Week 1: Tier-A sickle cell signature via 2-study concordance

Implement and run the Week 1 disease-signature pipeline:
- src/disease.py: Welch t-test + BH DE (microarray), probe->symbol
  collapse, cross-study concordance filter, 2-study provenance schema
- scripts/week1_explore.py: download GSE35007 + GSE16728, DE + concordance
- scripts/week1_finalize.py: mygene ID mapping + persist signature
- tests/test_disease.py: synthetic-data tests for DE/collapse/concordance
- docs/data_sources.md: chosen datasets, group defs, reproduction steps

Result: sickle_cell_signature_v1.json (gitignored), Tier A, 250 up /
227 down genes from 671 concordant (GSE35007 Illumina whole blood SS/AA +
GSE16728 Affymetrix whole blood patient/control). Documented caveats:
missing HbF axis (globin depletion) and reticulocyte composition confound.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

tests/test_disease.py

@@ -0,0 +1,122 @@
+"""Tests for the Week 1 signature-construction logic on synthetic data.
+
+These verify the DE / probe-collapse / concordance math without touching the network, so the
+pipeline is trustworthy before it is pointed at real GEO studies.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pandas as pd
+import pytest
+
+from src.disease import (
+    DISEASE_LABEL,
+    HEALTHY_LABEL,
+    ConcordanceSummary,
+    SignatureProvenance,
+    StudyProvenance,
+    build_signature,
+    collapse_probes_to_symbols,
+    compute_differential_expression,
+    concordance_filter,
+)
+from src.provenance import ConfidenceTier
+
+
+def _synthetic_study(seed: int, n_per_group: int = 12) -> tuple[pd.DataFrame, pd.Series]:
+    """Genes x samples matrix where UP is higher and DOWN is lower in disease."""
+    rng = np.random.default_rng(seed)
+    genes = ["UP1", "UP2", "DOWN1", "DOWN2", "NOISE1", "NOISE2"]
+    samples = [f"d{i}" for i in range(n_per_group)] + [f"h{i}" for i in range(n_per_group)]
+    groups = pd.Series([DISEASE_LABEL] * n_per_group + [HEALTHY_LABEL] * n_per_group, index=samples)
+
+    base = rng.normal(8.0, 0.3, size=(len(genes), len(samples)))
+    df = pd.DataFrame(base, index=genes, columns=samples)
+    disease = groups == DISEASE_LABEL
+    df.loc["UP1", disease] += 3.0
+    df.loc["UP2", disease] += 2.0
+    df.loc["DOWN1", disease] -= 3.0
+    df.loc["DOWN2", disease] -= 2.0
+    return df, groups
+
+
+def test_welch_de_recovers_direction_and_significance():
+    expr, groups = _synthetic_study(seed=1)
+    de = compute_differential_expression(expr, groups, method="welch")
+
+    assert de.loc["UP1", "log_fc"] > 0 and de.loc["UP1", "qvalue"] < 0.05
+    assert de.loc["DOWN1", "log_fc"] < 0 and de.loc["DOWN1", "qvalue"] < 0.05
+    # Pure-noise genes should not be significant.
+    assert de.loc["NOISE1", "qvalue"] > 0.05
+
+
+def test_compute_de_rejects_unlabelled_samples():
+    expr, groups = _synthetic_study(seed=2)
+    with pytest.raises(ValueError):
+        compute_differential_expression(expr, groups.iloc[:-3], method="welch")
+
+
+def test_collapse_probes_keeps_highest_mean_expression():
+    de = pd.DataFrame(
+        {"log_fc": [1.0, 2.0, -1.0], "pvalue": [0.1, 0.2, 0.3], "qvalue": [0.1, 0.2, 0.3]},
+        index=["probeA1", "probeA2", "probeB1"],
+    )
+    probe_to_symbol = pd.Series({"probeA1": "GENEA", "probeA2": "GENEA", "probeB1": "GENEB"})
+    expr = pd.DataFrame(
+        {"s1": [1.0, 100.0, 5.0], "s2": [1.0, 100.0, 5.0]},
+        index=["probeA1", "probeA2", "probeB1"],
+    )
+    collapsed = collapse_probes_to_symbols(de, probe_to_symbol, expression_for_ranking=expr)
+
+    assert set(collapsed.index) == {"GENEA", "GENEB"}
+    # probeA2 has the higher mean expression, so its log_fc (2.0) should win.
+    assert collapsed.loc["GENEA", "log_fc"] == 2.0
+
+
+def test_concordance_filter_keeps_only_agreeing_genes():
+    de_a = pd.DataFrame(
+        {"log_fc": [2.0, -2.0, 1.5, 0.1], "qvalue": [0.001, 0.001, 0.2, 0.001]},
+        index=["UP1", "DOWN1", "WEAK", "DISAGREE"],
+    )
+    de_b = pd.DataFrame(
+        {"log_fc": [1.8, -2.2, 1.4, -0.1], "qvalue": [0.002, 0.002, 0.2, 0.002]},
+        index=["UP1", "DOWN1", "WEAK", "DISAGREE"],
+    )
+    keep, summary = concordance_filter(de_a, de_b)
+
+    assert set(keep.index) == {"UP1", "DOWN1"}  # WEAK fails q-cut; DISAGREE flips sign
+    assert keep.loc["UP1", "log_fc"] == pytest.approx(1.9)  # mean of the two
+    assert keep.loc["UP1", "qvalue"] == pytest.approx(0.002)  # max of the two
+    assert isinstance(summary, ConcordanceSummary)
+    assert summary.n_genes_tested == 4 and summary.n_concordant == 2
+    assert summary.n_up == 1 and summary.n_down == 1
+
+
+def test_build_signature_splits_directions_and_respects_top_n():
+    concordant = pd.DataFrame(
+        {
+            "log_fc": [3.0, 2.0, 1.0, -1.0, -2.0],
+            "qvalue": [0.001, 0.002, 0.003, 0.004, 0.005],
+        },
+        index=["UP1", "UP2", "UP3", "DOWN1", "DOWN2"],
+    )
+    prov = SignatureProvenance(
+        studies=[
+            StudyProvenance(geo_accession="GSE1", n_disease=12, n_healthy=12,
+                            platform="P", tissue="whole blood", method="welch"),
+            StudyProvenance(geo_accession="GSE2", n_disease=15, n_healthy=11,
+                            platform="P", tissue="whole blood", method="welch"),
+        ],
+        concordance=ConcordanceSummary(n_genes_tested=100, n_concordant=5, n_up=3, n_down=2),
+        created_date="2026-06-23",
+    )
+    sig = build_signature(
+        concordant, prov, tier=ConfidenceTier.A,
+        tier_rationale="Two-study concordance", limitations=["cell-composition confound"],
+        top_n=2,
+    )
+
+    assert [g.gene for g in sig.up_regulated] == ["UP1", "UP2"]  # top 2 up by qvalue
+    assert [g.gene for g in sig.down_regulated] == ["DOWN1", "DOWN2"]
+    assert sig.confidence_tier == ConfidenceTier.A