and against the full SILVA or ITS databases (“Database”) using BLASTN at 97% and 100% identity cutoffs. “Unmatched” sequences did not match an expected sequence or the SILVA/ITS databases at 97% identity or greater. Dotted lines indicate the total number of ASVs expected, accounting for 16S copy variation within genomes. A) Human Microbiome Project mock community; B) Extreme dataset; C) Fungal ITS1 mock community; D) Zymomock community.
Figure 2: Relative abundances of taxa generated by each denoising method for four different mock communities
All ASVs that matched with expected sequences at 97% or greater identity were assigned taxonomy using a BLASTN search against the expected sequences provided for each the Extreme, Human Microbiome Project, and Zymomock mock communities. All ASVs that matched an expected species with 97% or greater identity to the UNITE database were classified as expected sequences. Non_reference refers to the abundance of ASVs that did not match expected sequences with 97% or greater identity. A) Human Microbiome Project mock community; B) Extreme dataset - it is important to note that due to the low abundance of some organisms in the Extreme dataset they were not displayed in this figure; C) fungal ITS1 mock community; D) Zymomock community.
Figure 3: Intra-sample distances between denoising methods based on a real soil community
A) The weighted UniFrac distances between the same biological samples based on ASVs outputted by each of the different methods. B) The Bray-Curtis dissimilarity distances between the same biological samples based on genera outputted by the three methods after being classified with the RDP classifier. Deblur tends to be slightly more dissimilar when compared to the other two methods. C) Principal coordinates analysis of the weighted UniFrac distances of all the samples in the real soil dataset generated by each method. The three different profiles generated for each biological sample are colour-coded and are joined by an interconnecting line. D) Non-metric multidimensional scaling plot that displays the Bray-Curtis dissimilarity profiles of all the samples in the real soil dataset generated by each method. The three different profiles generated for each biological sample are colour-coded and are joined by an interconnecting line.
Figure 4: Run time and memory usage of each denoising method on a dataset of varying size
The time in seconds A) and memory in megabytes B) to run varying amounts of reads through the three different methods. Note time is on a log10 scale.
Supplemental Figure 1: Removal of low abundance ASVs removes many unmatched sequences from Deblur- and DADA2-generated ASVs
Amplicon sequence variants (ASVs) were run through an abundance filtering at 0.1% and then were compared to a database of full-length amplicon sequences for just the microbes supposedly in the community (“Expected”) and against the full SILVA or ITS databases (“Database”) using BLASTN at 97% and 100% identity cutoffs. “Unmatched” sequences did not match an expected sequence or the SILVA 16S rRNA gene database at 97% identity or greater. Dotted lines indicate the total number of ASVs expected, accounting for 16S gene-copy variations within genomes. A) Human Microbiome Project mock community; B) Extreme dataset; C) Fungal ITS1 mock community; D) Zymo mock community.
Supplemental Figure 2: DADA2 finds more rare organisms than Deblur or UNOISE3
Rank-abundance curves for ASVs (A) and classified species (B) generated from the soil dataset using the DADA2, Deblur and UNOISE3 methods. ASVs were classified using the RDP classifier against the Greengenes (13_8) database.
Supplemental Figure 3: Filter stringency does not affect relative abundance data drastically
The Human Microbiome Project mock community was run using DADA2, UNOISE3, and Deblur at varying stringency filters (low, medium and high). Resulting relative abundance profiles are shown for A) DADA2, B) Deblur and C) UNOISE3.
Supplemental Figure 4: Intra-sample distances between methods based on intestinal biopsy samples from pediatric Crohn’s disease patients and controls
A) The weighted UniFrac distances between the same biological samples based on ASVs outputted by each of the different methods. B) The Bray-Curtis dissimilarity distance between the same biological samples based on genera outputted by the three methods after being classified with the RDP classifier. C) Principal coordinates analysis of the weighted UniFrac distances of all the samples in the real soil dataset generated by each method. The three different profiles generated for each biological sample are colour-coded and are joined by an interconnecting line. D) Non-metric multidimensional scaling plot that displays the Bray-Curtis dissimilarity profiles of all the samples in the real soil dataset generated by each method. The three different profiles generated for each biological sample are colour-coded and are joined by an interconnecting line.
Supplemental Figure 5: Intra-sample distances between methods based on mouse exercise associated fecal samples
A) The weighted UniFrac distances between the same biological sample based on ASVs outputted by each of the different methods. B) The Bray-Curtis dissimilarity distance between the same biological samples based on genera outputted by the three methods after being classified with the RDP classifier. C) Principal coordinates analysis of the weighted UniFrac distances of all the samples in the real soil dataset generated by each method. The three different profiles generated for each biological sample are colour-coded and are joined by an interconnecting line. D) Non-metric multidimensional scaling plot that displays the Bray-Curtis dissimilarity profiles of all the samples in the real soil dataset generated by each method. The three different profiles generated for each biological sample are colour-coded and are joined by an interconnecting line.
Supplemental Figure 6: There are outlier genera that drastically differ in relative abundance between Deblur and the other denoising methods
ASVs were classified using the RDP classifier against the Greengenes (13_8) database. Relative abundances of each genus were than compared between methods and differences were plotted in a histogram. A) Relative abundance differences by genus between DADA2 and Deblur. B) Relative abundance differences by genus between DADA2 and UNOISE3. C) Relative abundance differences by genus between Deblur and UNOISE3.
Supplemental Figure 7: Top 5 genera driving differences between Deblur and the other two denoising tools in the soil dataset
Boxplots of the relative abundances per sample of five of the classified genera that had relative abundance differences greater than 1% between Deblur and both DADA2 and UNOISE3. Deblur calls more reads that were unclassified at the kingdom and class levels than DADA2 or UNOISE3. A) ASVs only classified at the Bacteria kingdom level. Deblur tends to find higher abundances of these ASVs. B) ASVs only classified at the Verrucomicrobia phylum level. Deblur finds higher abundances of these ASVs. C) ASVs only classified at the Spartobacteria class level. DADA2 and UNOISE3 find more of these ASVs than Deblur. D) ASVs classified at the Gp1 order level of the Acidobacteria_Gp1 class. E) ASVs classified at the Granulicella order level of the Acidobacterta_Gp1 class. Strikingly these two classifications share opposite relationships where Deblur finds more ASVs in the Gp1 order and DADA2 and UNOISE3 find more ASVs in the Granulicella order.