Previously checkpoints only logged a message but never wrote to DuckDB,
causing all progress to be lost on process kill. Now every batch_size
genes (default 500) are persisted to a literature_partial table.
On restart the CLI loads partial results and resumes from where it
left off. The partial table is cleaned up after successful completion.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add sensitivity module imports and exports
- Preserve existing negative_controls exports from Plan 01
- All sensitivity functions and constants now importable from usher_pipeline.scoring
- Add perturb_weight() function with renormalization to maintain sum=1.0
- Add run_sensitivity_analysis() for parameter sweep across all layers and deltas
- Add summarize_sensitivity() for stability classification
- Add generate_sensitivity_report() for human-readable output
- Default perturbations: ±5% and ±10% with stability threshold 0.85
- Create ReproducibilityReport dataclass with all metadata fields
- Implement generate_reproducibility_report function
- Extract parameters from PipelineConfig (scoring weights, data versions)
- Capture software environment (Python, polars, duckdb versions)
- Build filtering steps from ProvenanceTracker
- Compute tier statistics from tiered DataFrame
- Support optional validation metrics
- to_json: write as indented JSON for machine-readable format
- to_markdown: write with tables and headers for human-readable format
- 7 tests covering all report fields, formats, and edge cases
- Add matplotlib>=3.8.0 and seaborn>=0.13.0 to dependencies
- Create visualizations.py with 3 plot functions and orchestrator
- plot_score_distribution: histogram colored by confidence tier
- plot_layer_contributions: bar chart of evidence layer coverage
- plot_tier_breakdown: pie chart of tier distribution
- Use Agg backend for headless/CLI safety
- All plots saved at 300 DPI with proper figure cleanup
- 6 tests covering file creation, edge cases, and return values
- Add confidence tier classification (HIGH/MEDIUM/LOW) based on composite_score and evidence_count
- Add supporting_layers and evidence_gaps columns per gene
- Use vectorized polars expressions for performance
- Configurable thresholds for tier assignment
- Created score_cmd.py following evidence_cmd.py pattern
- Orchestrates full scoring pipeline: known genes -> composite scores -> QC -> validation
- Options: --force, --skip-qc, --skip-validation for flexible iteration
- Registered score command in main CLI group
- Displays comprehensive summary with quality flag distribution
- Create validation.py with known gene ranking validation
- validate_known_gene_ranking: PERCENT_RANK window function over all genes
- Computes median percentile, top quartile count/fraction for known genes
- generate_validation_report: human-readable text output with formatted table
- Update __init__.py to export run_qc_checks, validate_known_gene_ranking, generate_validation_report
- Create join_evidence_layers() with LEFT JOIN preserving NULLs from all 6 evidence tables
- Implement compute_composite_scores() with NULL-preserving weighted average (weighted_sum / available_weight)
- Add quality_flag classification based on evidence_count (sufficient/moderate/sparse/no_evidence)
- Include per-layer contribution columns for explainability
- Add persist_scored_genes() to save scored_genes table to DuckDB
- Log summary stats: coverage, mean/median scores, quality distribution, NULL rates
- Create protein features data model with domain, coiled-coil, TM, cilia motifs
- Implement fetch.py with UniProt REST API and InterPro API queries
- Implement transform.py with feature extraction, motif detection, normalization
- Implement load.py with DuckDB persistence and provenance tracking
- Add CLI protein command following evidence layer pattern
- Add comprehensive unit and integration tests (all passing)
- Handle NULL preservation and List(Null) edge case
- Add get_steps() method to ProvenanceTracker for test compatibility
- Add localization subcommand to evidence command group
- Implement checkpoint-restart pattern for HPA download
- Display summary with evidence type distribution
- Create 17 unit and integration tests (all pass)
- Test HPA parsing, evidence classification, scoring, and DuckDB persistence
- Fix evidence type terminology (computational vs predicted) for consistency
- Mock HTTP calls in integration tests for reproducibility
- Define LocalizationRecord model with HPA and proteomics fields
- Implement fetch_hpa_subcellular to download HPA bulk data
- Implement fetch_cilia_proteomics with curated reference gene sets
- Implement classify_evidence_type (experimental vs computational)
- Implement score_localization with cilia proximity scoring
- Implement process_localization_evidence end-to-end pipeline
- Create load_to_duckdb for persistence with provenance
- PipelineStore class for DuckDB-based storage
- save_dataframe/load_dataframe for polars and pandas
- Checkpoint system with has_checkpoint and metadata tracking
- Parquet export capability
- Context manager support
