Strain Tracking Reveals the Determinants of Bacterial Engraftment in the Human Gut Following Fecal Microbiota Transplantation

Abstract

Fecal microbiota transplantation (FMT) from healthy donor to patient is a treatment for microbiome-associated diseases. Although the success of FMT requires donor bacteria to engraft in the patient's gut, the forces governing engraftment in humans are unknown. Here we use an ongoing clinical experiment, the treatment of recurrent Clostridium difficile infection, to uncover the rules of engraftment in humans. We built a statistical model that predicts which bacterial species will engraft in a given host, and developed Strain Finder, a method to infer strain genotypes and track them over time. We find that engraftment can be predicted largely from the abundance and phylogeny of bacteria in the donor and the pre-FMT patient. Furthermore, donor strains within a species engraft in an all-or-nothing manner and previously undetected strains frequently colonize patients receiving FMT. We validated these findings for metabolic syndrome, suggesting that the same principles of engraftment extend to other indications.

Keywords: C. difficile; Clostridium difficile; FMT; bacterial engraftment; fecal microbiota transplant; fecal transplant; human microbiome; human microbiota; strain inference; strain tracking.

Figure 1:. Study design for the use of fecal microbiota transplantation to treat recurrent Clostridium difficile infection.

Nineteen patients with recurrent Clostridium difficile infection were treated with feces from one of four donors. Stool samples were collected and FMTs were performed at the indicated time points. For each patient, we show the class of antibiotics they were most recently treated with and the success of the overall treatment.

Figure 2:. A machine learning model predicts the gut microbiota of the post-FMT patient.

Samples were clustered according to their species compositions (dendrograms). Samples from the same patient are connected with colored arcs. (A) Donor samples (white dots) do not cluster with the corresponding post-FMT patient samples (black dots). (B) Pre-FMT patient samples (white dots) do not cluster with the corresponding post-FMT samples (black dots). (C) The receiver operating characteristic (ROC) curve for the model of mg-OTU presence. (D) Predicted abundances were correlated to the measured abundances in the post-FMT samples. (E) Samples from post-FMT patients (black dots) cluster perfectly with their predicted values (white dots). (F) The mean relative importance of each feature across both models. Because sequencing depth was only used in the model of OTU presence, only this feature importance is reported.

Figure 3:. The abundance and phylogeny of bacterial species are the strongest determinants of bacterial engraftment.

(A) The abundances of mg-OTUs in the donor are strongly correlated to their abundances in the post-FMT patient. (B) The abundances of mg-OTUs in the patient are strongly correlated before and after FMT. (C) The partial dependence of engraftment on each taxonomic order, reflecting the orders’ marginal effects on the probability of engraftment in the reduced model. Orders are arranged on the bacterial taxonomy, with phylum labels on the right.

Figure 4:. Strain Finder outperforms ConStrains on simulated metagenomic alignments generated from a set of Escherichia coli genomes.

We simulated 16 metagenomic alignments of 8 samples each, with varying numbers of strains (N = 4, 8, 12, and 16) and depths of coverage (N = 25, 100, 500, and 1,000X). ConStrains was run on each sample separately (CS Model 1) and on all samples combined (CS Model 2). (A) The weighted UniFrac distances from the true strain profiles to the predictions of Strain Finder, ConStrains, and Null Model 1. (B) The weighted UniFrac distances from the true strain profiles to the predictions of Strain Finder, both ConStrains models, and Null Model 2. Asterisks denote significant comparisons (*** = p-value < 1e-10) as determined by a Wilcox test. N.S. denotes non-significant comparisons.

Figure 5:. Complete sets of donor strains and previously undetected strains engraft in the patient after FMT.

(A) The strain compositions of mg-OTUs in the pre-FMT patient, the donor, and the post-FMT patient. The frequencies of strains that are unique to the donor, unique to the patient, shared by the donor and the patient, and undetected are shown on the right. (B) Strain specificity of mg-OTUs in the donor (N = 3,090) and the pre-FMT patient (N = 2,024). Strain specificity is measured as the log-ratio of (i) the distance from the donor (or pre-FMT patient) to the post-FMT patient, and (ii) the distance from an unrelated donor (or unrelated pre-FMT patient) to the post-FMT patient. (C) Across mg-OTUs, the percentages of strains from the donor (N = 3,090) and the pre-FMT patient (N = 2,024) that were transferred to the post-FMT patient. (D) The fraction of each community that is unique to the donor, unique to the patient, present in the donor and the patient, and previously undetected. Treatment success is provided on the right.

Figure 6:. Engraftment modeling is accurate in a meta-analysis of five FMT trials for the treatment of recurrent C. difficile.

Models of bacterial engraftment were trained on the 16S rRNA sequence data from five FMT trials for the treatment of recurrent C. difficile infection. (A) ROC curves for the predictions of OTU presence in each of the five FMT datasets (see legend in panel B). (B) Statistics describing the performance of each model, including the number of patients in the dataset, the AUC (for predictions of OTU presence), the r-squared (for predictions of OTU abundance), and the percent of predictions that correctly cluster with their target samples (see Methods). (C) Heatmap showing the partial dependence of the models of OTU presence on each bacterial order. High values indicate that the taxon has a favorable impact on engraftment.

Figure 7:. Shared phylogenetic principles drive models of engraftment for recurrent C. difficile infection and metabolic syndrome.

Partial dependence reflects the impact of each bacterial order on the predicted frequency of bacterial engraftment, with high values indicating that a taxon has a favorable impact on engraftment. Partial dependence values estimated for recurrent C. difficile infection and metabolic syndrome are strongly correlated (Kendall’s tau = 0.50, p-value < 1e-10). Prevalent gut commensals, such as Lactobacillales and Clostridiales, had consistently positive impacts on predicted levels of engraftment, while bacteria that are rarely found in the gut, such as Caulobacterales and Chromatiales, had consistently negative effects.