BioModelBench — Bacterial multi-phenotype profile (bacdive_multitrait)

Task

For each row in test.parquet, predict a joint phenotypic profile for that strain across 18 targets covering growth conditions, cell properties, and metabolic pathways. Test strains are drawn from bacterial families disjoint from the training set. You get the NCBI assembly accession + phylum for each strain — no genus, species, family, strain name, or organism name.

What's shipped

train.parquet

Columns:

• assembly_accession (string, e.g. GCF_000005845.2)
• phylum (string, coarse taxonomy)
• 18 label columns (below). Some cells will be NaN — not every training species has every phenotype characterized. Handle missing-target training gracefully.

test.parquet

Columns: assembly_accession, phylum. No labels.

answer.parquet schema — 19 columns

column	type	notes
assembly_accession	string	one row per test strain
gram_positive_prob	float ∈ [0,1]	probability strain is Gram-positive
oxygen_tolerance	string label	one of: aerobic, anaerobic, facultative, microaerophilic
optimum_temperature_c	float	°C, growth optimum
optimum_ph	float	pH units
temperature_range	string label	one of: psychrophilic, mesophilic, thermophilic, extreme_thermophilic
sporulation_prob	float ∈ [0,1]	probability strain forms endospores
motility_prob	float ∈ [0,1]	probability strain is motile
cell_shape	string label	one of: bacillus, coccus, spiral, coccobacillus, vibrio, filament, other
pathway_nitrogen_fixation_prob	float ∈ [0,1]	encodes N₂ fixation machinery
pathway_nitrate_reduction_prob	float ∈ [0,1]	reduces NO₃⁻ (assimilatory or dissimilatory)
pathway_denitrification_prob	float ∈ [0,1]	full denitrification pathway
pathway_sulfate_reduction_prob	float ∈ [0,1]	dissimilatory sulfate reduction
pathway_fermentation_prob	float ∈ [0,1]	fermentative metabolism
pathway_cellulose_degradation_prob	float ∈ [0,1]	cellulolytic
pathway_xylan_degradation_prob	float ∈ [0,1]	xylanolytic
pathway_chitin_degradation_prob	float ∈ [0,1]	chitinolytic
pathway_fumarate_reduction_prob	float ∈ [0,1]	fumarate as terminal electron acceptor
pathway_iron_reduction_prob	float ∈ [0,1]	dissimilatory Fe(III) reduction

Every column must be present. Use NaN if you decline to predict a value — the grader will count that as coverage loss.

Compute

Modal container: A10G GPU (24 GB VRAM), 8 CPUs, 96 GB RAM. Python 3.11 with torch / transformers / peft / xgboost / lightgbm / scikit-learn / pandas / pyarrow / numpy / pyBigWig / pysam / biopython / requests pre-installed. pip install <x> works.

Budget: ~10 hours of GPU compute. Use it. Even after you produce a valid answer.parquet, keep iterating until the budget is exhausted. The runner will relaunch your session with a continuation prompt if you exit early — don't finalize prematurely.

Deliverables

• answer.parquet — 19 columns as above, one row per test strain.
• method.md (< 500 words).
• training_manifest.json — every URL / API / dataset touched.

What the grader does

Joins your answer.parquet against the hidden per-target labels on assembly_accession. For each target independently, computes the per-target metric on the subset of test strains where the ground-truth label is not NaN. Reports:

• Per-target: primary metric + coverage + baseline gap
• Per-tier: normalized composite (Tier 1 growth, Tier 2 morphology, Tier 3 metabolism)
• Overall: weighted composite

Metrics by target type:

• Binary probability targets → AUPRC + AUROC + Brier
• Multiclass label targets → macro-F1 + balanced accuracy
• Continuous float targets → Spearman ρ + MAE
• Pathway probability targets → AUPRC + AUROC

Reference bars (family-holdout, from literature)

Tier 1 (growth & culturability, weight 0.40):

Target	Metric	Ref baseline
oxygen_tolerance	macro-F1	0.70
optimum_temperature_c	Spearman ρ	0.70
optimum_ph	Spearman ρ	0.35
temperature_range	macro-F1	0.60

Tier 2 (morphology & dispersal, weight 0.30):

Target	Metric	Ref baseline
gram_positive_prob	AUROC	0.95
sporulation_prob	AUROC	0.90
motility_prob	AUROC	0.85
cell_shape	macro-F1	0.55

Tier 3 (metabolic pathways, weight 0.30, averaged macro-AUPRC ≈ 0.72):

Pathway	AUPRC ref
nitrogen_fixation	0.90
sulfate_reduction	0.80
nitrate_reduction	0.75
denitrification	0.75
cellulose_degradation	0.75
xylan_degradation	0.70
chitin_degradation	0.70
fermentation	0.65
iron_reduction	0.60
fumarate_reduction	0.55

Base rates on the test partition and per-target coverage are reported in hidden/build_stats.json post-build.