Bio Metagenomics Metaphlan

Version Compatibility

Reference examples tested with: MetaPhlAn 4.1+, Bowtie2 2.5.3+, minimap2 2.26+, pandas 2.2+.

Before using code patterns, verify installed versions match. If versions differ:

• CLI: metaphlan --version then metaphlan --help to confirm flag names and defaults
• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

The marker DATABASE version is the experimental variable. Results track the index (e.g. mpa_vJun23_CHOCOPhlAnSGB_202403 vs the live vJan25 build); MetaPhlAn 3 and MetaPhlAn 4 databases are not interchangeable. Pin --index and report it like a reagent lot. Two flags were renamed in 4.2: --bowtie2out -> --mapout and --bowtie2db -> --db_dir (the --input_type value bowtie2out likewise becomes mapout); unknown-fraction estimation flipped from opt-in (--unclassified_estimation) to on-by-default (--skip_unclassified_estimation to disable). Confirm against metaphlan --help.

MetaPhlAn Profiling

"Who is in my metagenome, by cell fraction?" -> Detect which clades' private marker genes are present, average their per-marker coverage, and normalize to a genome-size-aware relative abundance - so the percentage is a fraction of cells, not of reads.

• CLI: metaphlan reads_1.fq.gz,reads_2.fq.gz --input_type fastq --index mpa_vJun23_CHOCOPhlAnSGB_202403 -o profile.txt --mapout sample.bz2

Scope: marker-gene species/SGB profiling and its alternatives (mOTUs3, sourmash gather). K-mer read classification -> kraken-classification. Strain-resolved SNV haplotypes -> strain-tracking. Functional profiling -> functional-profiling. Compositional stats and plotting -> metagenome-visualization. 16S amplicon -> the microbiome category.

The Single Most Important Modern Insight -- A MetaPhlAn Percentage Is a Cell Fraction, Not a Read Fraction

A MetaPhlAn percentage estimates what fraction of the CELLS in the community belong to a clade - a genome-size-normalized taxonomic abundance. A Kraken/Bracken percentage estimates what fraction of the READS came from a clade - a sequence abundance. There is no sample-independent conversion between them, because sequence abundance under-estimates small-genome microbes and over-estimates large-genome ones by a factor that depends on the whole community's genome-size distribution (Sun 2021 Nat Methods 18:618). Therefore:

• Never merge MetaPhlAn percentages with Kraken/Bracken percentages into one table, correlate them, or benchmark one against the other. Disagreement between them is expected even when both are correct.
• Marker profiling is not "classify every read." It detects which clades' PRIVATE markers are present (default presence gate: reads cover roughly 20% of a clade's markers) and averages their per-marker coverage. Most reads are never assigned - by design, not failure.

Mnemonic: markers measure WHO is there (cells); k-mers measure HOW MUCH DNA is there (reads).

SGBs: the Unit Is Species-Level, and uSGBs Quantify the Unnamed

MetaPhlAn 4's atomic taxon is the SGB (species-level genome bin, a ~95% ANI cluster), not an NCBI species. A kSGB contains a cultured reference genome and gets a Latin name; a uSGB is defined only from MAGs (>=5 required) and is reported with a placeholder ID and no name. Quantifying uSGBs - taxa with no reference genome - is MetaPhlAn 4's headline advance over MetaPhlAn 3 and explains ~20% more gut reads, >40% more in under-characterized environments (Blanco-Miguez 2023 Nat Biotechnol 41:1633). Consequences: an unnamed t__SGB... row is a real quantified taxon - do not drop it; one named species can split into several SGBs; MetaPhlAn 3 species profiles and MetaPhlAn 4 SGB profiles are not row-compatible (use sgb_to_gtdb_profile.py for GTDB names). The t__ tier is the SGB, NOT a strain - strain resolution is StrainPhlAn (-> strain-tracking).

Tool Taxonomy

Decision Tree by Scenario

Basic Profiling

# Paired-end reads are passed as ONE comma-separated argument (MetaPhlAn treats them as two
# single-end files - it does not use insert/pairing info). Pin the index for reproducibility.
metaphlan reads_R1.fastq.gz,reads_R2.fastq.gz \
    --input_type fastq \
    --index mpa_vJun23_CHOCOPhlAnSGB_202403 \   # pin it; DB version is a batch variable
    --nproc 8 \
    --mapout sample.map.bz2 \                    # cache the read->marker mapping (pre-4.2: --bowtie2out)
    --output_file profile.txt

Re-Profile from the Mapping Cache (the real operational lever)

Goal: Try different analysis types, levels, or estimator settings without realigning.

Approach: Save the mapping once with --mapout, then re-run from it with --input_type mapout (pre-4.2: bowtie2out). Realignment is the expensive step; everything downstream is free.

metaphlan sample.map.bz2 --input_type mapout \
    --tax_lev s \           # k,p,c,o,f,g,s,t (t = SGB tier)
    --stat_q 0.2 \          # quantile-truncated robust mean of per-marker coverages: drop top/bottom 20%, average the middle 60%
    --output_file profile_species.txt

--stat_q down-weights markers in HGT/mobile and conserved cross-clade regions; the default 0.2 is a sensible robust mean. Changing it changes the reported abundances - report it if it is changed. Long reads (4.1+) route to minimap2 with --long_reads.

The Unknown Fraction Rescales Everything

Relative abundance sums to 100% only over DETECTED clades. With unknown estimation OFF (pre-4.2 default), known taxa absorb 100% and the database-absent community is invisible - overstating every known taxon. With it ON (4.2 default), an UNCLASSIFIED row appears and every known abundance shrinks proportionally. In soil/marine/rumen the unknown fraction can be the largest "taxon" in the sample.

# 4.2 default includes the UNCLASSIFIED row. To force it on pre-4.2: --unclassified_estimation
# For SAM input, pass --nreads <total> or the unknown fraction is wrong.
metaphlan reads.fastq.gz --input_type fastq -o profile.txt   # 4.2: UNCLASSIFIED row present by default

Pre-4.2-default and 4.2-default outputs are not comparable abundances - mixing them is a hidden batch effect.

Merge and Convert

# All inputs MUST come from the SAME database index or rows mismatch silently.
merge_metaphlan_tables.py profiles/*_profile.txt > merged_abundance.txt
sgb_to_gtdb_profile.py -i merged_abundance.txt -o merged_gtdb.txt   # recover GTDB names for SGBs

Per-Method Failure Modes

MetaPhlAn percentages merged with Kraken percentages

Trigger: putting MetaPhlAn and Bracken abundances in one matrix or correlating them. Mechanism: cell fraction vs read fraction - different quantities (Sun 2021). Symptom: "tools disagree," spurious scatter, broken ML/differential-abundance features. Fix: keep them separate; if harmonizing, convert via genome length (Bracken counts / genome length, renormalize) and accept it is approximate.