JointShrink: Joint Bayesian shrinkage in Python

Many analyses estimate an effect in each of several conditions, such as the effect of a genetic variant on gene expression in different tissues. The estimates differ in precision, but the underlying effects may share patterns across conditions. Multivariate adaptive shrinkage (MASH) learns these patterns and uses them to estimate effects jointly (Urbut et al., 2019). We present JointShrink, a Python package that combines covariance-pattern learning, mixture fitting and posterior inference for supplied effect estimates. Gaussian and native multivariate Student-t observation models support posterior contrasts and reuse of a fitted prior. When researchers instead supply outcomes and predictors, a separate Bayesian regression model shares coefficient patterns across outcomes. It supports Gaussian, binary, count, categorical and survival likelihoods, with Gaussian group effects. Analytic checks and selected comparisons with installed R mashr assess numerical agreement; constructed examples illustrate the fitted patterns, uncertainty and predictions.