Microbial Source Tracking

The core mathematical engine of tools like SourceTracker and FEAST is a Gibbs sampler operating on a Bayesian mixture model. The model treats an unknown 'sink' community as a convex combination of known 'source' communities, estimating the proportion of each source contributing to the sink. This probabilistic approach naturally handles uncertainty and provides posterior distributions for source contributions rather than single point estimates.

Source tracking can operate on two distinct feature spaces:

• Taxonomic composition: Uses Operational Taxonomic Units (OTUs), Amplicon Sequence Variants (ASVs), or species-level relative abundances as input features. This is the most common approach and works well when sources have distinct community profiles.
• Functional gene profiles: Uses the abundance of gene families (e.g., KEGG Orthology groups) rather than taxonomic labels. This approach can be more robust when tracking sources across environments where taxonomy varies but metabolic function is conserved.

FEAST (Fast Expectation-mAximization microbial Source Tracking) improves upon earlier methods by using an Expectation-Maximization (EM) algorithm instead of Gibbs sampling. This provides orders-of-magnitude speed improvements, scaling to thousands of sources and millions of sequences. FEAST also introduces a source-sink similarity metric that quantifies the overall relatedness between a sink and each potential source before decomposition.

A critical preprocessing step in modern MST workflows is sink-seeded source selection. Rather than including all available sources in the mixture model, which can lead to overfitting and spurious assignments, this method first identifies a subset of sources that share taxa with the sink sample. This is often implemented using FracMinHash sketches to rapidly estimate the containment index between the sink and each candidate source, retaining only those above a threshold.

A key advantage of Bayesian MST methods is the explicit modeling of an 'unknown' source. The model includes a latent source category that accounts for taxa observed in the sink but absent from all provided sources. The estimated proportion assigned to this unknown category serves as a diagnostic: a high unknown proportion indicates that the true source environment was not sampled, preventing false attribution to an incorrect but available source.

Rigorous validation of MST pipelines requires in silico mock communities with known mixing proportions. Synthetic sink samples are generated by computationally mixing reads from pure-culture or single-source metagenomes at defined ratios. The MST tool's predicted proportions are then compared against the ground truth using metrics like root-mean-square error (RMSE) and Pearson correlation, quantifying accuracy across a range of mixture complexities and sequencing depths.