bio-stats-ml-reporting

Bio Stats ML Reporting

Aggregate results, train ML models, and produce reports with validated references.

Instructions

Tool guides and versions: docs/README.md.

1. Join outputs in DuckDB v1.1+ and build feature tables. Arrow / DuckLake integration is the recommended bridge into ML pipelines for large datasets.
2. Train baseline models and evaluate with cross-validation.
   • CPU baseline: scikit-learn v1.5+ for linear/tree/clustering baselines; XGBoost v2.1.4+ for gradient boosting.
   • GPU node available (CUDA): set device="cuda" on XGBoost (native since v2.0) by default. For sklearn-compatible estimators (random forest, k-means, PCA, UMAP), use RAPIDS cuML as a drop-in replacement and record the device in the run log.
3. Generate reports and validate references.
   • Validate the prediction table with scripts/validate_predictions.py. Keep group identifiers and confounder labels in the table so the gate can detect sample/group leakage, class imbalance, calibration failure, batch-outcome imbalance, and performance that does not beat the majority-class null.
4. For exploratory omics projects, aggregate discovery evidence across the literature-derived analysis playbook, annotation, phylogenomics, viromics, and comparative-genomics outputs.
5. Comparative-axes rollup — join the per-axis comparison artifacts produced by upstream skills into a single comparative_axes_summary.tsv. The rollup must have one row per (query genome, axis) and include:
   • genome-property frontier (size, gene count, etc. — link to relative_genome_metrics.tsv and genome_size_frontier.tsv)
   • marker-gene census (link to marker_census.tsv)
   • family copy-number expansions/contractions (link to family_copy_number_comparison.tsv and family_expansion_candidates.tsv)
   • synteny / conserved neighborhoods (link to conserved_neighborhoods.tsv)
   • non-coding RNA census (link to ncRNA_census.tsv) Each row records observation, comparison baseline, literature reference, status (notable / conserved / artifact / negative), and a follow-up test.
6. Produce an interesting-findings section that ranks candidate discoveries relative to the literature-derived baseline and separates:
   • strong candidates with multiple evidence types
   • plausible candidates needing validation
   • likely artifacts or conserved lineage features
   • explicit negative findings where nothing notable was detected
7. Include the comparison baseline, literature context, confidence, and next discriminating analyses for each candidate.