ImagineG2P/g2p-poc-v1

IMAGINE-AI genotype→phenotype proof-of-concept set, v1

500 bacterial and archaeal type-strain genomes, one per species, each joined to curated BacDive phenotypes and a nitrogen-fixation label, prepared for Evo 2, MicroGlot and BacDive-AI. Built 21 Oct 2026 from the WVU-GENESIS repository at 53a6e78 (scripts/poc_set/). DOE GENESIS, PI Ember Morrissey, West Virginia University.

What is in it

Path	One row / file per	Use
labels/labels.parquet	genome	phenotype labels, N-fixation label, folds, provenance
labels/genomes.parquet	genome	genome-side QC and measurements (never use as labels)
genomes/<accession>.fna.gz	genome	RefSeq assembly as NCBI distributes it
proteins/<accession>.faa.gz	genome	RefSeq (PGAP) proteins
evo2/windows-*.parquet	8,192-bp window	Evo 2 8K-context checkpoints (288,815 windows, 2.27 Gbp)
microglot/contigs-*.parquet	gap-free contig piece	MicroGlot (windows at run time)
pfam/<accession>.tsv.gz	genome	InterProScan-format Pfam annotation for BacDive-AI
pfam/pfam_long.parquet	genome × Pfam family	protein-family presence matrix
bacdive_ai/predictions.parquet	genome	BacDive-AI probabilities for its eight traits

How the genomes were chosen

Start: the strict minimum genome+phenotype set — BacDive strains with their own genome link, NCBI-labelled as the same species, held locally, and at least 5 of 8 core traits (Gram, shape, motility, spore, oxygen, optimum temperature, pH, salt) from curated sources.

Gate	Strains left	Species
strict minimum set	4,207	4,065
files	4,181	4,041
refseq	4,077	3,950
ani	3,986	3,867
strain	3,965	3,851
checkm	2,973	2,881
contiguity	2,780	2,697
label	2,698	2,631

• files genome and protein FASTA held; refseq current RefSeq assembly in NCBI Datasets; ani NCBI's ANI taxonomy check is OK; strain a BacDive culture-collection number or designation matches NCBI's strain name, or both mark it the type strain; checkm NCBI-computed CheckM completeness ≥ 95 % and contamination ≤ 5 %; contiguity contig N50 ≥ 50 kb and ≤ 200 contigs; label N-fixation label defined, and a positive genome carries nitrogenase.
• One strain per species (best assembly level, then strain evidence, curated link, trait depth).
• Sample. All 25 positives and 7 assay negatives that pass, then presumed negatives drawn by phylum from 4,000 seeded random draws, keeping the draw whose phenotype distributions are closest to the gated pool's: mean total-variation distance 0.011 over phylum and eight traits (median random draw 0.021). Pool vs sample for every variable is in poc_selection_report.json in the source repository.
• Folds. 5 folds blocked by genus (no genus spans two folds), 5 positives per fold. split marks fold 0 as test. Diazotrophy is phylogenetically clustered: never split at random.

Nitrogen-fixation label

nfix_basis	Label	Genomes	Meaning
phenotype_assay_direct	1	12	ARA, ¹⁵N or nitrogenase activity measured on the organism
phenotype_claim	1	10	literature states the species fixes nitrogen
phenotype_assay_relative	1	3	relative or indirect fixation assay
assay_negative	0	7	direct assay found no fixation; survived the WP3 source audit
untagged_presumed	0	468	no fixer tag in any project source (project rule: untagged = non-fixer)

Read these before using the label:

1. Presumed negatives are presumed. 18 of them carry a full nitrogenase (nfix_genotype_conflict); they may be untested fixers. Drop them for a cleaner negative class.
2. Positives were required to carry nitrogenase (nifHDK, vnfDK or anfDK, from PGAP product names). Before the gates, 74 positive strains lacked it, led by Pseudomonas (18), Bacillus (9), Stutzerimonas (4), Burkholderia (3); such claims concern other isolates of the species or are false acetylene-reduction positives. Because of this filter every positive has nif genes, so a model can separate the classes by detecting nitrogenase; that is a test of nif detection, not of realized fixation.
3. Labels are species-level evidence attached to the type-strain genome. Realized fixation is strain- and condition-specific.
4. Evidence rows removed before labelling: 481 supplement cells holding a C/N ratio misread as positives, 18 "alternative nitrogenase" cells, rows a hand audit found wrong, and negatives the WP3 audit found invalid (mutants, controls, inhibitor treatments).
5. Species tagged only by FAPROTAX, ProTraits or GOLD pathway inference, and conflicted species, are excluded.

Phenotype columns

Categorical traits are curated BacDive values mapped to classes; where curated sources disagree the value is null and <trait>_conflict is true (the raw values are in <trait>_obs). oxygen_aerobic / oxygen_anaerobic say whether the organism grows with / without oxygen and survive most vocabulary disagreements ("aerobe" vs "facultative anaerobe"). Numeric optima are medians of curated optimum entries (*_spread = max − min across sources); salt is % w/v NaCl. bai_* are binary targets shaped like BacDive-AI's eight models (thermophile: optimum ≥ 45 °C; psychrophile: < 20 °C; acidophile: optimum pH < 6). These cut-offs are ours, not BacDive-AI's training definitions. Empty is not negative: null means no curated source reported it.

Leakage you must account for

• BacDive-AI training overlap. bacdive_ai_train_<model> is true when that strain is in the model's own training set (read from the model files): acidophile 130, aerobic 325, anaerobic 346, gram_positive 252, motile2 143, psychrophile 333, spore_forming 155, thermophile 350 of 500. Score BacDive-AI only on rows where the flag is false.
• Foundation-model pretraining. Evo 2 and MicroGlot were pretrained on public genome collections that probably include many of these assemblies. That is not label leakage, but report it.
• MicroGlot species conditioning gives the model the species identity, which predicts many traits by phylogeny alone. Run MicroGlot-plain alongside and compare.

Using it

Evo 2 (Brixi et al., Nature 2026, doi:10.1038/s41586-026-10176-5). 8K-context checkpoints (evo2_7b_base, evo2_1b_base, evo2_40b_base) take evo2/ windows; 1M-context checkpoints (evo2_7b, evo2_40b) can take whole contigs from genomes/. Only the 7B models run in bf16 without FP8 and a Hopper GPU. The README recommends intermediate layers (e.g. blocks.28.mlp.l3). Pool windows to a genome (mean, or attention/MIL over windows: a nif gene cluster is a few windows in a genome of hundreds, so a plain mean dilutes it).

import pandas as pd, torch
from evo2 import Evo2
m = Evo2("evo2_7b_base")
w = pd.read_parquet("evo2/windows-000.parquet")
ids = torch.tensor(m.tokenizer.tokenize(w.sequence[0]), dtype=torch.int).unsqueeze(0).cuda()
_, emb = m(ids, return_embeddings=True, layer_names=["blocks.28.mlp.l3"])

MicroGlot (Li et al., bioRxiv 2026, doi:10.64898/2026.09.22.753215; weights CC BY 4.0, needs CUDA + flash-attn). Pass sequence with species_name (NCBI names; 7 species are not in its vocabulary, species_in_microglot false — use MicroGlot-plain or its species encoder), window at 8,190 tokens as its README shows, mean over tokens, windows, then pieces per genome.

BacDive-AI (Koblitz et al., Commun Biol 2025, doi:10.1038/s42003-025-08313-3): python predict.py all pfam/<accession>.tsv.gz after gunzip. Pfam here is Pfam-A 38.2 with gathering cutoffs (pyhmmer); the models were trained on an earlier InterProScan Pfam. The model pickles need scikit-learn 1.4.0.

BacDive-AI baseline on this set

Scored only on strains outside each model's training set, against this set's bai_* targets (our cut-offs); balanced accuracy at p ≥ 0.5.

Model	Scored	Positives	Excluded (in training)	Balanced accuracy
acidophile	148	6	130	0.5
gram-positive	241	103	252	0.981
spore-forming	71	22	155	0.99
aerobic	144	138	325	0.924
anaerobic	116	30	346	0.617
thermophile	112	6	350	0.75
psychrophile	124	4	333	0.5
motile2+	330	158	143	0.778

• anaerobic: the model never calls a non-anaerobe anaerobic but misses most facultative anaerobes, which our target counts as anaerobic; its training label appears to mean obligate anaerobe.
• acidophile, thermophile, psychrophile: too few positives outside training (single digits) for the score to mean anything.
• These are baselines for the set's shape, not a benchmark: our targets are not BacDive-AI's definitions.

Provenance and licences

Phenotypes: BacDive (Schober et al., Nucleic Acids Res 2025, doi:10.1093/nar/gkae959), CC BY 4.0. Genomes and proteins: NCBI RefSeq. Pfam: CC0. Nitrogen-fixation evidence: the IMAGINE-AI literature mining and team curation, with the audits above. Selection report, funnel and code: WVU-GENESIS scripts/poc_set/ (private repository).