Files
usher-exploring/.planning/phases/06-validation/06-02-PLAN.md
2026-02-12 04:33:17 +08:00

9.8 KiB

phase, plan, type, wave, depends_on, files_modified, autonomous, must_haves
phase plan type wave depends_on files_modified autonomous must_haves
06-validation 02 execute 1
src/usher_pipeline/scoring/sensitivity.py
true
truths artifacts key_links
Sensitivity analysis perturbs each weight by +-5% and +-10% and measures rank stability
Spearman rank correlation is computed for top-100 genes between baseline and perturbed configurations
Weight perturbation renormalizes remaining weights to maintain sum=1.0 constraint
Rank stability assessment classifies each perturbation as stable (rho>=0.85) or unstable
path provides exports
src/usher_pipeline/scoring/sensitivity.py Parameter sweep sensitivity analysis with Spearman correlation
perturb_weight
run_sensitivity_analysis
summarize_sensitivity
from to via pattern
src/usher_pipeline/scoring/sensitivity.py src/usher_pipeline/scoring/integration.py compute_composite_scores import from usher_pipeline.scoring.integration import compute_composite_scores
from to via pattern
src/usher_pipeline/scoring/sensitivity.py scipy.stats spearmanr import from scipy.stats import spearmanr
from to via pattern
src/usher_pipeline/scoring/sensitivity.py src/usher_pipeline/config/schema.py ScoringWeights import from usher_pipeline.config.schema import ScoringWeights
Implement sensitivity analysis module for parameter sweep validation of scoring weights.

Purpose: Demonstrates that top candidate rankings are robust to reasonable weight perturbations (+-5-10%), satisfying success criterion 3 (rank stability). This is the core of the "are our results defensible?" validation.

Output: sensitivity.py module with weight perturbation, Spearman correlation analysis, and stability classification.

<execution_context> @/Users/gbanyan/.claude/get-shit-done/workflows/execute-plan.md @/Users/gbanyan/.claude/get-shit-done/templates/summary.md </execution_context>

@.planning/PROJECT.md @.planning/ROADMAP.md @.planning/STATE.md @.planning/phases/06-validation/06-RESEARCH.md

@src/usher_pipeline/scoring/integration.py @src/usher_pipeline/config/schema.py @src/usher_pipeline/persistence/duckdb_store.py

Task 1: Create sensitivity analysis module with weight perturbation and rank correlation src/usher_pipeline/scoring/sensitivity.py Create `src/usher_pipeline/scoring/sensitivity.py` with:
  1. EVIDENCE_LAYERS list constant: ["gnomad", "expression", "annotation", "localization", "animal_model", "literature"]

  2. DEFAULT_DELTAS list constant: [-0.10, -0.05, 0.05, 0.10]

  3. STABILITY_THRESHOLD float constant: 0.85 (Spearman rho threshold for "stable")

  4. perturb_weight(baseline: ScoringWeights, layer: str, delta: float) -> ScoringWeights function:

    • Get baseline weights as dict via baseline.model_dump()
    • Apply perturbation: w_dict[layer] = max(0.0, min(1.0, w_dict[layer] + delta))
    • Renormalize ALL weights so they sum to 1.0: divide each by total
    • Return new ScoringWeights instance
    • Raise ValueError if layer not in EVIDENCE_LAYERS
  5. run_sensitivity_analysis(store: PipelineStore, baseline_weights: ScoringWeights, deltas: list[float] | None = None, top_n: int = 100) -> dict function:

    • Default deltas to DEFAULT_DELTAS if None
    • Compute baseline scores via compute_composite_scores(store, baseline_weights)
    • Sort by composite_score DESC, take top_n genes as baseline ranking
    • For each layer in EVIDENCE_LAYERS, for each delta in deltas:
      • Create perturbed weights via perturb_weight()
      • Compute perturbed scores via compute_composite_scores(store, perturbed_weights)
      • Sort by composite_score DESC, take top_n genes
      • Inner join baseline and perturbed on gene_symbol to get paired scores
      • If fewer than 10 overlapping genes, log warning and record rho=None
      • Otherwise compute spearmanr() on paired composite_score columns
      • Record: layer, delta, perturbed_weights (as dict), spearman_rho, spearman_pval, overlap_count (how many of top_n genes appear in both), top_n
    • Return dict with keys: baseline_weights (dict), results (list of per-perturbation dicts), top_n, total_perturbations
    • Use structlog for progress logging (log each perturbation result)

    IMPORTANT: The compute_composite_scores function re-queries the DB each time. This is by design -- different weights produce different composite_score values from the same underlying evidence layer scores.

    For the Spearman correlation, join baseline_top_n and perturbed_top_n DataFrames on gene_symbol (inner join). Use the composite_score from each as the paired values. This measures whether the relative ordering of shared top genes is preserved.

  6. summarize_sensitivity(analysis_result: dict) -> dict function:

    • From the results list, compute:
      • min_rho, max_rho, mean_rho across all perturbations (excluding None values)
      • count of stable perturbations (rho >= STABILITY_THRESHOLD)
      • count of unstable perturbations (rho < STABILITY_THRESHOLD)
      • most_sensitive_layer: layer with lowest mean rho across its perturbations
      • most_robust_layer: layer with highest mean rho across its perturbations
    • overall_stable: bool = all non-None rhos >= STABILITY_THRESHOLD
    • Return dict with: min_rho, max_rho, mean_rho, stable_count, unstable_count, total_perturbations, overall_stable, most_sensitive_layer, most_robust_layer
  7. generate_sensitivity_report(analysis_result: dict, summary: dict) -> str function:

    • Follow the formatting pattern from generate_validation_report() in validation.py
    • Show table: Layer | Delta | Spearman rho | p-value | Stable?
    • Show summary: overall stability verdict, most/least sensitive layers
    • Include interpretation text

Use structlog, polars, scipy.stats.spearmanr imports. Import compute_composite_scores from usher_pipeline.scoring.integration, ScoringWeights from usher_pipeline.config.schema, PipelineStore from usher_pipeline.persistence.duckdb_store. Run: `cd /Users/gbanyan/Project/usher-exploring && python -c " from usher_pipeline.scoring.sensitivity import perturb_weight, EVIDENCE_LAYERS, STABILITY_THRESHOLD, DEFAULT_DELTAS from usher_pipeline.config.schema import ScoringWeights

Test weight perturbation

w = ScoringWeights() p = perturb_weight(w, 'gnomad', 0.10) p.validate_sum() # Must not raise print(f'Original gnomad: {w.gnomad}, Perturbed: {p.gnomad:.4f}') assert p.gnomad > w.gnomad, 'Perturbed weight should be higher'

Test renormalization

total = p.gnomad + p.expression + p.annotation + p.localization + p.animal_model + p.literature assert abs(total - 1.0) < 1e-6, f'Weights must sum to 1.0, got {total}'

Test edge: perturb to near-zero

p_low = perturb_weight(w, 'gnomad', -0.25) p_low.validate_sum() assert p_low.gnomad >= 0.0, 'Weight must not go negative'

print('All perturb_weight tests passed') "` exits 0 sensitivity.py exists with perturb_weight (renormalizing), run_sensitivity_analysis (computing Spearman rho for top-N genes across all layer/delta combinations), summarize_sensitivity (stability classification), and generate_sensitivity_report (formatted output). Weights always renormalize to sum=1.0 after perturbation.

Task 2: Export sensitivity module from scoring package src/usher_pipeline/scoring/__init__.py Update `src/usher_pipeline/scoring/__init__.py` to add imports and exports for the sensitivity module:

Add imports:

from usher_pipeline.scoring.sensitivity import (
    perturb_weight,
    run_sensitivity_analysis,
    summarize_sensitivity,
    generate_sensitivity_report,
    EVIDENCE_LAYERS,
    STABILITY_THRESHOLD,
)

Add to all list: "perturb_weight", "run_sensitivity_analysis", "summarize_sensitivity", "generate_sensitivity_report", "EVIDENCE_LAYERS", "STABILITY_THRESHOLD"

NOTE: Plan 01 may have already updated init.py to add negative_controls exports. If so, ADD the sensitivity imports alongside those -- do not remove them. Read the file first to check current state. Run: cd /Users/gbanyan/Project/usher-exploring && python -c "from usher_pipeline.scoring import perturb_weight, run_sensitivity_analysis, summarize_sensitivity, generate_sensitivity_report, EVIDENCE_LAYERS, STABILITY_THRESHOLD; print(f'EVIDENCE_LAYERS: {EVIDENCE_LAYERS}'); print(f'STABILITY_THRESHOLD: {STABILITY_THRESHOLD}'); print('All sensitivity exports OK')" exits 0 All sensitivity analysis functions and constants are importable from usher_pipeline.scoring. Existing exports from negative_controls (Plan 01) are preserved.

- `python -c "from usher_pipeline.scoring.sensitivity import perturb_weight; from usher_pipeline.config.schema import ScoringWeights; w = ScoringWeights(); p = perturb_weight(w, 'gnomad', 0.05); p.validate_sum(); print('OK')"` -- weight perturbation works and renormalizes - `python -c "from usher_pipeline.scoring import run_sensitivity_analysis, summarize_sensitivity, generate_sensitivity_report"` -- all exports available - `python -c "from usher_pipeline.scoring.sensitivity import STABILITY_THRESHOLD; assert STABILITY_THRESHOLD == 0.85"` -- threshold configured

<success_criteria>

  • perturb_weight correctly perturbs one layer and renormalizes to sum=1.0
  • run_sensitivity_analysis computes Spearman rho for all layer x delta combinations
  • summarize_sensitivity classifies perturbations as stable/unstable
  • generate_sensitivity_report produces human-readable output
  • All functions exported from scoring package </success_criteria>
After completion, create `.planning/phases/06-validation/06-02-SUMMARY.md`