COMPUTE, DON'T DESCRIBE

NOT for (use other skills instead):

• Running the statistical analysis itself --> Use tooluniverse-statistical-modeling or tooluniverse-rnaseq-deseq2
• Pure gene enrichment without prior analysis --> Use tooluniverse-gene-enrichment
• Pure literature review --> Use tooluniverse-literature-deep-research
• Single variant interpretation --> Use tooluniverse-variant-interpretation

Step 1: Statistical Results to Biological Questions

Map each type of significant finding to the right biological question:

Key principle: Do not stop at "gene X is significant." Ask: significant in what context? Through what mechanism? With what downstream consequence?

Step 2: Multi-Database Evidence Integration

For each significant finding, query multiple sources and synthesize. The pattern:

1. Literature evidence: Search PubMed/EuropePMC for published studies linking your finding to the phenotype. Look for meta-analyses and systematic reviews first.
2. Genetic association evidence: Query GWAS Catalog or OpenTargets to check whether genetic evidence independently supports the association.
3. Pathway context: Query KEGG, Reactome, or WikiPathways to place the finding in a biological pathway. Identify upstream regulators and downstream effectors.
4. Interaction networks: Query STRING or BioGRID for protein-protein interactions. Look for whether your significant genes cluster in the same network neighborhood.
5. Clinical relevance: Check ClinVar for variant clinical significance, DGIdb or ChEMBL for druggability, or ClinicalTrials.gov for ongoing interventions.

Evidence grading (grade each piece of evidence):

Step 3: Causal Reasoning

Statistical association is not causation. Apply these reasoning frameworks:

DAG construction: Before interpreting, sketch the causal directed acyclic graph (DAG).

• Identify potential confounders (common causes of exposure and outcome) -- these must be adjusted for.
• Identify potential mediators (on the causal path) -- do NOT adjust for these if estimating total effect.
• Identify colliders (common effects) -- conditioning on colliders introduces bias.

Triangulation: The same finding supported by different methods with different biases strengthens causal inference.

• Observational association + Mendelian randomization + animal experiment = strong triangulated evidence
• If MR contradicts observational data, suspect confounding in the observational study

Mendelian randomization logic: Genetic variants (instruments) are assigned at conception, so they are not confounded by lifestyle or reverse causation. If a genetic variant that increases exposure X also increases disease Y, this supports X causing Y. Check instrument strength (F-statistic > 10), exclusion restriction (variant affects Y only through X), and pleiotropy (MR-Egger intercept).

Mediation analysis: If gene G is associated with both exposure and outcome, ask: does the exposure effect on outcome go through G? Use the finding's pathway context (Step 2) to propose mediators, then check if adjusting for the mediator attenuates the effect.

Step 4: Cross-Validation

Before reporting a finding as robust, attempt to falsify it:

1. Replication: Search literature and datasets (DataCite, GEO, ArrayExpress) for independent datasets where the same finding can be tested. A finding that replicates in an independent cohort is much stronger.
2. Biological plausibility: Does the mechanism make biological sense? Check if animal or cell models support it (PubMed search for "[gene] knockout [phenotype]" or "[chemical] exposure [cell type]").
3. Genetic support: Check if GWAS evidence supports the direction of effect. If your analysis says gene X is protective but GWAS shows risk alleles increase X expression, there is a contradiction to resolve.
4. Dose-response: If available, check whether the effect increases with dose. A dose-response relationship strengthens causal inference.
5. Negative controls: If possible, test the same analysis on a finding where you expect no association. If the negative control also shows an association, suspect a methodological artifact.

Step 5: Actionable Reporting

Structure the integrated report as follows:

Evidence Summary Table

For each significant finding, produce one row:

Strength Assessment

• Strong: Statistical significance + biological mechanism + independent replication + genetic support
• Moderate: Statistical significance + biological mechanism + partial replication
• Weak: Statistical significance + plausible mechanism but no independent support
• Suggestive: Statistical trend + computational prediction only

Knowledge Gaps and Next Steps

• Which findings lack replication? Propose specific datasets to test in.
• Which mechanisms are inferred but not experimentally validated? Propose experiments.
• Which findings have conflicting evidence? State the contradiction explicitly.
• Generate testable hypotheses: "If finding X is causal, then [experimental prediction]."

Clinical or Public Health Implications

• State whether the finding is actionable now or requires further validation.
• If druggable, identify existing therapeutics and their development stage.
• If a biomarker, assess sensitivity/specificity for clinical utility.