Skip to content

Changelog

This page gives the short release story. For the full raw changelog, see CHANGELOG.md.

0.9.8

Accelerate MESuSiE covariance EM without pruning SNPs/configurations or changing the model, priors, precision, tolerances or iteration budgets. Evaluate the EM Gaussian likelihood with Cholesky log determinants and multivariate posterior moments with latent square-root conditioning. Preserve the independent objective guard; this is not an LD repair or a guarantee of convergence.

Keep EM as the default and --mesusie-optimizer native as the original-arithmetic sensitivity control. Runtime diagnostics record the Gaussian implementation, BLAS library and requested environment settings without imposing them. See MESuSiE limitations.

0.9.7

Require nonnegative finite ELBO increments for convergence in Python SuSiE, MultiSuSiE and MESuSiE. Use float64 in the MultiSuSiE wrapper and default MESuSiE to dimension-general covariance EM. Honor MESuSiE's requested outer tolerance and CS coverage, with actual iteration counts and inner-optimizer diagnostics. These changes do not promise calibrated fine-mapping under LD mismatch or universal population-order invariance.

0.9.6

Corrected native convergence metadata for SuSiEx and MESuSiE. SuSiEx reads FAIL, NULL, or normal CS rows directly from its .cs file and rejects missing or malformed output. MESuSiE uses its final two computed ELBO values with the original difference threshold of 0.001 instead of defaulting a missing flag to True. Insufficient or nonfinite evidence is False; finite negative changes still satisfy the native criterion. No ELBO trace is exported.

MultiSuSiE, PIP/CS parsing, and existing empty-on-nonconvergence behavior are unchanged. Corrected flags can activate that existing behavior. This patch does not add convergence columns to CLI outputs or change purity-filter metadata propagation. See native convergence metadata.

0.9.5

Added --ld-weighting ess|se to meta and pipeline, with ESS as the default. Both modes match GWAS and LD within cohorts, recompute matched IVW statistics, and merge LD with geometric per-SNP normalization. ESS is no longer the legacy pairwise average, so use fresh output directories. Mode manifests and matching audits make the change explicit and protect resume from incompatible outputs.

Also fixed meta output metadata when the input TSV uses a different column order. Adaptive L, float16 LD storage, and the absence of default PSD repair are unchanged. See matched LD weighting.

Known limitation: INS validation found population-order sensitivity in MESuSiE and MultiSuSiE. Keep population order fixed when comparing results; this release does not resolve that sensitivity or change their fitting algorithms.

0.9.4

Fixed the adaptive L (--adaptive-max-causal) fallback logic.

When a larger L failed, the loop used to restart from initial - 1 and discard earlier successful runs (5 → 10 → 15 fails → 4). It now falls back one step at a time from the failed L (15 fails → 14 → 13 → …), reuses an earlier valid result when it reaches that L, never exceeds L = 20, and never accepts a non-converged result as success. See Adaptive L.

0.9.3

Documented adaptive L (--adaptive-max-causal) in more detail.

The guide now explains how CREDTOOLS changes max_causal, when it increases or decreases L, how non-convergence and purity filtering affect the retry loop, and which tools support the mechanism.

0.9.0

Added SuSiE-inf support through --tool susie_inf.

Use it when you want a SuSiE 2.0-style model that can account for a broad background component in addition to sparse signals.

0.8.0

Added SuSiE-ash support through --tool susie_ash.

This is useful for loci where a few strong effects sit on top of weaker background effects.

0.7.0

Added CARMA support through --tool carma.

CARMA is R-based and includes its own outlier-aware modeling behavior.

0.6.x

Improved convergence handling and adaptive max_causal behavior.

The main user-facing change: adaptive runs can retry more cleanly when a tool does not converge.

0.5.x

Improved QC, outlier handling, heterogeneity output, and credible set post-processing.

Highlights:

  • C1b and adaptive QC options,
  • cleaned loci outputs after outlier removal,
  • better per-locus credible set columns,
  • heterogeneity summaries from meta-analysis.

Earlier Releases

Earlier releases added MESuSiE, broad wrapper test coverage, preprocessing tests, and the current CLI structure.

Use GitHub releases or the root changelog when you need exact patch-level details.