microbiome

Microbiome Analysis

Comprehensive guidelines for amplicon and shotgun metagenomics, with emphasis on compositional data analysis, diversity metrics, differential abundance, and functional profiling.

COMPOSITIONALITY WARNING — READ FIRST

Microbiome count data is compositional. Sequencing produces counts that sum to a constant (library size). Each sample's counts represent relative proportions, NOT absolute abundances. This has profound statistical consequences:

• An increase in one taxon's proportion forces decreases in others, even if their absolute abundance is unchanged.
• Standard statistical tests (t-test, Wilcoxon, correlation) on raw counts or proportions produce spurious correlations and false positives.
• Methods designed for gene expression (DESeq2, edgeR) assume count distributions that do NOT hold for compositional, zero-inflated microbiome data.

You MUST use compositional-aware methods for all statistical analyses. This is not optional — it is a fundamental requirement of the data type. See the Differential Abundance section below for approved methods.

Method-Selection Decision Tree

1. 16S/ITS Amplicon Sequencing

Demultiplexed FASTQ
  → DADA2 (R): filterAndTrim → learnErrors → dada → mergePairs → makeSequenceTable → removeBimeraDenovo
    → ASV table + taxonomy assignment (assignTaxonomy with SILVA for 16S, UNITE for ITS)
      → Import into phyloseq (R) for downstream analysis

Alternative: QIIME2 artifacts
  → Import pre-processed QIIME2 feature table + taxonomy
    → Convert to phyloseq for R-based analysis or pandas DataFrame for Python

• DADA2 is installed, so raw demultiplexed FASTQ is a supported entry point: the chain above runs end to end through removeBimeraDenovo and yields an ASV table, the ASV sequences, and per-step read tracking. Taxonomy assignment is the one step that may not complete — it needs a training set, which is a separate question covered below. QIIME2 is not installed, so that branch is import-only: bring an exported feature table rather than expecting a .qza to be read natively. A pre-computed feature table (BIOM, QIIME2 export, or counts matrix plus taxonomy) is still a perfectly good entry point when that is what you were handed — the rest of this pack (phyloseq, vegan, mia, ANCOM-BC2, ALDEx2, MaAsLin2, biom-format, scikit-bio) runs from it normally.
• DADA2 produces ASVs (amplicon sequence variants) — single-nucleotide resolution, no OTU clustering needed.
• ALWAYS inspect the error rate learning plots. Poor error models produce unreliable ASVs.
• SILVA is the standard reference for 16S; UNITE for ITS (fungal). Resolve the release from the inventory rather than assuming a version.
• SILVA (16S/18S) and UNITE (ITS) training sets are in the reference inventory, but as opt-in downloads rather than part of a default install — so they are resolvable, and may or may not actually be present. Resolve them by what they are before planning on them. If they are absent, report that taxonomy assignment cannot be run, say they are available to provision, and hand back the ASV table, sequences, and read tracking — which are complete and useful without it. Do not invent a path, do not substitute a general-purpose sequence database, and do not drop the taxonomy step silently.