Towards precision quantification of contamination in metagenomic sequencing experiments

Abstract

Metagenomic next-generation sequencing (mNGS) experiments involving small amounts of nucleic acid input are highly susceptible to erroneous conclusions resulting from unintentional sequencing of occult contaminants, especially those derived from molecular biology reagents. Recent work suggests that, for any given microbe detected by mNGS, an inverse linear relationship between microbial sequencing reads and sample mass implicates that microbe as a contaminant. By associating sequencing read output with the mass of a spike-in control, we demonstrate that contaminant nucleic acid can be quantified in order to identify the mass contributions of each constituent. In an experiment using a high-resolution (n = 96) dilution series of HeLa RNA spanning 3-logs of RNA mass input, we identified a complex set of contaminants totaling 9.1 ± 2.0 attograms. Given the competition between contamination and the true microbiome in ultra-low biomass samples such as respiratory fluid, quantification of the contamination within a given batch of biological samples can be used to determine a minimum mass input below which sequencing results may be distorted. Rather than completely censoring contaminant taxa from downstream analyses, we propose here a statistical approach that allows separation of the true microbial components from the actual contribution due to contamination. We demonstrate this approach using a batch of n = 97 human serum samples and note that despite E. coli contamination throughout the dataset, we are able to identify a patient sample with significantly more E. coli than expected from contamination alone. Importantly, our method assumes no prior understanding of possible contaminants, does not rely on any prior collection of environmental or reagent-only sequencing samples, and does not censor potentially clinically relevant taxa, thus making it a generalized approach to any kind of metagenomic sequencing, for any purpose, clinical or otherwise.

Keywords: DNA; DNA contamination; Metagenomics; Microbiota; Regression analysis; Sequence analysis.

Fig. 1

Contaminant sequencing reads are inversely proportional to sample mass. For each of n = 32 HeLa input masses (present in triplicate), sequencing reads for the total ERCC set (n = 92 different transcripts) are normalized per million (rpm) and presented in green; sequencing rpm aligning to the E. coli genome are presented in blue; and sequencing rpm aligning to the S. cerevisiae genome are presented in red. The linear regressions associating sample input mass with ERCC, E. coli, and S. cerevisiae are described with the adjusted R² and p value

Fig. 2

Precision quantification of microbial contamination in sequencing experiments. For each of n = 32 HeLa input masses (measured in triplicate), microbial contaminants were identified if the inverse linear relationship associating log₁₀-transformed rpm of any given microbe with the log₁₀-transformed sample mass demonstrated an adjusted R² ≥ 0.7. By solving the equation contaminant mass/ERCC mass = contaminant reads/ERCC reads, the estimated mass of each contaminant in each sample was calculated. The top contaminating taxa were E. coli (2.59 ± 0.67 ag), S. cerevisiae (1.02 ± 0.30 ag), S. maltophilia (0.61 ± 0.49 ag), unspecified cloning vector (0.43 ± 0.17 ag), and A. xylosoxidans (0.40 ± 0.27 ag), respectively. The estimated mass of all contaminants (excluding human and low-quality reads) in each sample was 9.1 ± 2.0 ag

Fig. 3

Identification of outliers among contaminant microbes. Left: for each of n = 97 serum sample RNA input masses, sequencing reads for the total ERCC set (n = 92 different transcripts) are normalized per million (rpm) and presented in green; sequencing rpm aligning to the E. coli genome are presented in blue; and sequencing rpm aligning to the S. maltophilia genome are presented in grey. The linear regressions associating sample input mass with ERCC, E. coli, and S. cerevisiae are described with the adjusted R² and p value. Right: a histogram of the studentized residual for each observation informing the linear regression between log₁₀-transformed sequencing reads (E. coli in blue, S. maltophilia in grey) and log₁₀-transformed sample input mass. Studentized residuals approximate a near-normal distribution between − 2 and + 2 such that outliers can be rapidly identified (red)