How Scientists Track Live Gut Microbes After Fecal Transplants and Probiotic Drug Therapies

Live microbial therapeutics, including fecal microbiota transplantation (FMT) and defined live biotherapeutic products (LBPs), have produced two approved drugs for recurrent C. difficile infection. Early signals exist for ulcerative colitis, immunotherapy-related colitis, cancer, and decolonization of antibiotic-resistant organisms, but results are inconsistent. This perspective argues that progress toward designed consortia needs standardized measures of drug composition, potency, dose, and strain engraftment. Such data could show whether clinical response depends on total engraftment, specific strains, transferred functions, or other factors such as transient exposure to microbial molecular patterns.

The paper first defines a bacterial strain. Isolates from one person's gut form tight genomic clusters, and different people's isolates are separated by deep branches. Within-strain isolates typically share at least 96–98% of kmer content, which corresponds to roughly 99.8–99.9% average nucleotide identity (ANI). Thresholds can be tightened to avoid false positives or loosened to avoid false negatives. Because most gut species have large strain populations, strains can usually be tracked across thousands of recipients with little chance of a coincidental match.

The author then reviews detection approaches. Genus-level (16S rRNA) or species-level (shallow metagenomic) profiling is easiest and works when the recipient's microbiome is very depleted, as in rCDI. It is confounded by antibiotic suppression, environmental sources, and recipients who already carry donor taxa. Culturing donor and recipient isolates with whole-genome comparison is the gold standard, as in studies that re-isolated 48 donor strains from five recipients weeks after FMT. It is labor-intensive, unscalable, and limited to culturable organisms.

Genome-based metagenomic tracking has two forms. In closed ecosystems (in vitro communities, gnotobiotic mice), shared genomic regions are masked and reads mapped to unique regions give relative abundance. This is straightforward but lacks statistical presence/absence calling and fails when the community is incompletely known. In open ecosystems, the rarest 1–5% of each strain genome, judged against large genome and metagenome collections, serve as identifiers. Detection uses depth relative to negative controls or depth–breadth Poisson relationships. This is sensitive even at under 1X coverage and suits defined LBPs, biobank donors, and metagenome-assembled genomes (MAGs), though it needs strain-specific training and MAGs can contain assembly errors.

Metagenome-only methods infer strains from single-nucleotide variant overlap against marker genes or reference genomes. They make use of abundant existing datasets, but need deep sequencing (about 10X coverage; a genome at 0.1% abundance would need ~300 million paired-end reads), have poor sensitivity below 1% abundance, are harder to interpret without an isolate, and cannot readily count strains per species. The text provided ends mid-discussion, so later sections on trial design are not covered here.