Skip to content

GenoLeWM-FX Borzoi baseline and saturation report

Generated by tools.research.fx_borzoi_baselines at 2026-06-11T13:00:00Z.

Parent epic: #266. Baseline gate: #270. Model gate: #271.

Decision: go_residual_model.

This report is a baseline and saturation gate only. It is not a model-quality result, clinical result, deployment-readiness result, broad VEP superiority claim, broad Carbon comparison, useful-planning claim, ground-truth biology claim, or exact fipip overlap claim.

Reproduce

uv run python -m tools.research.fx_borzoi_baselines \
  --source-manifest configs/fx/borzoi_rescue_sources.json \
  --cache-manifest docs/research/fx-borzoi-cache-manifest.json \
  --output-json docs/research/fx-borzoi-baseline-report.json \
  --output-md docs/research/fx-borzoi-baseline-report.md

Machine-readable report: fx-borzoi-baseline-report.json.

Baselines

Baseline Family AUROC AUPRC Balanced accuracy
constant_negative constant 0.5000 [0.5000, 0.5000] 0.0971 [0.0879, 0.1173] 0.5000 [0.5000, 0.5000]
label_prior_constant label_prior 0.5000 [0.5000, 0.5000] 0.0971 [0.0879, 0.1173] 0.5000 [0.5000, 0.5000]
maf_source_only source_only 0.5057 [0.4731, 0.5382] 0.1024 [0.0895, 0.1200] 0.4987 [0.4851, 0.5128]
ld_score_source_only source_only 0.5077 [0.4761, 0.5417] 0.1026 [0.0894, 0.1196] 0.5013 [0.4830, 0.5201]
tss_dist_source_only source_only 0.5089 [0.4776, 0.5401] 0.1091 [0.0919, 0.1264] 0.5188 [0.4999, 0.5391]
direct_traitgym_native_borzoi direct_borzoi 0.6446 [0.6082, 0.6787] 0.2214 [0.1794, 0.2641] 0.5790 [0.5540, 0.6005]
source_logistic_probe source_probe 0.5094 [0.4736, 0.5415] 0.1046 [0.0912, 0.1231] 0.5000 [0.5000, 0.5000]
borzoi_plus_source_logistic_probe borzoi_probe 0.6494 [0.6144, 0.6821] 0.2225 [0.1800, 0.2674] 0.5080 [0.5014, 0.5153]

Gate Decision

Strongest simple baseline: borzoi_plus_source_logistic_probe with AUPRC 0.2225 and AUROC 0.6494.

Saturation thresholds: AUPRC 0.8, AUROC 0.9.

Leakage Audit

Check Value
Duplicate variant keys 0
Train/holdout key overlap 0
Train chromosomes 1, 2, 4, 5, 6, 7, 8, 9, 10, 12, 13, 15, 18, 19, 21, 22
Holdout chromosomes 3, 11, 14, 16, 17, 20
fipip exact join status not_run_full_table_not_staged
Carbon baseline status not_applicable_no_carbon_scores_in_cache

Blockers

No #270 blockers remain; the residual-model issue may proceed.

Issue Action Reason
#270 close-completed The baseline gate is not saturated and direct Borzoi has holdout signal.
#271 open-next A residual model may proceed against the strongest simple baseline.