Multivariate fine-mapping with mvSuSiE

This vignette shows how to jointly fine-map the genetic effects of one molecular trait measured across multiple contexts.

Learning goals

After completing this vignette, you will be able to:

• distinguish cross-context mvSuSiE from single-context SuSiE and cross-gene fine-mapping;

• prepare matched genotype, phenotype and covariate inputs;

• run the current qtl_dataset_construct+mnm workflow; and

• interpret context-specific posterior effects together with shared variant PIPs.

Background and method

When the same molecular trait is measured in several tissues, cell types or experimental conditions, analyzing each context separately discards information about shared genetic effects. The mnm workflow uses mvSuSiE to model the contexts jointly. It estimates a multivariate effect for each candidate variant and uses patterns learned across contexts to distinguish effects that are shared, context-specific or too weak to resolve.

The workflow first harmonizes genotype, phenotype and covariate data in a QtlDataset. It then fine-maps each requested trait across its available contexts. By default, mvSuSiE uses its canonical prior. A previously fitted TWAS-weight object can be supplied with --prior-twas-weights to provide a data-driven prior; this is optional and should not be confused with running a separate mr.mash analysis inside this workflow.

Use this route for one trait across contexts. Use ordinary susie_twas for one trait in one context, and use mnm_genes when the scientific unit is multiple genes within the same locus.

Worked example

The example fine-maps ENSG00000130538 jointly in two contexts from the same individuals.

Role	Example path	What it supplies
Genotype	input/finemapping/protocol_example.genotype.chr22.bed	PLINK bed/bim/fam trio
Phenotype manifest	input/finemapping/protocol_example.pheno_manifest_multicontext.tsv	One manifest row per context and its phenotype BED file
Covariates	input/finemapping/protocol_example.covariates.tsv	Covariates for the shared samples
Association windows	input/finemapping/protocol_example.association_windows.bed	Coordinates for the requested trait

Context rows must refer to the same trait, genome build and sample population. Sample identifiers must agree across all files.

Run cross-context fine-mapping

qtl_dataset_construct+mnm constructs the harmonized dataset and performs one joint mvSuSiE fit for the selected trait.

Timing: TBD

sos run pipeline/mnm_regression.ipynb qtl_dataset_construct+mnm \
  --name protocol_example \
  --cwd output_mnm \
  --genoFile input/finemapping/protocol_example.genotype.chr22.bed \
  --phenoFile input/finemapping/protocol_example.pheno_manifest_multicontext.tsv \
  --covFile input/finemapping/protocol_example.covariates.tsv \
  --customized-association-windows input/finemapping/protocol_example.association_windows.bed \
  --region-name ENSG00000130538 \
  --transpose-covariates \
  -j 1

The command is expected to create:

• output_mnm/qtl_dataset/protocol_example.qtl_dataset.rds — harmonized genotype, phenotype and covariate data.

• output_mnm/multivariate_fine_mapping/protocol_example.ENSG00000130538.multicontext_bvsr.rds — the cross-context QtlFineMappingResult.

The most useful optional settings are:

Option	Purpose
--contexts	Restrict the fit to named contexts; empty uses all available contexts
--mvsusie-max-iter	Maximum mvSuSiE iterations; default 200
--prior-twas-weights	Optional preceding TWAS-weight RDS used to derive a data-driven prior
--prior-weights-min	Lower bound applied to supplied prior weights
--mvsusie-method-args	Additional mvSuSiE arguments as JSON

Command reference

Check the current interface before adapting the command to a new study.

sos run pipeline/mnm_regression.ipynb -h

Results and interpretation

Cross-context fine-mapping result

The output is a QtlFineMappingResult. The result table identifies the fitted contexts and method; each entry contains the mvSuSiE fit and an attached topLoci table.

Component	Interpretation
Result metadata	Trait, contexts and fitted method
topLoci	Variant PIPs, posterior effect summaries and credible-set membership
mvSuSiE fit	Context-specific posterior effects for each single-effect component
Credible-set purity	Whether variants in a credible set are sufficiently correlated to support localization

suppressPackageStartupMessages(library(pecotmr))

result_file <- "output_mnm/multivariate_fine_mapping/protocol_example.ENSG00000130538.multicontext_bvsr.rds"
result <- readRDS(result_file)

cat("Class:", class(result), "\n")
print(result[, setdiff(names(result), "entry"), drop = FALSE])

entry <- result$entry[[1]]
top_loci <- getTopLoci(entry)
top_loci <- top_loci[order(top_loci$pip, decreasing = TRUE), ]
keep <- intersect(c("variant_id", "pip", "cs_95", "cs_95_purity", "method"),
                  names(top_loci))
head(top_loci[, keep, drop = FALSE], 8)

Rank variants by PIP, then evaluate credible-set size and purity. The multivariate fit can strengthen evidence when effects are consistent across contexts, but a high joint PIP does not mean that the effect has the same magnitude in every context. Inspect the context-specific posterior effects before labeling a variant as shared or context-specific. Broad, low-purity credible sets indicate insufficient localization rather than a single confirmed causal variant.

Limitations and common pitfalls

• All context-specific phenotypes must represent the same trait and use compatible genomic coordinates.

• mvSuSiE does not correct sample mismatches; genotype, phenotype and covariate identifiers must be harmonized first.

• A missing phenotype in some contexts changes the information available to the joint model and should be documented.

• A data-driven prior is optional. Do not supply an unrelated TWAS-weight object merely to enable the option.

• The toy dataset demonstrates the interface but is too small for biological conclusions.

• Joint evidence does not replace examination of context-specific effect sizes and uncertainty.

Next steps

For well-localized signals, compare posterior effects across contexts and annotate the highest-PIP variants. Use the cross-gene mnm_genes workflow only when the next question concerns several genes in the same locus, and use downstream post-processing to summarize credible sets across analyses.