0.5.0

Added

• bt.io.write_h5mu and bt.io.read_h5mu: save a MuData to an .h5mu file
  and read it back with each TreeData modality's trees kept (mudata alone
  writes a TreeData as an AnnData). biotapy now declares h5py as a
  dependency.

0.4.1

Fixed

• bt.ml.embed(..., "mgm"): the first embedding in a process no longer
  differs from later ones by up to 2e-4 on CPUs running several threads;
  biotapy fills oneMKL's CPU-type cache with one serial call before MGM's
  forward pass (pytorch/pytorch#188792).

0.4.0

Added

• bt.io.to_mudata: combine data types measured on the same samples into one
  MuData, keeping the samples every modality has (in the first one's order,
  with a warning naming how many each lost). Modality names: taxa,
  function, function_by_taxon, metabolites, host.
• bt.tl.mmvec: which metabolites go with which microbe (mmvec, Morton et
  al. 2019), fitted by scikit-bio, as a microbes x metabolites table of
  row-centred log probabilities.
• bt.ml.PrevalenceFilter and bt.ml.CLR: scikit-learn transformers, so a
  prevalence filter and CLR are fitted inside each cross-validation fold.
  They pass scikit-learn's estimator checks and give bt.pp.filter_features's
  and bt.pp.clr's results. Relative abundance is scikit-learn's own
  Normalizer(norm="l1").
• bt.ml.to_torch: a PyTorch dataset over a table's samples, densifying one
  row at a time, with integer or float labels from an obs column. It needs
  the new extra biotapy[torch] (torch>=2.9).
• bt.ml.embed: one embedding per sample from a pretrained model, returned or
  stored in obsm["X_<model>"]. Models are plugins: a package registers one
  in the entry-point group biotapy.embeddings, and biotapy checks what it
  returns.
• MGM, the Microbial General Model (Zhang et al. 2026), as the first such
  model, behind the new extra biotapy[mgm] (torch>=2.9,
  transformers>=5). Its pretrained weights (MIT) are downloaded once from
  the microformer-mgm 0.5.8 wheel on PyPI and checked against its SHA-256;
  biotapy's embeddings are within 2e-6 of MGM's own code.
• bt.datasets.biocrust: mmvec's soil example, microbes and metabolites of
  a desert biocrust after wetting, downloaded and cached on first use.
• Guide pages for multi-omics and machine learning, and three tutorials:
  multi-omics with mmvec, leak-free cross-validation on the HMP2 cohort, and
  MGM embeddings of GlobalPatterns.

Changed

• The pseudocount warning of bt.pp.clr and bt.pp.philr ends with the
  step that gave it, (pp.clr) or (pp.philr).
• bt.pp.filter_features raises TypeError naming min_prevalence or
  min_total when the threshold is a bool or not a number.
• bt.pp.tax_glom, bt.fn.func_glom and bt.pl.bar's fill sum a table
  of bools or integers narrower than 64 bits in 64 bits, as NumPy's sum
  does (unsigned stays unsigned): an int8 table no longer wraps around, and
  a bool table is counted instead of saturating at True. An int32 table
  now gives int64 sums.
• bt.da.ancombc2 documents that scikit-bio 0.7.4 can fail to estimate the
  bias with two samples in the reference level, where R returns results;
  biotapy raises naming scikit-bio
  (scikit-bio#2631).
• The Coming-from-R page marks the phyloseq calls biotapy does not cover yet
  "not in 0.4".

0.3.0

Added

• bt.pp.clr: centred log-ratio transform into layers["clr"], equal to
  vegan::decostand(x, "clr", pseudocount = 0.5).
• bt.pp.philr: PhILR balances along the phylogeny into obsm["X_philr"],
  equal to philr::philr with its default weights.
• bt.da.linda, bt.da.ancombc2: differential abundance by LinDA (a port
  of MicrobiomeStat::linda) and ANCOM-BC2 (through scikit-bio), both in
  Python and compared with R in golden tests.
• bt.da.aldex2, bt.da.maaslin3: ALDEx2 and MaAsLin 3's abundance model,
  run in R through rpy2. They need R, the R package, and the new extra
  biotapy[r] (rpy2>=3.6.8, GPL-2.0-or-later, built against your R).
• One result table for every method: log2 effect, se, pvalue,
  Benjamini-Hochberg qvalue over the features the method tested,
  direction, method and contrast. Methods take group, covariates
  and reference instead of a formula, and never filter features.
• bt.da.consensus and bt.pl.consensus: count, per feature, the methods
  that call it and whether they agree on its direction, as a table and as a
  dot matrix.
• A differential abundance guide, a page per method, and a tutorial on the
  GlobalPatterns genera.
• asv benchmarks for pp.philr, da.linda and da.ancombc2, with their
  baselines in the docs.

Changed

• The readers infer x_kind "counts" only for non-negative whole numbers,
  and the functions that need counts (pp.rarefy, tl.alpha's
  observed_features and chao1, weighted tl.unifrac) refuse a table with
  a negative value.
• bt.pl.bar and bt.pl.heatmap raise ValueError for a table with a
  negative value, such as layers["clr"], instead of drawing it as
  abundances.
• The Coming-from-R page marks the phyloseq calls biotapy does not cover yet
  "not in 0.3".

0.2.0

Added

• bt.io.read_humann: read HUMAnN 3 and 4 tables (gene families, reactions
  or pathway abundance, also after regrouping or renormalising) into a
  MuData with a community modality, "function", and a per-taxon one,
  "function_by_taxon".
• bt.io.read_metaphlan: read MetaPhlAn 3 and 4 profiles, one or merged, as
  relative abundances of the leaf clades with the seven taxonomic ranks.
• bt.io.read_picrust2 / bt.io.read_picrust2_traits: read PICRUSt2's
  predicted metagenomes and pathways, with the per-ASV contributions as the
  per-taxon modality, and its per-ASV gene copy numbers. PICRUSt2 itself is
  not needed.
• bt.fn.load_hierarchy: read a function hierarchy (HUMAnN or PICRUSt2
  mapping files, or your own) from a local file.
• bt.fn.func_glom: aggregate functions to one level of a hierarchy, as
  humann_regroup_table does.
• bt.fn.renorm: renormalise a function table to copies per million or
  relative abundance, as humann_renorm_table does. func_glom and renorm
  are tested against HUMAnN 3.9's own output.
• bt.fn.contributions and bt.pl.contributions: one function's abundance
  per taxon in every sample, as a table and as stacked bars.
• bt.fn.functional_redundancy: taxonomic diversity, functional diversity
  and functional redundancy per sample (Tian et al. 2020).
• bt.datasets.toy_humann: a tiny in-memory function table for examples and
  tests.
• bt.datasets.enzyme: the ENZYME EC hierarchy (CC BY 4.0), downloaded once
  and cached.
• bt.datasets.hmp2: the HMP2 inflammatory bowel disease cohort's pathways,
  species and metadata, one stool metagenome per participant, downloaded once
  and cached.
• A function guide and a tutorial on the HMP2 cohort.
• asv benchmarks for fn.func_glom, io.read_humann and
  fn.functional_redundancy, with their baselines in the docs.
• New runtime dependency: mudata>=0.4, for function tables.

Changed

• bt.pp.relative divides each value by its sample total, summed in float64:
  a sample whose total is subnormal no longer gives inf, and a float32 table
  is no longer off by about 1e-7.
• bt.pl.bar no longer gives a group the grey that marks missing values, so
  with eight or more groups plus missing values the colours change.
• The Coming-from-R page marks the phyloseq calls biotapy does not cover yet
  "not in 0.2".
• The docs build gives each notebook cell up to 5 minutes, enough for the
  function tutorial to download the HMP2 tables on a cold cache.

0.1.0

Added

• bt.io.read_biom / bt.io.write_biom: read and write BIOM tables, both the
  1.0 (JSON) and 2.1 (HDF5) dialects.
• bt.io.read_qiime2: read QIIME 2 .qza artifacts (table, taxonomy, tree,
  sample metadata), with no QIIME 2 install needed.
• bt.io.read_dada2: read a DADA2 sequence table (CSV, TSV or .rds), with
  optional taxonomy and tree.
• bt.io.read_phyloseq: read a phyloseq object saved from R (.rds/.RData),
  with no R or rpy2 install needed.
• bt.datasets.toy: a tiny in-memory dataset for examples and tests.
• bt.datasets.global_patterns / bt.datasets.enterotype: phyloseq's example
  datasets, downloaded once and cached on disk with pooch.
• bt.pp.relative: per-sample relative abundance, added as a new layer.
• bt.pp.tax_glom: aggregate features that share a lineage to a taxonomic
  rank.
• bt.pp.filter_features / bt.pp.filter_samples: keep features by
  prevalence and total reads, and samples by depth.
• bt.pp.rarefy: subsample every sample to the same depth, without
  replacement.
• bt.datasets.esophagus: phyloseq's esophagus dataset, with a tree.
• bt.tl.alpha: observed features, Shannon, Simpson, Chao1 and Faith's PD.
• bt.tl.beta / bt.tl.unifrac: Bray-Curtis and Jaccard distances, and
  unweighted and weighted UniFrac.
• bt.tl.pcoa / bt.tl.nmds: principal coordinates and non-metric
  multidimensional scaling of a stored distance matrix.
• bt.tl.permanova: PERMANOVA of a stored distance matrix.
• R golden parity tests for pp.relative and pp.tax_glom against phyloseq
  on real data.
• New runtime dependencies: rdata, xarray, pooch, scikit-bio,
  biom-format, treedata, networkx, scipy, pandas.
• New runtime dependency: scikit-learn>=1.8, for tl.nmds.
• bt.pl.bar, bt.pl.heatmap, bt.pl.richness, bt.pl.ordination and
  bt.pl.scree: phyloseq's plots, drawn from the results tl and pp
  store, returning matplotlib axes.
• A Coming-from-R page, generated from every public docstring's R
  equivalent, and a test that keeps those docstrings parseable.
• Tutorials: getting started, and phyloseq's analysis vignette redone with
  biotapy. Every notebook now runs on each docs build.
• asv benchmarks on a synthetic 5,000 x 50,000 table, with the 0.1
  baselines in the docs.
• New runtime dependency: matplotlib>=3.8, for pl, imported only when a
  plot is drawn.
• New runtime dependency: threadpoolctl>=3.5, already required by
  scikit-learn, to limit tl.permanova's OpenMP threads.

Changed

• Feature-changing functions (pp.filter_features, pp.rarefy,
  pp.tax_glom) keep X and drop the ordination summaries
  uns["biotapy"]["pcoa"] and ["nmds"] along with the other derived slots.
• CI builds the docs, executing every notebook, and fails when total line
  coverage drops below 90%.
• bt.tl.permanova runs scikit-bio's OpenMP permutation test on one thread: on
  a machine with busy cores, its default thread pool was up to thousands of
  times slower.
• Package metadata declares the license as the SPDX expression BSD-3-Clause
  (PEP 639) and adds development status, audience and topic classifiers.
• The sdist no longer ships the tests that check the repository itself (CI
  config and knowledge bundle), so its test suite passes on its own.

0.0.1

Name reservation: package skeleton, CI, documentation site and development rules. No public functions yet.