Skip to content

Latest commit

 

History

History
264 lines (182 loc) · 8.38 KB

File metadata and controls

264 lines (182 loc) · 8.38 KB

Stability and Replication

Overview

Robust research findings must demonstrate both internal stability (consistent under perturbation) and external validity (replicable in independent data). This document describes validation approaches for the pathway framework.


Why Stability Matters

In high-dimensional biological data:

  • Many analytical choices are arbitrary
  • Small changes can produce different results
  • Overfitting to specific cohorts is common
  • Publication bias favors novel findings

Stability testing and replication protect against false discoveries and overinterpretation.


Internal Stability Testing

1. Bootstrap Resampling

Procedure:

  1. Resample individuals with replacement (B times, typically B=1000)
  2. Recompute pathway scores and clustering for each bootstrap
  3. Assess consistency of:
    • Pathway rankings
    • Subtype assignments
    • Key associations

Metrics:

  • Bootstrap confidence intervals for pathway scores (95% CI)
  • Cluster stability: proportion of times pairs co-cluster (threshold: ≥0.80 for stable clusters)
  • Rank correlation of pathway importance across bootstraps (Spearman ρ ≥ 0.70 indicates stability)

2. Subsample Stability

Procedure:

  1. Randomly hold out X% of samples (e.g., 20%)
  2. Refit models on remaining 80%
  3. Repeat K times
  4. Measure consistency across subsamples

Useful for:

  • Detecting results driven by small subsets
  • Identifying influential observations
  • Estimating effective sample size

3. Parameter Sensitivity

Procedure:

  1. Vary key parameters within reasonable ranges:
    • Variant weighting schemes
    • Network propagation α
    • Number of clusters K
  2. Assess stability of main conclusions

Stable results: Key findings should persist when parameters vary by ±20% from defaults. Specifically:

  • Variant weighting: α ± 0.2 should not change top-10 pathway rankings
  • Network propagation: α ± 0.1 should maintain >80% overlap in enriched genes
  • Cluster number K: results should be interpretable for K ± 1

4. Initialization Sensitivity

For methods with random initialization (clustering, VAE):

  1. Run multiple times with different seeds
  2. Compare solutions across runs
  3. Report only findings consistent across initializations

Cross-Cohort Replication

Replication Philosophy

Replication is assessed at the pathway and subtype level, not by exact gene overlap.

Level Expectation
Individual variants Will differ
Specific genes May differ
Pathway enrichments Should be directionally consistent (Spearman ρ ≥ 0.60)
Subtype structure Should show similar patterns (ARI ≥ 0.50, Jaccard ≥ 0.40)

Replication Procedure

  1. Train: Develop model on discovery cohort
  2. Freeze: Fix all parameters and thresholds
  3. Apply: Run frozen pipeline on replication cohort
  4. Compare: Assess consistency of findings

What to Compare

Pathway-level:

  • Rank correlation of pathway disruption scores (Spearman ρ ≥ 0.60 for strong replication)
  • Overlap of top-K pathways (Jaccard index ≥ 0.50 for top-20 pathways)
  • Directional consistency: ≥75% of significantly disrupted pathways show same direction of effect

Subtype-level:

  • Similar number of clusters (±1 cluster tolerance)
  • Relative size consistency: largest cluster size differs by <15% between cohorts
  • Consistent pathway signatures per subtype (cosine similarity ≥ 0.70)
  • Comparable phenotype associations (if available): correlation r ≥ 0.50

Partial Replication

Full replication is rare. Consider:

  • Directional consistency: Same direction of effect (≥70% agreement on top pathways)
  • Biological coherence: Same biological themes even if specific pathways differ (semantic similarity ≥ 0.60 using GO hierarchy)
  • Effect size attenuation: Smaller effects in replication (expected 20-40% attenuation due to winner's curse)

Negative Controls

1. Permuted Labels

Procedure:

  1. Randomly permute case/control labels (or outcome variables)
  2. Rerun full analysis
  3. Results should be null (no significant pathways)

Detects: Confounding, data leakage, analysis bugs

2. Randomized Gene Sets

Procedure:

  1. Replace pathway definitions with random gene sets of same size
  2. Rerun pathway scoring
  3. Real pathways should outperform random sets

Detects: Spurious aggregation effects, pathway size artifacts

3. Batch-Only Predictors

Procedure:

  1. Build model using only batch/technical variables
  2. Should not predict biological outcomes

Detects: Batch effects masquerading as biology

4. Held-Out Chromosomes

Procedure:

  1. Exclude one chromosome from training
  2. Test if predictions generalize

Detects: Overfitting to specific genomic regions


Reporting Standards

Required for Main Findings

Element Description
Bootstrap CIs 95% confidence intervals from resampling
Stability metrics Cluster stability, rank correlation
Effect sizes Not just p-values
Sample sizes Per group, after QC

Required for Replication

Element Description
Cohort description Demographics, ascertainment, sequencing
Frozen parameters All settings fixed from discovery
Comparison metrics Quantitative replication assessment
Discrepancies Honest reporting of differences

Required Negative Controls

Control Outcome
Permutation test P-value from null distribution
Random gene sets Comparison to real pathways
Batch prediction Should be non-predictive

Common Pitfalls

1. Circular Analysis

Problem: Using the same data for discovery and validation

Solution: Strict separation of training/test sets; pre-registration of hypotheses

2. P-Hacking

Problem: Testing many comparisons, reporting only significant ones

Solution: Correct for multiple testing; report all tests performed

3. HARKing

Problem: Hypothesizing After Results are Known

Solution: Distinguish exploratory vs. confirmatory analyses; pre-register when possible

4. Survivorship Bias

Problem: Only reporting successful replications

Solution: Publish failures; use registered reports

5. Overfitting

Problem: Complex models that don't generalize

Solution: Regularization; cross-validation; simpler models


Checklist for Robust Results

  • Bootstrap confidence intervals computed
  • Cluster stability assessed (if clustering used)
  • Parameter sensitivity tested
  • At least one negative control performed
  • Cross-cohort replication attempted (or acknowledged as limitation)
  • All analytical choices documented
  • Effect sizes reported (not just p-values)
  • Multiple testing correction applied
  • Limitations explicitly stated

Quantitative Thresholds Summary

Metric Threshold Interpretation
Cluster stability (co-clustering) ≥ 0.80 Pair of samples clusters together ≥80% of bootstraps
Rank correlation (pathways) Spearman ρ ≥ 0.70 Stable pathway importance across bootstraps
Cross-cohort pathway correlation Spearman ρ ≥ 0.60 Strong replication of pathway scores
Top-K pathway Jaccard ≥ 0.50 At least half of top-20 pathways overlap
Directional consistency ≥ 75% Same enrichment direction in replication
Adjusted Rand Index (ARI) ≥ 0.50 Moderate cluster agreement across cohorts
Subtype Jaccard overlap ≥ 0.40 Acceptable sample assignment agreement
Pathway signature similarity Cosine ≥ 0.70 Similar subtype pathway profiles
Phenotype correlation r ≥ 0.50 Comparable clinical associations
Effect size attenuation 20-40% Expected reduction in replication cohort
Parameter sensitivity ±20% Results stable within this range

Summary

Validation Type Purpose Key Methods
Bootstrap Uncertainty quantification Resampling with replacement
Stability Robustness to perturbation Subsampling, parameter variation
Negative controls Detect artifacts Permutation, random gene sets
Replication External validity Independent cohorts

Next Steps