ancestryaudit 0.3.10

Bias detection and correction framework for genomic cancer AI

AncestryAudit

Bias detection and correction framework for genomic cancer AI.

Detects ancestry-linked performance gaps in copy number variation (CNV)-based cancer classifiers, applies supervised fine-tuning correction, and generates structured audit reports.

Developed from research on ancestry bias in TCGA-LIHC/STAD classification (Yergaliyeva, 2026).

---

Installation

pip install ancestryaudit
# or from source:
pip install -e .

---

Input Format

AncestryAudit works on any CNV feature matrix:

Format	Shape	Notes
np.ndarray	(n_samples, n_genes)	Continuous copy-number values
pd.DataFrame	(n_samples, n_genes)	Column names = gene identifiers

Column values: continuous copy number (e.g. TCGA ABSOLUTE pipeline output, where 2.0 = normal diploid, >2 = amplification, <2 = deletion).

Labels: binary integer (0 or 1), one per sample.

Models: any scikit-learn compatible estimator with fit / predict interface.

---

Quick Start

import numpy as np
from sklearn.linear_model import LogisticRegression
from ancestryaudit import AncestryAuditFramework

# NOTE: Replace with your real CNV matrices
# The example below uses synthetic data for illustration only
rng = np.random.RandomState(42)
X_western = rng.randn(200, 50)
y_western  = (X_western[:, 0] + X_western[:, 1] > 0).astype(int)
X_asian    = rng.randn(80, 50)
y_asian    = (X_asian[:, 0] + X_asian[:, 1] > 0).astype(int)

framework = AncestryAuditFramework()

# Step 0: Check data sufficiency first
power = framework.power_analysis(n_source=200, n_target=80, expected_gap_pp=3.0)

# Step 1: Detect ancestry-linked performance gap
report = framework.audit(
    LogisticRegression(max_iter=1000),
    X_western, y_western,
    X_asian,   y_asian
)
print(f"Gap: {report.gap_pp:.2f}pp, p={report.p_value:.4f}")
print(f"Recommendation: {report.recommendation}")

Output:

Gap: +2.39pp, p=0.0069
Recommendation: correction_required

---

Full Pipeline

from ancestryaudit import AncestryAuditFramework
from sklearn.linear_model import LogisticRegression

framework = AncestryAuditFramework(
    random_state=42,
    n_bootstrap=1000,
    threshold_pp=2.0,   # minimum gap to trigger correction_required
    threshold_p=0.05    # maximum p-value to trigger correction_required
)

# ── Step 1: Filter population-stratification noise (optional) ──────────────
X_western_filtered, kept_genes, filter_log = framework.filter_stratification_noise(
    X_western_df,   # pd.DataFrame with gene names as columns
    gene_list       # list of gene name strings
)
print(f"Removed {filter_log['n_removed']} junk genes, kept {filter_log['n_kept']}")

# ── Step 2: Audit ──────────────────────────────────────────────────────────
audit_report = framework.audit(
    LogisticRegression(max_iter=1000),
    X_western_filtered, y_western,
    X_asian_filtered,   y_asian
)
print(audit_report)
# AuditReport(gap=+2.39pp, p=0.0069, d=1.52, null_CI=[-2.10, 2.15],
#             recommendation='correction_required')

# ── Step 3: Correct ────────────────────────────────────────────────────────
if audit_report.recommendation == "correction_required":
    corrected_model, correction_report = framework.correct(
        LogisticRegression(max_iter=1000),
        X_western_filtered, y_western,
        X_asian_labeled,    y_asian_labeled,  # labeled Asian samples
        n_samples=75                          # how many to include
    )
    print(correction_report)
    # CorrectionReport(delta=+3.51pp, p=0.0012, n_used=75, all_positive=True)

# ── Step 4: Validate ───────────────────────────────────────────────────────
validation_report = framework.validate(
    corrected_model,
    X_asian_holdout, y_asian_holdout   # never seen during correction
)
print(validation_report)
# ValidationReport(pre_gap=+2.39pp, post_gap=-1.12pp, improvement=+3.51pp)

# ── Step 5: Report ─────────────────────────────────────────────────────────
report_dict = framework.generate_report("my_audit_report.json")
framework.summary()

---

API Reference

AncestryAuditFramework

Method	Description	Returns
audit(model, X_source, y_source, X_target, y_target)	Detect gap	AuditReport
filter_stratification_noise(X, gene_list)	Remove OR/pseudogene columns	(X_filtered, kept_genes, filter_log)
correct(model, X_source, y_source, X_target_labeled, y_target_labeled, n_samples)	Fine-tune correction	(corrected_model, CorrectionReport)
validate(corrected_model, X_holdout, y_holdout)	Post-correction audit	ValidationReport
generate_report(save_path)	Full JSON report	dict
summary()	Print pipeline summary	str

AuditReport fields

Field	Type	Description
gap_pp	float	Accuracy gap in percentage points (positive = source better)
p_value	float	Two-sided p-value from label-permutation test
cohen_d	float	Effect size
null_ci	tuple	2.5/97.5 percentiles of permutation null (NOT a CI on the gap)
source_accuracy	float	Model accuracy on held-out source data
target_accuracy	float	Model accuracy on target data
n_source	int	Source sample count
n_target	int	Target sample count
recommendation	str	"correction_required" or "no_action"

CorrectionReport fields

Field	Type	Description
delta_pp	float	Mean accuracy improvement on target holdout (pp)
p_value	float	McNemar p-value (exact binomial when b+c<25, chi-square with continuity correction otherwise)
n_used	int	Target samples used (min of n_samples and available)
refit_robustness	dict	mean/min/max delta_pp across 5 refits (model random_state varied, split held fixed) — tests training stochasticity only
all_positive	bool	True if all 5 refits show positive delta_pp (not just the primary fixed split)
direction_confirmed	bool	True if McNemar p<0.05 AND fine-tuned model beat baseline (c>b)
baseline_accuracy	float	Source-only accuracy on full target
corrected_accuracy	float	Estimated corrected accuracy on full target

ValidationReport fields

Field	Type	Description
pre_gap	float	Performance gap before correction (pp)
post_gap	float	Performance gap after correction (pp)
correction_magnitude	float	pre_gap - post_gap
improvement_pp	float	Accuracy improvement on target (pp)
pre_accuracy_target	float	Target accuracy before correction
post_accuracy_target	float	Target accuracy after correction

---

Filtering Details

filter_stratification_noise removes three gene categories that are known to reflect population-level genetic drift rather than cancer biology:

• Olfactory receptor genes (OR*) — CNV in these clusters varies by ancestral migration history, not cancer type
• Pseudogenes (*P, *P1, *P2, …) — non-functional, high population-stratification signal
• Uncharacterized loci (names containing .) — clone-based placeholder identifiers with no interpretable biological information

Required Methods disclosure: The feature space was defined using all samples prior to train/test split, which constitutes a bounded form of data snooping. No label information was used in this step (Kaufman et al., 2012).

---

Expected Results

Disclaimer: Audit results depend on the model, train/test split, and preprocessing pipeline provided. Results will differ from published paper figures, which used a 7-algorithm ensemble with specific PCA preprocessing. The library is designed for arbitrary input — directional consistency (not numerical identity) with paper results is the correct validation criterion.

Validated reproduction test (Yergaliyeva, 2026):

When provided with the exact paper train/test split (White n=338 train, n=113 test; Asian n=242 evaluation) and gene-aligned PCA features, the library reproduces paper results with 0.003pp numerical precision:

Metric	Paper	Library
Mean PGI	+2.39pp	+2.393pp
Algorithms positive	7/7	7/7
Direction	positive	positive

All 6 API methods (import, audit, correct, validate, report, filter) pass independent correctness tests on synthetic CNV data.

Note on train/test splits: Provide the full dataset and let the framework handle splitting internally. Passing only the training portion causes the framework to re-split a subset, producing different model boundaries and non-comparable gap values.

Statistical power at TCGA scale: power_analysis(n_source=451, n_target=242, expected_gap_pp=3.0) — the realistic gap range for CNV data — returns 11.0% power (UNDERPOWERED). Reaching 80% power requires n_target≈798 (~3.3× the current n=242) and n_source≈1596 (~3.5× the current n=451); both sides must scale together and the needed ratio (2:1 source:target) differs slightly from the current study's ratio (1.86:1). TCGA-LIHC/STAD sample sizes are not sufficient to detect a 1-5pp ancestry-linked gap at conventional power.

---

Citation

If you use AncestryAudit in research, please cite:

Yergaliyeva, D. (2026). Ancestry-linked bias in genomic cancer AI:
Transfer learning correction for East Asian populations.
[Manuscript in preparation]

---

License

MIT License. Copyright (c) 2026 Dana Yergaliyeva.