Data card — mouse 16S microbiome

About the experiment

This teaching dataset comes from a classic murine gut microbiome study used in the Schloss lab MiSeq SOP and repackaged for machine learning by the Quadram Institute. The underlying experiment is observational and longitudinal, not a randomized trial: lab mice were sampled repeatedly after weaning to see how the fecal bacterial community changes over time.

Biological setup (Schloss cohort). Researchers collected fresh feces from mice on many days after weaning (the full published study follows animals for up to a year). The scientific motivation is to understand normal variation in the gut microbiome and whether community composition is stable or shifts between life stages — for example, comparing an early post-weaning period (rapid growth) with a later period when weight and microbiota are more stable. Our CSVs are a subset of that larger resource, chosen so a classroom can download tables quickly and still face a realistic p ≫ n prediction problem.

What was measured. For each fecal sample, DNA was amplified with primers targeting the V4 region of the 16S rRNA gene and sequenced on an Illumina MiSeq (paired-end reads). Bioinformatic processing (in the original work, typically with mothur) turns millions of short reads into a feature table: each column is an OTU (operational taxonomic unit — a cluster of similar 16S sequences treated as one bacterial lineage), and each cell is the number of reads assigned to that OTU in that sample. Optional taxonomy maps OTU IDs to names (genus, family, etc.).

What you hold in R. After loading, each row is one sample (one mouse on one Day), and predictors are sparse count data (many zeros, few very abundant OTUs). The main classification label in the tutorial is Label: Early vs Late, a coarse time grouping (early post-weaning days vs later days — in this subset, Early samples are mostly Day ≤ ~65, Late samples mostly Day ≥ ~125). You can also model Sex (F / M). The same mouse reappears under Individual across days, so samples are not independent.

How these CSVs were produced. The Quadram workshop distributes already aggregated tables (metadata.csv, otutab_raw.csv) derived from the Schloss/MiSeq pipeline, so you do not re-run mothur in our labs. You join metadata to the OTU matrix, transpose so samples are rows, optionally filter rare OTUs, and log1p-transform counts before PCA or modeling — the same statistical habits as for gene counts, but with ecology-flavoured column names.

Citation. If you use this data in a report, cite the Schloss MiSeq SOP / Kozich et al. 2013 AEM paper (see the mothur wiki) and note that you accessed the Quadram GitHub mirror on the date you downloaded it.

How to load

Public CSVs from the Quadram Institute workshop (see URLs below). Join metadata and OTU table on sample ID; transpose to samples x OTUs. Task details on the lab exercises page.

File URL
Metadata https://raw.githubusercontent.com/quadram-institute-bioscience/datasciencegroup/main/4_machine_learning/mouse-16s/metadata.csv
OTU table https://raw.githubusercontent.com/quadram-institute-bioscience/datasciencegroup/main/4_machine_learning/mouse-16s/otutab_raw.csv
Taxonomy (optional) https://raw.githubusercontent.com/quadram-institute-bioscience/datasciencegroup/main/4_machine_learning/mouse-16s/taxonomy.csv

Source repo: Quadram — ML with microbiome (Python workshop; we use R + tidymodels).

Where we use this

Caveats

ImportantLeakage warning

Samples are repeated measures per mouse (Individual). Do not treat a random row split as your final answer without discussing it.

  • Research-grade (Day 4 solution): group_vfold_cv with group = Individual.
  • Day 2 intro track: initial_split on rows — instructor explains why this can inflate scores.

Data card

1. Source

Schloss mouse 16S — longitudinal fecal microbiome cohort processed from Illumina MiSeq amplicon data (mothur MiSeq SOP). We use the Quadram datasciencegroup mirror (CSV feature table + metadata) for teaching. Not collected for this summer school; reused to practise high-dimensional count classification and interpretation.

2. Outcome

  • Label: factor Early vs Late (community shift over time in the experiment).
  • Sex: F / M (Day 1 afternoon also models this outcome).

3. Predictors

  • OTU abundances (thousands of columns after transpose); typically log1p transformed in recipes.
  • Metadata columns (e.g. Individual, library size) — use for grouping and QC, not always as predictors.

4. Sample size

Roughly 282 samples and ~13 mice (Individual) after joining metadata and OTU table (exact OTU count depends on prevalence filtering; default loader keeps OTUs in ≥ 10% of samples → on the order of 300+ columns). p >> n. Library sizes and sparsity vary; explore in Day 1 tasks.

5. Leakage risks

  • Random row CV when the same mouse (Individual) appears in train and test inflates performance.
  • Normalizing or filtering OTUs on the full table before splitting leaks.
  • Document any simplification (e.g. one row per mouse) on your write-up.

6. Subgroups

  • Individual (mouse ID) — primary grouping for honest resampling.
  • Sex, Day — secondary stratification or second outcome (Day 1).

Back to all datasets · Lab exercises