Case study: cFMD

TabFlux ships a worked example that needs no data of your own: the curated Food Metagenomic Data repository (cFMD), a public compilation of food metagenomes, profiled from reads with MetaPhlAn 4. The feature is the species-level genome bin (SGB).

One question runs through the four pages below. Which food category did this sample come from — dairy, meat, fish, fermented_beverages and so on — given nothing but its SGB profile.

Which counts mean what

Release v1.3.2 lists 109 datasets and about 4,000 samples, 107 of which ship a profile table. The case study uses a filtered subset of that: drop the datasets with fewer than ten samples, keep the categories with fifty samples or more, and 3,680 samples across 57 datasets in 7 categories remain. Every result on these pages is measured on those 3,680.

Why this example

• Public. Every number on this site can be reproduced from a versioned tag.
• Grouped. Samples arrive in study-sized blocks: whole studies stay together in the five outer folds, so nothing is scored on a study it was also trained on.
• Unbalanced. Dairy alone is 2,467 of the 3,680 samples, the smallest category 75. Answering "dairy" to everything already scores 0.67 plain accuracy, so balanced accuracy — chance level 1/7 = 0.143 — is the column to read. TabPFN reaches 0.662 there, XGBoost 0.634, the random forest 0.626.

The four pages

• The database — what cFMD is, how releases differ, how TabFlux reads it.
• Running it — the configuration, and why the design is what it is.
• Results — what TabFlux finds, and what the honest design costs.
• Comparing learners — TabPFN, random forest and XGBoost under one decision rule.