Skip to content

Multi-INTACT

Summary

Multi-INTACT extends the single-molecular-trait INTACT method to jointly evaluate multiple "gene products" of a candidate gene — such as transcript abundance and protein abundance — combining canonical-correlation-based TWAS/PWAS evidence with colocalization evidence in a Bayesian framework. For each gene, it answers two questions: whether it is a putative causal gene (PCG) for a trait through any measured gene product, and, if so, which gene products (expression, protein, or both) drive the effect. Applied to GWAS of 1,408 metabolites, it identified 52–109% more causal genes than single-omics (protein-alone or expression-alone) analysis.[1]

Motivation

Standard TWAS and colocalization analyses each examine one molecular phenotype (usually gene expression) at a time, but eQTLs often fail to explain disease heritability that instead operates through splicing, RNA decay, or protein abundance QTLs, which show minimal eQTL overlap. TWAS is also vulnerable to inflated false positives from LD hitchhiking (a causal variant in LD with, but distinct from, the eQTL), while pure colocalization is well-calibrated but statistically conservative. Multi-INTACT is designed to combine complementary molecular layers to recover both power and calibration.

Method

Building on an instrumental-variables/TWAS framework extended to multiple endogenous variables (e.g., expression E and protein P), Multi-INTACT:

  1. Fine-maps eQTL, pQTL, and GWAS data, then separately colocalizes eQTL-GWAS and pQTL-GWAS signals.
  2. Builds composite genetic instruments (TWAS/PWAS weights) for each gene product from the fine-mapping step, and computes their canonical correlation with the GWAS trait.
  3. Combines the canonical correlation evidence with colocalization evidence via an empirical-Bayes prior (following Dempster's rule of combination, as in the original single-trait INTACT), requiring at least one gene product to show at least modest colocalization (gene-level colocalization probability > 0) as a minimum condition for PCG classification.
  4. Compares posterior probabilities across all 2^p causal models for p gene products (for two products: null, expression-only, protein-only, or both), using an EM algorithm to estimate the empirical prior over models by pooling across all candidate genes, then reports a per-gene-product relevance probability by marginalizing over posterior model probabilities.

The complement of a gene's posterior PCG probability is a local false discovery rate, enabling standard Bayesian FDR control.

Performance

  • Simulations (GTEx chromosome 5 genotypes, 1,198 genes, 100 simulated datasets, ~20% true PCGs): at 5% target FDR, Multi-INTACT achieved higher power than single-gene-product methods (TWAS, colocalization, single-trait INTACT) while maintaining calibrated type I error — plain TWAS showed severely inflated false discovery from LD hitchhiking, and colocalization alone was overly conservative. Summary-statistics-only Multi-INTACT (using population reference-panel LD) gave results nearly identical to individual-level-data analysis, though this represents a best-case scenario with well-matched LD.
  • Gene-product identification: the EM-estimated proportions of causal mechanisms (expression-only, protein-only, both) fell within the interquartile range of ground truth in simulation, and true causal mechanisms consistently received the highest average posterior probability among the three alternatives.
  • Metabolite application: integrating GTEx expression and UK Biobank protein QTL data against GWAS of 1,408 plasma metabolites, single-omics PWAS alone implicated 2,217 genes and colocalization alone only 170, with ~95% of PWAS-only genes showing no colocalization evidence — consistent with widespread LD hitchhiking in PWAS. Single-trait INTACT (combining PWAS and colocalization) implicated 293 PCGs; Multi-INTACT, incorporating both expression and protein evidence jointly, inferred 52–109% more causal genes than either single-omics analysis alone, with both gene products relevant for most gene nominations.

Availability

Software (INTACT R package, includes Multi-INTACT): github.com/jokamoto97/INTACT (AGPL-3.0; also on Bioconductor). Reproducibility code and data: github.com/jokamoto97/multi_intact_paper (archived on Zenodo).

See Also

Citations

[1] Okamoto, J., Yin, X., Ryan, B., Chiou, J., Luca, F., Pique-Regi, R., Im, H.K., Morrison, J., Burant, C., Fauman, E.B., Laakso, M., Boehnke, M., & Wen, X. (2025). Multi-INTACT: integrative analysis of the genome, transcriptome, and proteome identifies causal mechanisms of complex traits. Genome Biology, 26, 19.