SuShiE🍣

SuShiE (Sum of Shared Single Effect) is a Python package for multiancestry SNP fine-mapping, estimating effect size correlations across ancestries, and computing ancestry-specific prediction weights using either individual-level or summary-level data for molecular or complex traits.

- We detest usage of our software or scientific outcome to promote racial discrimination.

SuShiE is described in

Improved multiancestry fine-mapping identifies cis-regulatory variants underlying molecular traits and disease risk.

Zeyun Lu, Xinran Wang, Matthew Carr, Artem Kim, Steven Gazal, Pejman Mohammadi, Lang Wu, James Pirruccello, Linda Kachuri, Alexander Gusev, Nicholas Mancuso.

Nature Genetics. July, 2025.

Check here for full documentation.

Installation | Example | Notes | Version History | Support | Other Software

Installation

Install the released package from PyPI:

pip install sushie
# or using uv
uv pip install sushie

To install from a source checkout:

git clone https://github.com/mancusolab/sushie.git
cd sushie
pip install .

or with uv:

uv sync

Get Started with Example

SuShiE software is very easy to use:

For fine-mapping using individual-level data:

cd ./data/
sushie finemap --pheno EUR.pheno AFR.pheno --vcf vcf/EUR.vcf vcf/AFR.vcf --covar EUR.covar AFR.covar --output ./test_result

For fine-mapping using summary-level data:

cd ./data/
sushie finemap --summary --gwas EUR.gwas AFR.gwas --vcf vcf/EUR.vcf vcf/AFR.vcf --sample-size 489 639 --gwas-header chrom snp pos a1 a0 z --output ./test_result

It can perform:

• SuShiE: multi-ancestry fine-mapping accounting for ancestral correlation
• Single-ancestry SuSiE (Sum of Single Effect)
• Independent SuShiE: multi-ancestry SuShiE without accounting for correlation
• Meta-SuSiE: single-ancestry SuSiE followed by meta-analysis
• Mega-SuSiE: single-ancestry SuSiE on row-wise stacked data across ancestries (individual-level data only)
• cis-molQTL effect size correlation estimation
• cis-SNP heritability estimation (individual-level data only)
• Cross-validation for SuShiE prediction weights (individual-level data only)
• Convert prediction results to FUSION format, thus can be used in TWAS

See here for more details on how to use SuShiE.

If you want to use in-software SuShiE inference function, you can use following Python code as an example:

from sushie.infer import infer_sushie
# Xs is for genotype data, and it should be a list of numpy array whose length is the number of ancestry.
# ys is for phenotype data, and it should also be a list of numpy array whose length is the number of ancestry.
infer_sushie(Xs=X, ys=y)
# Or summary-level data
# lds is for LD data, and it should be a list of p by p numpy array whose length is the number of ancestry.
# zs is for GWAS data, and it should be a list of numpy array whose length is the number of ancestry/
infer_sushie_ss(lds=LD, zs=GWAS, ns=np.array([100, 100]))

You can customize this function with your own ideas!

Troubleshooting

Installation Issues

Import errors with JAX:

# Ensure you have compatible versions
pip install --upgrade jax jaxlib

Runtime Issues

Out of memory errors:

• Reduce the number of SNPs by shortening your analysis region
• Use --max-select to limit SNPs for purity computation

Slow performance:

• Use --no-update to skip prior updates if convergence is slow
• Reduce --max-iter for initial testing

LD matrix errors (summary-level data):

• Ensure LD matrices are positive semi-definite
• Check that SNP order matches between GWAS and LD files
• Verify sample sizes are correctly specified with --sample-size

For more help, see the full documentation or open an issue.

Notes

• SuShiE currently only supports continuous phenotype fine-mapping for individual-level data.

Version History

Support

For any questions, comments, bug reporting, and feature requests, please contact Zeyun Lu (zeyun_lu@dfci.harvard.edu) and Nicholas Mancuso (nmancuso@usc.edu), and open a new thread in the Issue Tracker.

Other Software

Feel free to use other software developed by Mancuso Lab:

• jaxQTL: a single-cell eQTL mapping tool using highly efficient count-based model (i.e., negative binomial or Poisson).
• MA-FOCUS: a Bayesian fine-mapping framework using TWAS statistics across multiple ancestries to identify the causal genes for complex traits.
• SuSiE-PCA: a scalable Bayesian variable selection technique for sparse principal component analysis
• twas_sim: a Python software to simulate TWAS statistics.
• FactorGo: a scalable variational factor analysis model that learns pleiotropic factors from GWAS summary statistics.
• HAMSTA: a Python software to estimate heritability explained by local ancestry data from admixture mapping summary statistics.
• Traceax: a Python library to perform stochastic trace estimation for linear operators.