Release Overview

This major release introduces susieR 2.0, a complete architectural redesign that addresses the code duplication and fragmented architecture of the original SuSiE implementation as a result of years of continued development, while adding new features and various performance optimizations, all while maintaining backward compatibility.

susieR 2.0 eliminates the duplicative architecture of the original implementation through a unified framework built on modular design principles. The user-facing interface remains largely unchanged, but the implementation now uses generic functions with data-type specific backends through R’s S3 dispatch system. A key design insight is that most SuSiE extensions reduce to customizations in either Bayes factor computation (per single-effect regression) or residual variance estimation (on the model objective), and the S3 dispatch architecture exploits this separation so that each extension is a local change without modifying the core IBSS algorithm. This architecture has enabled the integration of multiple SuSiE extensions, and serves as the backbone for the refactored mvsusieR package for multi-trait fine-mapping (Zou et al. 2026). Beyond architectural improvements, this release introduces new features including unmappable effects modeling, stochastic LD reference with mixture panel support, enhanced computational speed, and improved IBSS refinement procedures.

New Features

• SuSiE-ash Model: New adaptive shrinkage framework (Stephens 2017) for unmappable effects, extending SuSiE to handle moderately strong but unmappable effects. SuSiE-ash achieves calibration intermediate between standard SuSiE and the more conservative SuSiE-inf under polygenic architectures while maintaining comparable power.
• SuSiE-inf Integration: Implementation of SuSiE-inf (Cui et al. 2024), which models infinitesimal effects alongside sparse causal signals to improve fine-mapping calibration when genetic architecture includes polygenic backgrounds.
• Normal-Inverse-Gamma residual variance prior: New prior option (Servin and Stephens, 2007) on residual variance estimates in single effect regression (SER) to improve credible set coverage and calibration, particularly useful for small sample studies (Denault et al. 2025). Implemented for both individual-level and summary statistics backends.
• Stochastic LD Reference: Support for using genotype matrices directly, including stochastic genotypes in place of LD correlation matrices in summary statistics models (Sun et al., in preparation), with linear algebra optimizations that cache decompositions and exploit the LD structure for speedup.
• RSS LD Quality Control: New empirical Bayes LD bias correction with residual artifact diagnostics for summary-statistics fine-mapping using imperfect or population-mismatched LD references (Sun et al., in preparation).
• Mixture of Reference Panels: Support for combining multiple reference correlation panels with adaptive weight optimization on the susie_rss() sufficient-statistics path (Sun et al., in preparation)
• Beta-Binomial slot activity prior: A new slot_prior_betabinom() option that regularizes slot inclusion, letting the model adapt the effective number of effects rather than treating all L slots as real (Scott and Berger 2010).
• Single-effect summary-statistics interface: New susie_ser() interface for fitting a one-effect credible-set model directly from z-scores or marginal effect estimates and standard errors, without supplying or constructing an LD matrix. Credible sets use the Maller et al. (2012) one-effect model with the SparsePro attainable-coverage filter (Zhang et al. 2023); users can call susie_get_cs(fit, Xcorr = R, ...) afterward to apply LD-based purity filtering.
• mvsusieR Backbone: The S3 dispatch architecture serves as the shared backbone for mvsusieR, enabling our multi-trait fine-mapping methods to build directly on the susieR 2.0 framework with their own S3 dispatch class.
• mfsusieR Backbone: The S3 dispatch architecture serves as the shared backbone for mfsusieR, enabling our multi-trait functional fine-mapping methods to build directly on the susieR 2.0 framework with their own S3 dispatch class.

Enhancements

• SuSiE-inf speedup: Exploits low-rank structure when variables far outnumber samples (p ≫ n), accelerating SuSiE-inf on both the individual-level and summary-statistics backends.
• Performance improvements for RSS model: algebraic optimizations improves RSS models, with further gains on large genomic regions when reference genotype matrices replace LD correlation matrices.
• Correlation-based credible set extension: New cs_extension_corr argument that optionally extends each credible set to absorb every variable whose absolute correlation with a set member exceeds the threshold (near-perfectly correlated proxies). Works from either the genotype matrix X or the correlation matrix Xcorr.
• Wald vs. score z-score handling: New z_method argument ("wald" or "score") on susie_rss() and susie_rss_lambda() controlling the z-score PVE adjustment. The default "wald" rescales Wald-test z-scores onto the model’s residual-variance scale; "score" leaves z-scores that already come from a score test or linear mixed model GWAS unadjusted, avoiding double-shrinkage.
• Median-based purity filtering: New median_abs_corr option that filters credible sets by the median absolute correlation among their variables.
• Attainable Coverage: Post-hoc credible set coverage adjustment as a rough alternative when LD matrices are unavailable for purity filtering.
• Block Coordinate Ascent Refinement: Improved refinement procedure that escapes local optima through alternative initializations and enables residual variance estimation with missing data in multi-trait fine-mapping via mvsusieR.
• Convergence Criteria: PIP-based convergence option alongside traditional ELBO convergence, particularly useful for methods where ELBO computation is expensive.
• Verbose Diagnostics: Detailed diagnostic output for monitoring model fitting behavior.
• Resource-aware purity subsampling: The n_purity default is now "auto", adaptively capping the number of credible-set variables used in purity calculations based on a memory and compute budget rather than a fixed count.
• Residual Variance Estimation: Method of Moments option alongside Maximum Likelihood for improved stability.
• Unit Test Coverage: Comprehensive unit tests covering 100% of code.