B-229 Interlaboratory study using spike-ins and control samples to assess sample extraction and sequencing biases for metagenomics workflows

Abstract Background Sequencing bias remains a major obstacle to comparing metagenomic sample analyses and data reuse. Methodological variables impact the data, causing significant challenges to interpretating results from otherwise similar studies. However, controls in the form of spike-ins and common samples provide mechanisms for comparing between workflows to characterize biases. These allow methodological discrepancies to be resolved ahead of interlaboratory studies, and/or provide data to potentially reconcile differences from individual workflows. We initiated a small two-part interlaboratory study (ILS) to examine two questions: (a) does DNA extraction (5 methods) impact apparent sample composition; and (b) do DNA library/sequencing protocols (3 methods) have biases? Methods For ILS-a, participants were given 8 total samples (4x sample #1, 1x samples #2-#5) consisting of human stool spiked with a mixture of whole cells (200k/uL total of S. aureus, S. enterica, E. coli, L. monocytogenes, P. aeruginosa) and DNA internal controls (34k genome/uL ea. A. hydrophila & L. pneumophila). Participants extracted the DNA and returned the samples to NIST for sequencing. For ILS-b, participants were given 8 total samples consisting of DNA extracted from the same human stool samples in (a), spiked with DNA internal controls at ∼50k copy/uL/strain and spike-ins at ∼75k copy/ul/strain (see above). Labs generated DNA libraries, sequenced the samples, and returned the fastq files to NIST for processing. For (a) and (b), kraken2 was used to taxonomically classify the reads, reporting at the genus level. Relative abundance (reads / total reads) and Normalized abundance (reads / internal control reads) were used to examine the spike-ins and native taxa across the 5 samples. Results ILS-a (extraction) showed significant extraction bias between no change and 5-fold, with the spike-ins and native taxa mimicking similar trends in Gram +/- behavior. ILS-b (DNA) also showed significant bias vs. genome GC-content from different DNA library preparations (see Figure). These biases were reproducible between labs. Within-lab reproducibility of the 4 sample #1 replicates was 10-16% (a) and 9-18% (b), and the spike-in controls’ normalized abundances were consistent within lab across the 5 samples. This showed that the biases were sample composition-independent, and the biases were both reproducible and systematic. Conclusion Spike-ins and common-sample controls elucidate biases (and harmonization) between workflows, and indicate where data will likely have comparability challenges. The biases observed with the spike-ins were similar to the native taxa, such that a small number of well-characterized organisms helped account for biases across many native taxa. Hence, even small numbers of spike-ins provide a useful tool for assessing method bias, and indicate when more thorough method characterization may improve data intercomparability.