Long Reads Enable Accurate Estimates of Complexity of Metagenomes

Abstract

Although reduced microbiome diversity has been linked to various diseases, estimating the diversity of bacterial communities (the number and the total length of distinct genomes within a metagenome) remains an open problem in microbial ecology. We describe the first analysis of microbial diversity using long reads without any assumption on the frequencies of genomes within a metagenome (parametric methods) and without requiring a large database that covers the total diversity (non-parametric methods). The long read technologies provide new insights into the diversity of metagenomes by interrogating rare species that remained below the radar of previous approaches based on short reads. We present a novel approach for estimating the diversity of metagenomes based on joint analysis of short and long reads and benchmark it on various datasets. We estimate that genomes comprising a human gut metagenome have total length varying from 1.3 to 3.5 billion nucleotides, with genomes responsible for $50\%$ of total abundance having total length varying from only 40 to 60 million nucleotides. In contrast, genomes comprising an aquifer sediment metagenome have more than two-orders of magnitude larger total length (${\approx }840$ billion nucleotides).

Appendix

TruSPAdes Assemblies of MOCK, GUT, and SEDI Datasets. The TruSeq SLR technology generates accurate and long virtual reads derived from pools of short reads [27, 32, 52]. It is based on fragmenting genomic DNA into large segments (${\approx }10$ kb long) and forming random pools of the resulting segments (each pool contains ${\approx }300$ segments). Next, these fragments are amplified, sheared, and marked with a barcode that is unique to the pool. Afterwards, they are sequenced using the standard Illumina short reads technology. All short reads originating from the same barcode are assembled together resulting in a set of long contigs (this step is called the SLR barcode assembly). Ideally, the result of such sequencing effort for a single barcode is the collection of approximately 300 fragments (each fragment is ${\approx }10$ kb long) from a genome forming 300 long virtual reads. SLRs have low mismatch rate (about $0.1\%$), extremely low indel rate, and few misassemblies [3].

Table 2 presents results of barcode assembly of MOCK, GUT and SEDI datasets with truSPAdes.

Table 2. Results of truSPAdes assemblies of MOCK, GUT and SEDI datasets. Long SLRs are defined as SLRs longer than 6 kb.

Analyzing the CAMI and CROHN Datasets. In addition to datasets described in the main text, we also analyzed a larger synthetic dataset and four human microbiome datasets from a patient suffering from the Crohn’s disease.

The CAMI dataset is a simulated dataset generated by the “Critical Assessment of Metagenome Interpretation” (CAMI) initiative aimed at evaluating various approaches to analyzing metagenomes (http://www.cami-challenge.org/). We used a CAMI dataset simulated from 225 genomes and containing 150 million 100bp paired-end reads with mean insert size of 180bp (the errors in simulated reads are modelled after Illumina HiSeq). We simulated 50 thousand SLRs in the same way as for the SYNTH dataset. The total length of the reference genomes for this dataset is ${\approx }820$ Mb and its de Bruijn complexity is ${\approx }770$ Mb. Figure 4 shows that our estimator works well for the CAMI dataset.

The CROHN datasets are four human gut microbiome datasets from a patient with Crohn’s disease. These datasets (CROHN1, CROHN2, CROHN3, CROHN4) represent a metagenomics time series collected at 12-28-2011, 04-29-2013, 11-16-2014 and 06-29-2015, respectively. Each of these datasets includes one Illumina paired-end library and one SLR library. Number of short reads in these datasets ranges from 150 to 230 millions with mean insert size ${\approx }400$ bp for all datasets. The number of SLRs ranges from 17 to 50 thousand. Assembly efforts for these datasets resulted in contigs of length 242, 172, 225 and 275 Mb for CROHN1, CROHN2, CROHN3, and CROHN4 datasets respectively.

We estimated metagenome capacity for CROHN1, CROHN2, CROHN3, and CROHN4 datasets as 3.5, 2.0, 2.4, and 3.2 Gb, respectively. Values of M50 were estimated as 41, 61, 25, and 45 Mb, respectively, while values of M90 were estimated as 230, 490, 240, 250 Mb respectively. These estimates reveal large variations in metagenome capacity during the course of disease that go well beyond what can be estimated using short read assemblies. Correlation between metagenome capacity and antibiotics treatments for this metagenomics time series will be discussed elsewhere.

Fig. 4.

Estimated frequency histograms and abundance plots for CAMI (left) and , , , datasets (right). The distribution of heights (frequencies) of individual genomes within a metagenome was obtained based on alignments of short reads to SLRs. For the CAMI dataset, we compared the constructed plots with the blue plot representing the reference genomes with known abundancies.

Estimating Metagenome Capacity Using Long Error Prone SMS Reads. Although SMS reads (e.g., reads generated using Pacific Biosciences and Oxford Nanopores technologies) are still rarely used for analyzing metagenomes [9], they have a potential to be widely used in future metagenomics projects when their cost reduces and when the read until technology [30] developed by Oxford Nanopores becomes widely available. Below we show how to extend our approach for estimating the metagenome complexity using SMS reads.

SMS reads present an attractive alternative to TSLRs since their average length is higher and since they feature a uniform coverage depth that is not affected by the GC content. However, alignment of short Illumina reads against error-prone SMS reads is a more challenging task than their alignment against accurate TSLRs. We addressed this complication using the bowtie2 alignment tool [24] with specially selected parameters aimed at alignment of short Illumina reads against error-prone SMS reads (-D 40 -R 3 -N 0 -L 17 -i S,1,0.50 –rdg 1,3 –rfg 1,3 –score-min L,-0.6,-1 -a). However, even using these custom parameters, bowtie2 fails to detect alignments of ${\approx }20\%$ of Illumina reads, resulting in an underestimation of the heights of long reads. To compensate for this effect, we applied an adjustment factor $\frac{100}{100-20}=1.25$ to artificially inflate the heights in our formula for estimating the metagenome capacity.

Currently, there is a shortage of publicly available hybrid metagenomics datasets (containing both Illumina and SMS reads). Ideally, Illumina and SMS reads for such datasets should be generated at the same time so that the abundances of individual genomes within a metagenome are the same for Illumina and SMS reads, implying that the depth of coverage by Illumina reads is proportional to the depth of coverage by SMS reads. In practice, since the SMS reads for these datasets were often generated as an afterthought, Illumina and SMS reads for the publicly available hybrid metagenomics datasets are generated at different time points and prepared for sequencing using different sample preparation protocols. Thus, since metagenome composition is changing and is subject to blooms [33], the existing hybrid datasets do not necessarily feature the proportional depths of coverage by Illumina and SMS reads. Our analysis revealed that the fractions of Illumina and SMS reads aligned to each of the reference genomes for publicly available hybrid synthetic metagenomic dataset may differ by two orders of magnitude. This difference in the genome coverages by short and long reads in the publicly available hybrid metagenomics datasets makes our approach inapplicable to the currently available hybrid metagenomics datasets.