bacdive_genomes

Bacterial genome and protein sequence sets drawn from BacDive, paired with manifest tables carrying BacDive IDs, INSDC accessions and phenotype labels.

Repository layout

16S_logistic_regression/   16S k-NN baseline + model-comparison pipeline (code, 16S
                           tree, per-phenotype selection tables); has its own README.md

genomefiles/     nucleotide genomes (.fna) + their manifest tables
  decanoate.zip
  decanoate_manifest.csv
  oxygenthreshold_genomes.zip
  oxygen_tolerance_manifest500.csv
  pooled_allpheno.zip
  remaining_pooled.zip

proteinfiles/    predicted protein sequences (.faa), one archive per phenotype
  acetate.zip  butyrate.zip  cellulose.zip  decanoate.zip  ethanol.zip
  formate.zip  hexanoate.zip  propionate.zip  xylan.zip
  remaining_pooled_faa.zip

comparison_out/  model-comparison result tables
  all_phenotypes_model_comparison_ver3_vsknn_summary.csv   all 9 phenotypes x 23 models
  acetate/  butyrate/  cellulose/  decanoate/  ethanol/    per-phenotype tables
  formate/  hexanoate/  propionate/  xylan/

Note on the two decanoate.zip files. genomefiles/decanoate.zip holds nucleotide genomes (.fna); proteinfiles/decanoate.zip holds protein sequences (.faa) for the same 1,237 genomes. They are different files, not duplicates.

genomefiles/

Archive	Contents	Files	Manifest
decanoate.zip	decanoate/<accession>.fna	1,237	decanoate_manifest.csv
oxygenthreshold_genomes.zip	genomes/<accession>.fna	500	oxygen_tolerance_manifest500.csv
pooled_allpheno.zip	pooled_allpheno/<accession>.fna	1,437	(none — see below)
remaining_pooled.zip	remaining_pooled/<accession>.fna	1,572	(none — see below)

pooled_allpheno.zip

A single de-duplicated pool of the genomes behind the substrate-utilisation phenotypes, for use as one combined training/evaluation set instead of eight overlapping ones. 1,437 unique genomes, 2.0 GB compressed / 7.2 GB uncompressed, laid out flat as pooled_allpheno/<accession>.fna.

The pool is exactly the union of the eight non-decanoate proteinfiles/ sets, so every accession in it carries at least one phenotype label:

Phenotype set	Genomes in the pool	Phenotype set	Genomes in the pool
acetate	684	hexanoate	69
butyrate	69	propionate	550
cellulose	478	xylan	100
ethanol	168	(decanoate)	291
formate	335

Genomes may appear under more than one phenotype — 822 carry exactly one of the eight labels, 355 carry two, and 260 carry three or more (up to all eight).

decanoate is only partly represented. The pool is built from the other eight phenotypes; the 291 decanoate genomes above are there because they also carry one of those labels. The 946 genomes that are decanoate-only are not in pooled_allpheno.zip. 944 of those 946 are in remaining_pooled.zip (below). For the full decanoate set (1,237 genomes) use decanoate.zip and its manifest.

This archive ships without a manifest. To attach labels, join it to the phenotype archives in proteinfiles/ on the accession, or to decanoate_manifest.csv on dest_file for the decanoate subset.

remaining_pooled.zip

A second pool of 1,572 genomes (2.5 GB compressed / 8.5 GB uncompressed), laid out flat as remaining_pooled/<accession>.fna. It shares no accessions with pooled_allpheno.zip, so the two pools together cover 3,009 distinct genomes. 944 of its genomes are also in decanoate.zip. Combined with the 291 decanoate genomes in pooled_allpheno.zip, that covers all but 2 of the 1,237 decanoate genomes. Like pooled_allpheno.zip, it ships without a manifest. Its protein sequences are in proteinfiles/remaining_pooled_faa.zip.

How the CSVs link to the archives

Each manifest row describes exactly one genome file inside its paired archive. The column that carries the filename differs between the two manifests (pooled_allpheno.zip has no manifest and is not covered here):

• decanoate_manifest.csv — the dest_file column holds the file name as it appears in the archive. Row dest_file = GCA_004310665.1.fna is the archive member decanoate/GCA_004310665.1.fna. All 1,237 rows match one-to-one.

• oxygen_tolerance_manifest500.csv — the genome_id column holds the accession; append .fna to get the archive member. Row genome_id = GCA_002973605.1 is the archive member genomes/GCA_002973605.1.fna. All 500 rows match one-to-one. This file also has a dest_file column, but it is blank for 108 of the 500 rows — use genome_id as the join key here, not dest_file.

Caveat: genome_source is not a filename column. It records where the assembly was retrieved from (e.g. ncbi_dataset/data) and is the same value for nearly every row.

Manifest columns

Shared by both manifests:

Column	Meaning
ID	BacDive strain ID
insdc_genome_accession	INSDC assembly accession, unversioned (GCA_003151135)
assembly_dir	Versioned accession (GCA_003151135.1)
accession_category	How the accession was resolved (e.g. single_insdc_ncbi)
labels_16s	16S accession joined to the BacDive ID (KY612316__159837)
source_file	Original NCBI file name as downloaded
dest_file	Renamed file as stored in the archive
n_versions_on_disk	Number of assembly versions available
genome_source	Provenance of the download (e.g. ncbi_dataset/data)
source_note	Free-text note; usually empty

Manifest-specific:

• decanoate_manifest.csv — decanoate: the phenotype call for decanoate utilisation.
• oxygen_tolerance_manifest500.csv — genome_id (join key), oxygen_tolerance (BacDive term, e.g. obligate aerobe), oxygen_class (collapsed label, e.g. aerobe), fna_path (absolute path on the machine the set was built on), label_source.

proteinfiles/

Predicted proteomes for nine carbon-source / substrate-utilisation phenotypes. Every phenotype archive follows the layout <phenotype>/faa/<accession>.faa:

Archive	Genomes	Archive	Genomes
acetate.zip	684	formate.zip	335
butyrate.zip	69	hexanoate.zip	69
cellulose.zip	478	propionate.zip	550
decanoate.zip	1,237	xylan.zip	100
ethanol.zip	168

The accession in each .faa file name is the same versioned assembly accession used in the manifests, so protein files join back to a manifest on dest_file with the extension swapped: manifest dest_file = GCA_004310665.1.fna corresponds to the archive member decanoate/faa/GCA_004310665.1.faa. Only decanoate currently has a manifest in this repository; the other eight phenotypes ship as archives alone.

remaining_pooled_faa.zip holds the proteomes for the 1,572 genomes in genomefiles/remaining_pooled.zip, one per genome with matching accessions. Its layout is different: files sit directly under faa/<accession>.faa, with no phenotype folder (3.5 GB uncompressed).

16S_logistic_regression/

Code and inputs for the 16S-vs-genome benchmark that produced comparison_out/: the driver run_16s_comparison_decanoate_ver3.sh, the stage modules (clade-threshold sweep, CV folds, 16S k-NN baseline, model comparison, TNF features, summary table), the GG2-derived 16S tree bacdive_16s_gg2.nwk, bacdive_sequence_info_16S.csv, and one <pheno>_selection.csv for each of the nine phenotypes. See 16S_logistic_regression/README.md for how to run it.

comparison_out/

Ver3 comparison of 23 genome-derived models against a 16S k-NN baseline (knn16s_nested), for all nine phenotypes, under seven cross-validation split schemes, with 10 repeats each. The tables come from two runs of the same driver (SLURM job IDs are in parentheses):

Run	What it computed	Driver
2026-10-03 (47386403–47386416, plus 47383648 for hexanoate)	stage 6 for the 18 embedding arms; the Pfam and TNF arms run alongside them	run_16s_comparison_decanoate_ver3.sh
2026-10-12 (47438583–47438594)	stage 6 for the lr_mmseq arm, then the combined per-phenotype summaries	run_16s_comparison_decanoate_ver3_gpu.sh

Start here: all_phenotypes_model_comparison_ver3_vsknn_summary.csv

This is every model on every phenotype in one table: 9 phenotypes × 23 arms × 7 schemes × 2 metrics = 2,898 rows. It is the nine per-phenotype <pheno>/<pheno>_model_comparison_ver3_vsknn_summary.csv files stacked, plus a phenotype column and two columns that line the clade scheme up across phenotypes:

• scheme: the scheme name as stored. The clade scheme carries a phenotype-specific threshold, so its name changes by phenotype: acetate group_clade105, butyrate group_clade164, cellulose group_clade75, decanoate group_clade70, ethanol group_clade106, formate group_clade86, hexanoate group_clade85, propionate group_clade83, xylan group_clade99.
• scheme_kind: the same value with the clade threshold removed (group_clade), so stratified, group_clade and group_gtdb_{phylum,class,order,family,genus} line up across all nine phenotypes.
• clade_threshold: that threshold (blank for the other schemes).

Each row gives the arm's mean and 95% CI over 10 repeats, the same for the baseline (knn_*), the difference (delta_mean, delta_sd), the per-repeat BCa bootstrap results (boot_*), and source, the per-arm table the row came from.

Models (23 arms):

• 18 embedding arms: logistic regression on mean-pooled bacformer (l16, l30), ESM-C (l36), gLM2 (l17, l33), ProkBERT (max, mean) and Evo2 (blocks26_max, blocks31_max) embeddings, each raw and L2-normalised (_l2).
• lr_mmseq: MMseqs2 protein-cluster features.
• lr_pfam, rf_pfam, svm_pfam: Pfam domain features with logistic regression, random forest and SVM.
• lr_tnf: tetranucleotide frequency.

Metrics: roc_auc, mcc.

Per-phenotype folders

Each comparison_out/<pheno>/ holds the combined summary for that phenotype plus nine tables for each (scheme × source) pair: 7 schemes × 19 sources (18 embeddings and arm_lr_mmseq) = 133 pairs, or 1,197 tables. The file names are <pheno>_<kind>_<scheme>_ver3_vsknn_<source>[_<suffix>].csv:

File	Contents
model_comparison_…_<source>.csv	per-repeat metrics for every model in that run, including the baseline
…_summary.csv, …_metrics_summary.csv, …_practical_summary.csv	condensed views of that table
…_tests.csv, …_tests_summary.csv	paired arm-vs-baseline tests (effect, BCa CI, p-values with Holm / BH correction)
…_fold_effects.csv	per-fold, per-repeat differences between arm and baseline
pred_lr_embed_…csv	per-genome out-of-fold predictions (strain_id, fold, y_true, y_score)
disagreements_…csv	genomes on which the arm and the baseline disagree

The Pfam and TNF arms have no tables of their own. They ran inside the embedding runs: lr_pfam in bacformer_L30, svm_pfam in blocks26_max, lr_tnf in esmc_L36 and rf_pfam in glm2_L33. The source column in the summaries points to the right table.

Five folders (butyrate, ethanol, formate, hexanoate, xylan) also hold 14 lr_mmseq files from 2026-10-03, two per scheme: pred_lr_mmseq_<scheme>.csv (per-genome predictions) and mmseq_sensitivity_<scheme>.csv (per-repeat metrics). The matching files for the other four phenotypes were written on 2026-10-04 and are not uploaded.

Caveats

• lr_embed_bacformer_l30 appears twice. The 2026-10-12 lr_mmseq run re-ran it as the arm carrying lr_mmseq, so it shows up again in the arm_lr_mmseq tables. The re-run is not bit-identical: its 10-repeat means differ from the 2026-10-03 run by up to 0.0032 ROC-AUC and 0.0038 MCC. The summaries use the 2026-10-03 values (source = …_bacformer_L30).
• The acetate 2026-10-03 job is marked FAILED in SLURM, because the driver flagged group_clade105 × bacformer_L16. That run's own log ends with exit 0, and all nine of its tables are present and complete, so the flag appears spurious (with about 126 concurrent runs, wait most likely lost the exit status). Treat that cell with the same care as the others, no more.
• When group_gtdb_class and group_gtdb_phylum folds are the same in every repeat (folds_fixed_across_repeats = True), the CIs for those schemes reflect only model noise.
• The combined table was checked against the raw per-repeat tables. It has all 9 phenotypes × 23 arms × 7 schemes × 2 metrics with no gaps or duplicates, and every arm_mean and knn_mean equals the 10-repeat mean of its source table.

Usage

import zipfile, pandas as pd
from huggingface_hub import hf_hub_download

csv = hf_hub_download("Jigyasa3/bacdive_genomes",
                      "genomefiles/decanoate_manifest.csv",
                      repo_type="dataset")
zp  = hf_hub_download("Jigyasa3/bacdive_genomes",
                      "genomefiles/decanoate.zip",
                      repo_type="dataset")

manifest = pd.read_csv(csv)
with zipfile.ZipFile(zp) as z:
    row = manifest.iloc[0]
    seq = z.read(f"decanoate/{row.dest_file}").decode()