mims-harvard/tooluniverse-gwas-study-explorer

GWAS Study Deep Dive & Meta-Analysis

Compare GWAS studies, perform meta-analyses, and assess replication across cohorts

---

Overview

The GWAS Study Deep Dive & Meta-Analysis skill enables comprehensive comparison of genome-wide association studies (GWAS) for the same trait, meta-analysis of genetic loci across studies, and systematic assessment of replication and study quality. It integrates data from the NHGRI-EBI GWAS Catalog and Open Targets Genetics to provide a complete picture of the genetic architecture of complex traits.

Key Capabilities

1. Study Comparison: Compare all GWAS studies for a trait, assessing sample sizes, ancestries, and platforms
2. Meta-Analysis: Aggregate effect sizes across studies and calculate heterogeneity statistics
3. Replication Assessment: Identify replicated vs novel findings across discovery and replication cohorts
4. Quality Evaluation: Assess statistical power, ancestry diversity, and data availability

---

COMPUTE, DON'T DESCRIBE

When analysis requires computation (statistics, data processing, scoring, enrichment), write and run Python code via Bash. Don't describe what you would do — execute it and report actual results. Use ToolUniverse tools to retrieve data, then Python (pandas, scipy, statsmodels, matplotlib) to analyze it.

Domain Reasoning: Comparing Studies for the Same Trait

When comparing GWAS studies for the same trait, ask: do they replicate? The same lead SNPs appearing in independent studies is strong evidence of a true association. Different lead SNPs at the same locus may reflect LD differences between populations — they may tag the same causal variant. Different loci entirely may reflect different study designs, phenotype definitions, or population ancestry. Before concluding that a finding failed to replicate, check whether the SNP was even genotyped or imputed in the replication cohort.

LOOK UP DON'T GUESS: effect sizes, p-values, allele frequencies, and LD structure for specific loci. Do not assume a SNP present in one study is present in another — use gwas_get_associations_for_snp to retrieve cross-study data. Do not infer LD blocks from genomic proximity; use credible sets from Open Targets for fine-mapping results.

---

Use Cases

1. Comprehensive Trait Analysis

Scenario: "I want to understand all available GWAS data for type 2 diabetes"

Workflow:

• Search for all T2D studies in GWAS Catalog
• Filter by sample size and ancestry
• Extract top associations from each study
• Identify consistently replicated loci
• Assess ancestry-specific effects

Outcome: Complete landscape of T2D genetics with replicated findings and population-specific signals

2. Locus-Specific Meta-Analysis

Scenario: "Is the TCF7L2 association with T2D consistent across all studies?"

Workflow:

• Retrieve all TCF7L2 (rs7903146) associations for T2D
• Calculate combined effect size and p-value
• Assess heterogeneity (I² statistic)
• Generate forest plot data
• Interpret heterogeneity level

Outcome: Quantitative assessment of effect size consistency with heterogeneity interpretation

Honesty rule (important): A real inverse-variance meta-analysis needs each study's beta + 95% CI. python_implementation.py parses these from the GWAS Catalog beta/or_value + range fields and only then pools effect sizes and computes Cochran's-Q I². When the matched associations don't report usable effect sizes (common), it returns method="descriptive", combined_beta=None, heterogeneity_i2=None, and combined_p_value = the smallest reported p (not a pooled p) — do NOT present a descriptive result as a formal meta-analysis or invent an I².

3. Replication Analysis

Scenario: "Which findings from the discovery cohort replicated in the independent sample?"

Workflow:

• Get top hits from discovery study
• Check for presence and significance in replication study
• Assess direction consistency
• Calculate replication rate
• Identify novel vs failed replication

Outcome: Systematic replication report with success rates and failed findings

4. Multi-Ancestry Comparison

Scenario: "Are T2D loci consistent across European and East Asian populations?"

Workflow:

• Filter studies by ancestry
• Compare top associations between populations
• Identify shared vs population-specific loci
• Assess allele frequency differences
• Evaluate transferability of genetic risk scores

Outcome: Ancestry-specific genetic architecture with transferability assessment

---

Statistical Methods

Meta-Analysis Approach

This skill implements standard GWAS meta-analysis methods:

Fixed-Effects Model:

• Used when heterogeneity is low (I² < 25%)
• Weights studies by inverse variance
• Assumes true effect size is the same across studies

Random-Effects Model (recommended when I² > 50%):

• Accounts for between-study variation
• More conservative than fixed-effects
• Better for diverse ancestries or methodologies

Heterogeneity Assessment:

The I² statistic measures the percentage of variance due to between-study heterogeneity:

I² = [(Q - df) / Q] × 100%

where Q = Cochran's Q statistic
      df = degrees of freedom (n_studies - 1)

Interpretation Guidelines:

• I² < 25%: Low heterogeneity → fixed-effects appropriate
• I² = 25-50%: Moderate heterogeneity → investigate sources
• I² = 50-75%: Substantial heterogeneity → random-effects preferred
• I² > 75%: Considerable heterogeneity → meta-analysis may not be appropriate

Sources of Heterogeneity

Common reasons for high I²:

1. Ancestry differences: Different allele frequencies and LD structure
2. Phenotype heterogeneity: Trait definition varies across studies
3. Platform differences: Imputation quality and coverage
4. Winner's curse: Discovery studies overestimate effect sizes
5. Cohort characteristics: Age, sex, environmental factors

Recommendations:

• Perform subgroup analysis by ancestry
• Use meta-regression to investigate sources
• Consider excluding outlier studies
• Apply genomic control correction

---

Study Quality Assessment

Quality Metrics

The skill evaluates studies based on:

1. Sample Size:

• Power to detect associations (80% power requires n > 10,000 for OR=1.2)
• Precision of effect size estimates
• Ability to detect modest effects

2. Ancestry Diversity:

• Single-ancestry vs multi-ancestry
• Population stratification control
• Transferability of findings

3. Data Availability:

• Summary statistics available for meta-analysis
• Individual-level data vs summary-level
• Imputation quality scores

4. Genotyping Quality:

• Platform density and coverage
• Imputation reference panel
• Quality control measures

5. Statistical Rigor:

• Genome-wide significance threshold (p < 5×10⁻⁸)
• Multiple testing correction
• Replication in independent cohort

Quality Tiers

Tier 1 (High Quality):

• n ≥ 50,000
• Summary statistics available
• Multi-ancestry or large single-ancestry
• Imputed to high-quality reference
• Independent replication

Tier 2 (Moderate Quality):

• n ≥ 10,000
• Standard GWAS platform
• Adequate power for common variants
• Some data availability

Tier 3 (Limited):

• n < 10,000
• Limited power
• May miss modest effects
• Use with caution

---

Best Practices

Before Meta-Analysis

1. Check phenotype consistency: Ensure studies measure the same trait
2. Verify ancestry overlap: High heterogeneity expected if ancestries differ
3. Harmonize alleles: Align effect alleles across studies
4. Quality control: Exclude low-quality studies or associations

Interpreting Results

1. Genome-wide significance: p < 5×10⁻⁸ (Bonferroni for ~1M independent tests)
2. Replication threshold: p < 0.05 in independent cohort
3. Direction consistency: Effect should be same direction across studies
4. Heterogeneity: I² > 50% suggests caution in interpretation

Common Pitfalls

❌ Don't:

• Meta-analyze without checking heterogeneity
• Ignore ancestry differences
• Over-interpret nominal p-values
• Assume replication failure means false positive

✅ Do:

• Always report I² statistic
• Perform sensitivity analyses
• Consider ancestry-stratified analysis
• Account for winner's curse in discovery studies

---

Limitations & Caveats

Data Limitations

1. Incomplete Overlap: Studies may analyze different SNPs
2. Cohort Overlap: Some cohorts participate in multiple studies (inflates significance)
3. Publication Bias: Significant findings more likely to be published
4. Winner's Curse: Discovery studies overestimate effect sizes
5. Imputation Quality: Varies across studies and populations

Statistical Limitations

1. Heterogeneity: High I² may preclude meaningful meta-analysis
2. Sample Size Differences: Large studies dominate fixed-effects models
3. Allele Frequency Differences: Same variant has different effects across ancestries
4. Linkage Disequilibrium: Fine-mapping needed to identify causal variants
5. Gene-Environment Interactions: Not captured in standard meta-analysis

Interpretation Guidelines

When I² > 75%:

• Meta-analysis results should be interpreted with extreme caution
• Investigate sources of heterogeneity systematically
• Consider ancestry-specific or subgroup analyses
• Descriptive comparison may be more appropriate than meta-analysis

When Studies Conflict:

• Check for methodological differences
• Verify phenotype definitions match
• Investigate population stratification
• Consider conditional analysis

---

Tools Used

GWAS Catalog API

• gwas_search_studies: Find studies by trait
• gwas_get_study_by_id: Get detailed study metadata
• gwas_get_associations_for_study: Retrieve study associations
• gwas_get_associations_for_snp: Get SNP associations across studies
• gwas_search_associations: Search associations by trait

Open Targets Genetics GraphQL API

• OpenTargets_search_gwas_studies_by_disease: Disease-based study search
• OpenTargets_get_gwas_study: Detailed study information with LD populations
• OpenTargets_get_variant_credible_sets: Fine-mapped loci for variant
• OpenTargets_get_study_credible_sets: All credible sets for study
• OpenTargets_get_variant_info: Variant annotation and allele frequencies

Glossary

Credible Set: Set of variants likely to contain the causal variant (from fine-mapping)

L2G (Locus-to-Gene): Score predicting which gene is affected by a GWAS locus License: Open source (MIT)