FAQ¶

Frequently asked questions about SuShiE.

General Questions¶

What is SuShiE?¶

SuShiE (Sum of Shared Single Effect) is a Bayesian fine-mapping method designed for multi-ancestry genetic studies. It identifies causal variants while accounting for effect size correlations across ancestries.

When should I use SuShiE vs single-ancestry SuSiE?¶

Use SuShiE when you have genetic data from multiple ancestries and want to:

Leverage shared genetic architecture across populations
Improve fine-mapping resolution through diverse LD patterns
Estimate effect size correlations across ancestries

Use single-ancestry SuSiE (via --meta or --mega flags) when:

You only have data from one ancestry
You want to compare results with multi-ancestry SuShiE

What data formats does SuShiE support?¶

Genotype data:

PLINK (.bed/.bim/.fam)
VCF (.vcf, .vcf.gz)
BGEN (.bgen)

Phenotype/Covariate data:

Tab-separated text files

Summary statistics:

Tab-separated GWAS files with Z-scores or effect sizes

Parameters¶

How do I choose the number of causal variants (L)?¶

The -L parameter specifies the maximum number of causal variants. Recommendations:

Start with -L 10 (default) for most fine-mapping analyses
Increase if you expect more causal variants in the region
The algorithm will only identify credible sets for true signals

What does the `--rho` parameter do?¶

The --rho parameter sets the prior correlation of effect sizes across ancestries:

--rho 0.0: Assumes independent effects (equivalent to --indep)
--rho 1.0: Assumes perfectly correlated effects
Default: Learns correlation from data (recommended)

When should I use `--no-update`?¶

Use --no-update to disable prior updates during inference:

When you have strong prior beliefs about effect distributions
For faster runtime on large datasets
When empirical Bayes updates cause convergence issues

Output Files¶

What is a credible set?¶

A credible set is a set of SNPs that contains the causal variant with high probability (default 95%). SuShiE outputs:

SNPs in each credible set
Posterior inclusion probabilities (PIPs)
Purity scores (minimum LD among SNPs in the set)

How do I interpret the PIP?¶

The Posterior Inclusion Probability (PIP) represents the probability that a SNP is causal given SNPs in the model:

PIP > 0.95: Strong evidence for causality
PIP 0.5-0.95: Moderate evidence
PIP < 0.5: Weak evidence

Higher PIPs in credible sets indicate better fine-mapping resolution.

What does purity mean?¶

Purity is the minimum absolute correlation (r²) between any pair of SNPs in a credible set:

High purity (>0.5): SNPs are in high LD, harder to distinguish
Low purity: SNPs are more independent, but set may contain false positives

The --purity threshold (default 0.5) filters out low-quality credible sets.

Performance¶

How can I speed up SuShiE?¶

Several options to improve performance:

Reduce genomic region size:
Reduce iterations:
```
sushie finemap --max-iter 100 ...
```
Limit SNPs for purity calculation:
```
sushie finemap --max-select 500 ...
```

How much memory does SuShiE need?¶

Memory usage depends on:

Number of samples (N)
Number of SNPs (P)
Number of ancestries (K)

Rough estimate: N × P × K × 8 bytes for genotype storage.

For large datasets, consider:

Analyzing smaller genomic regions
Running on a high-memory compute node

Errors¶

“LD matrix is not positive semi-definite”¶

This error occurs with summary-level data when the LD matrix has numerical issues:

Ensure LD was computed from a sufficiently large reference panel
Check for missing or mismatched SNPs
Try adding a small ridge regularization term

“No credible sets found”¶

This may indicate:

No significant signal in the region
Too stringent purity threshold (try --purity 0.1)
Convergence issues (check ELBO in output)

“Sample sizes don’t match”¶

For summary-level data, ensure:

--sample-size matches the number of ancestries
Sample sizes correspond to the correct GWAS files (in order)