Metagenomic source tracking after microbiota transplant therapy

Abstract

Reliable engraftment assessment of donor microbial communities and individual strains is an essential component of characterizing the pharmacokinetics of microbiota transplant therapies (MTTs). Recent methods for measuring donor engraftment use whole-genome sequencing and reference databases or metagenome-assembled genomes (MAGs) to track individual bacterial strains but lack the ability to disambiguate DNA that matches both donor and patient microbiota. Here, we describe a new, cost-efficient analytic pipeline, MAGEnTa, which compares post-MTT samples to a database comprised MAGs derived directly from donor and pre-treatment metagenomic data, without relying on an external database. The pipeline uses Bayesian statistics to determine the likely sources of ambiguous reads that align with both the donor and pre-treatment samples. MAGEnTa recovers engrafted strains with minimal type II error in a simulated dataset and is robust to shallow sequencing depths in a downsampled dataset. Applying MAGEnTa to a dataset from a recent MTT clinical trial for ulcerative colitis, we found the results to be consistent with 16S rRNA gene SourceTracker analysis but with added MAG-level specificity. MAGEnTa is a powerful tool to study community and strain engraftment dynamics in the development of MTT-based treatments that can be integrated into frameworks for functional and taxonomic analysis.

Keywords: Metagenomics; fecal microbiota transplant; metagenome-assembled genome; microbiota engraftment; microbiota transplant therapy; shotgun sequencing; ulcerative colitis; whole-genome sequencing.

Figure 1.

(a) Schema for randomized, double-blinded, placebo-controlled trial of microbiota transplant therapy (MTT) for participants with mild-to-moderate ulcerative colitis. Participants were administered placebo capsules or encapsulated MTT from a healthy donor daily for 8 weeks. Stool samples for metagenomic sequencing were collected from participants at baseline (week 0), and weeks 4, 8, and 12. (b) A visual depiction of pre-treatment (yellow), donor (blue), and post-treatment (green) strains as a Venn Diagram. While the engrafted donor strains constitute the focus of our manuscript, further information on MTT impacts can also be found in determining the likelihood of ambiguous strains being engrafted or persistent. (c) A visual depiction of the MAGEnTa pipeline. We begin with using metaSpades and Bowtie 2 to construct metagenome assembled genomes (MAGs) databases from the pre-mtt and donor samples. We then align the post-MTT sample with these databases to determine which reads match the donor and the pre-mtt (baseline) samples. Ambiguous reads matching both databases are evaluated by a Bayesian estimation to assign the likely source. Read source outputs are combined into proportions, assigning each read to engrafted, persistent, or novel/unknown (unmapped). This figure was created in BioRender.com. DB (database), MAGs (metagenome assembled genomes), MTT (microbiota transplant therapy).

Figure 2.

(a) Creation of 15 simulated datasets totaling 165 samples. For each of 15 donor and pre-mtt pairs in a CDI study, we obtain 11 samples comprised of different percentages of the donor and pre-microbiota transplant therapy (MTT) material. All samples are an even depth of 10 million reads. (b) From the 50% donor and 50% pre-mtt simulated samples, we also simulated the inclusion of an external, “contaminate” source. For this simulation, our external source is another healthy donor unrelated to the original donor and pre-MTT pair. This simulation seeks to represent other microbes introduced to the patient by external means such as food or environment. (c) Engraftment resulting from 15 simulations, displaying simulated (x-axis) vs. observed (y-axis), separated and colored by alignment identity. The expected outcome would be an x=y line as shown in gray, which the observed results closely follow. (d) The recall error of the algorithm per identity with the 50% donor and 50% pre samples, showing the false negative rate (type II error) per simulation, grouped by identity. Less stringent identity has a slightly lower type II error rate. (e) The accuracy error when 20% of the simulated 50% donor sample comes from a contaminate source in the form of an unrelated donor sample. The more stringent the identity, the fewer false positives (type I error).

Figure 3.

(a) Boxplot showing the percent of donor engraftment per group at each week. The greater spread in the intervention group is likely due to the variances in success of engraftment following the encapsulated MTT course. (b) Agreement between methods per sample. Whole-genome sequencing (WGS) MAGEnTa pipeline results are on the x-axis and 16S rRNA gene SourceTracker results on the y-axis. The intervention samples (blue) have high agreement (Pearson correlation, p < 0.001), and placebo samples have no agreement (Pearson correlation, p = 0.45); however, poor placebo sample agreement may be poor due to the low engraftment observed. (c) Percentage of donor engraftment for each patient and its matched placebo donor. Donor engraftment in the microbiota transplant therapy (MTT) group is significantly higher at weeks eight and 12 compared to placebo. Three donors were used and participants received donor material from only one donor. MTT is shown in blue and placebo in pink. Some participants are missing samples (predominantly from week 4), but the intervention of daily oral MTT began after collecting the baseline sample at week 0 and ran for the first 8 weeks, collecting samples every 4 weeks. (d) MetaBinner results from binning engrafted metagenome assembled genomes (MAGs) shown as a percentage of engrafted reads for weeks 4, 8, and 12 samples receiving material from Donor A. The top 10 MAG bins are distinctly colored to show relative abundance of reads aligning with these bins across all engrafted reads.