The database

The curated Food Metagenomic Data (cFMD; Carlino et al., 2024) is a public compilation of food metagenomes: release v1.3.2 lists 109 datasets and 4,125 samples, every one profiled with MetaPhlAn 4, which maps reads against marker genes and reports an abundance per species-level genome bin (SGB). The SGBs come from clustering assembled genomes, but that happened when the catalogue was built; these samples are not assembled or binned. cFMD also distributes StrainPhlAn strain profiles and HUMAnN functional profiles; TabFlux reads the taxonomic ones. It ships one tab-separated profile per dataset plus one metadata table, on GitHub, under a versioned tag. Two of the 109, KawaiT_2012 and RippF_2014, have metadata but no profile file, so datasets: [] reads the 107 that have one.

What a sample carries

metadata column	example values
dataset	the study: MASTER_WP4_CSIC_1, PasolliE_2020, …
macrocategory, category, type, subtype	food → dairy → cheese → soft cheese
fermented_non-fermented	a curated fermentation flag, carried alongside the category columns
country, sample_accession, run_accession	provenance

cFMD ships percentages: every sample sums to 100. The depth step (preprocessing.normalization, tss_log here) divides each sample by its own total, so on cFMD it rescales 100 to 1 before the log. The depth report shows 100 for every sample.

How the categories are distributed

Among the 65 datasets with at least ten samples (3,886 samples), fifteen categories occur, but seven of them hold fifty samples or more, and dairy alone is two thirds of everything:

category	samples	datasets
dairy	2,467	38
fermented_beverages	422	4
fruits_and_vegetables	224	10
meat	197	7
fermented_meat	154	7
fish	141	4
fermented_grains	75	3

Those seven categories are what the case study predicts: 3,680 samples across 57 distinct datasets, 11 of which carry more than one of the seven.

The datasets column is the one that constrains the design: a category living in three studies can lose all three to a single test fold, and is then missing from the training data of that fold.

Two facts about the compilation drive the design of the runs: 96 of the 109 datasets contain a single category, and the rare categories live in three or four datasets each.

Releases, and why the tag is pinned

The file format has moved between releases: v1.3.0 re-profiled every sample with a newer database and left a stray header line in some files, v1.3.1 removed the metadata rows that earlier profiles carried inside the table, v1.3.2 renamed the first column of the 22 new datasets to clade_name. TabFlux reads by position and keeps only the lineage rows, so all three parse; input.cfmd.ref is pinned to a tag so a run means one thing.

How TabFlux reads it

One anonymous GitHub API call lists the release. Each profile is downloaded once and cached by content hash under input.cfmd.cache_dir, so a rerun downloads nothing; when the API is unreachable and the data are already cached, the module works from its manifest. The profiles are merged on the lineage, samples and datasets below the size filters are dropped, and three files are written in TabFlux's own layout: counts, taxa (the lineage split into ranks, the SGB kept as the feature), metadata. Feature ids read Genus_species__SGB1234.