Appendix: Shotgun Metagenomics (MOSHPIT cocoa)

Appendix: Shotgun Metagenomics (MOSHPIT cocoa)#

A shotgun metagenomics dataset reaches q2-gglasso and q2-classo as a feature table like any other. The MOSHPIT cocoa fermentation study is the example. No feature table ships with this book for that study, and none of the commands below were run.

The dataset#

Cocoa bean fermentation, sampled as 14 shotgun metagenomes: two seed types (Forasteiro and Hybrid) at seven timepoints (0, 24, 48, 72, 96, 120 and 144 hours). Sequencing is Illumina HiSeq X Ten whole-genome shotgun, deposited under BioProject PRJNA552479.

\(2 \times 7 = 14\) gives exactly one sample per condition, with no within-cell replication, and 14 samples is a small number in an absolute sense regardless of how the conditions are arranged.

Where a feature table comes from#

MOSHPIT is the QIIME 2 distribution for shotgun metagenomics. Three routes through it end in a FeatureTable[Frequency]:

  • Read-based taxonomic profiling with Kraken2, refined by Bracken. Kraken2 assigns reads to taxa against a k-mer database; Bracken redistributes them to a chosen rank and produces per-sample abundance estimates. The result is a table of taxa × samples.

  • Read-based profiling with Kaiju, which classifies against a protein database instead. Same shape of output, different assignment behaviour.

  • Assembly and binning. Assemble the reads, bin the contigs into MAGs, dereplicate the bins across samples, then estimate per-sample abundance of each MAG — typically as TPM. The features are genome bins rather than named taxa.

Action names for those steps change between releases. Take them from the MOSHPIT documentation for the release you are running.

Compatibility with q2-gglasso and q2-classo#

Both plugins consume a FeatureTable and neither inspects its provenance:

  • qiime gglasso transform-features accepts FeatureTable[Composition | Frequency | Design]. A Bracken, Kaiju or MAG table registered as FeatureTable[Frequency] goes straight in, and the rest of the chain — calculate-covariance, solve-problem, summarize — is unchanged from Data Preparation.

  • qiime classo transform-features likewise takes the table as --i-features and the workflow proceeds as in Data Preparation.

Three caveats attach to shotgun-derived tables specifically:

  1. TPM is not a count. MAG-abundance tables are already normalized for genome length and depth. The compositional argument for a log-ratio transform still applies — TPM values are relative by construction — but --p-pseudo-count and --p-coef, which are written as count-like floors, no longer have an obvious scale. --p-coef (q2-classo) is registered as a Float, so you can set it below the smallest non-zero TPM value in your table. --p-pseudo-count (q2-gglasso) is registered as an Int, so it has a hard floor of 1 and cannot take the fractional value a TPM table usually needs. For such tables prefer qiime gglasso transform-features --p-transformation mclr, which log-transforms only the positive entries and needs no floor at all. mclr is a q2-gglasso option; qiime classo transform-features implements clr only, so on the classo side tune --p-coef.

  2. --i-taxonomy is still required by qiime gglasso transform-features and still unread. For a MAG table there may be no natural taxonomy to pass; any valid FeatureData[Taxonomy] will satisfy the input. See Troubleshooting.

  3. qiime classo add-taxa does need a real taxonomy, because trac builds its aggregation matrix from the hierarchy. Bracken output at a fixed rank is flat and gives the tree nothing to aggregate over; a MAG table needs taxonomic assignments attached first.

Data availability and sample size#

No feature table is published for this study. Producing one means running assembly and binning, or holding a local Kraken2 or Kaiju database — large reference resources that are impractical to bundle with a tutorial and slow to build from scratch. No files for this study appear in Download the Tutorial Data.

Fourteen samples cannot support covariance estimation. The empirical covariance matrix of \(n\) samples has rank at most \(n - 1\), so with \(n = 14\) every estimate lives in at most 13 dimensions while a Bracken table at species level has features in the thousands. A graphical lasso would still return a network, because that is what a penalized estimator does. The network would be almost entirely the penalty. The same applies to the log-contrast models — with one sample per condition, cross-validation folds contain one or two samples and the resulting error estimate is not usable.

Shotgun tables are compatible with both plugins. That compatibility is not a demonstration that they can be analyzed at this sample size. If you have a shotgun dataset with tens to hundreds of samples, the workflows in the high-dimensional Atacama chapters transfer directly — 16S rather than shotgun, but identical in every command.

Note

None of the MOSHPIT steps, tables or models described has been run against QIIME 2 2026.7 or any other release. No outputs, runtimes or database sizes are quoted.