NEWS

---

MeLSI 1.1.7

• Performance: the weak-learner metric optimizer (optimize_weak_learner_robust, the hot path of ensemble metric learning) is now a C++ kernel (melsi_opt_weak_learner in src/melsi_fstat.cpp, via Rcpp). Profiling showed this gradient-descent loop accounted for ~83% of total runtime; the C++ port runs it ~4.6x faster, for roughly ~2x end-to-end speedup, composing with the 1.1.6 F-statistic kernel and BPPARAM parallelism.
• Reproducibility: results are bit-faithful to the R implementation. The C++ loop draws its within/between-class pairs with R's own index sampler (R_unif_index), so it consumes the global RNG stream identically to the previous sample()-based loop. End-to-end F-statistics reproduce 1.1.6 to floating-point precision (~1e-15 relative) with no p-value changes.

MeLSI 1.1.6

• Performance: the PERMANOVA pseudo-F statistic is now computed by a fused C++ kernel (src/melsi_fstat.cpp, via Rcpp) that accumulates squared distances in a single pass over sample pairs, instead of materialising and squaring the full n x n distance matrix in R. This speeds up the F-statistic by ~4x and the overall analysis by roughly 1.3-2.2x (most for two-group / pairwise analyses), and composes multiplicatively with BPPARAM parallelism.
• Numerical note: results are statistically equivalent but not bit-identical to 1.1.5. Computing squared distances directly (no sqrt round-trip) and summing in a different order changes F-statistics at the ~1e-12 relative level. Across extensive testing no permutation p-value changed; in rare boundary cases a p-value could differ by a single permutation step (1/(n_perms+1)). Set a larger n_perms if exact reproducibility of borderline p-values matters.

MeLSI 1.1.5

• Dependencies: vegan moved from Imports to Suggests. The package computes the PERMANOVA F-statistic directly and no longer calls vegan::adonis2() at runtime, so vegan (and its transitive dependencies) is no longer installed for users; it is needed only to build the comparison section of the vignette.
• Cleanup: removed unused stats::aov, stats::as.dist, and stats::t.test imports and an unreachable t-test pre-filtering branch in the metric learner. Pre-filtering is handled by apply_conservative_prefiltering() as before.
• UX: permutation progress is now shown with a single progress bar instead of one message per permutation.
• UX: melsi() gains an optional seed argument for reproducible runs and now returns an object of class "melsi" with a print() method summarising the F-statistic, p-value, and top features. The returned object is still a list, so existing $ access is unchanged.
• All statistical results (F-statistics, p-values, feature weights) are numerically identical to 1.1.4; these changes do not alter the method.

MeLSI 1.1.4

• Documentation: document the BPPARAM argument to melsi() for parallel permutation testing. The argument has been available since the move to a BiocParallel backend; this release adds a worked example to the README and vignette and surfaces it in NEWS. Pass any BiocParallelParam object (for example MulticoreParam(workers = 8)) to distribute the permutation loop across cores; the default (NULL) runs permutations sequentially.

MeLSI 1.1.3

• Performance: additional ~22% speedup via diagonal vector storage in ensemble aggregation (eliminates B x p^2 matrix allocation per permutation) and rejection sampling in gradient optimizer (eliminates setdiff() allocation per iteration). Combined with 1.1.2, total speedup is ~2.9x over 1.1.1.

MeLSI 1.1.2

• Performance: 2.3x faster permutation testing via vectorized prefiltering, direct PERMANOVA F-statistic (replacing adonis2 overhead), and diagonal metric matrix optimization (replacing O(p^3) eigen decomposition with O(p) column scaling). Results are numerically identical; p-values unchanged.
• Suppress spurious NaN warnings from log2 fold-change on CLR-transformed data.

MeLSI 0.99.9

• Lower minimum R dependency to R >= 4.5.0 to match Bioconductor 3.23 build system

MeLSI 0.99.8

• Bump version to retrigger Bioconductor build report after CI configuration updates

MeLSI 0.99.7

• Switch CI to grimbough/bioc-actions to match the Bioconductor Build System environment

MeLSI 0.99.6

• Update CI to use R-devel and Bioconductor 3.23 targeting R >= 4.6.0

MeLSI 0.99.5

• Bump version to retrigger build after dependency updates

MeLSI 0.99.4

• Update minimum R dependency to R >= 4.6.0

MeLSI 0.99.3

• Address remaining Bioconductor reviewer checklist items for package acceptance
• Enable progress messages for pairwise comparisons in multi-group analysis

MeLSI 0.99.2

• Optimize vignette runtime to meet Bioconductor < 15 minute requirement
• Reduce example dataset size and permutation counts for faster vignette build

MeLSI 0.99.1

• Resolve Bioconductor pre-acceptance review issues: warnings, license NOTE, non-standard directories, and build artifacts
• Add Matrix to dependencies

MeLSI 0.99.0

Initial Bioconductor Submission

• Initial submission of MeLSI (Metric Learning for Statistical Inference) to Bioconductor
• Novel machine learning method for microbiome beta diversity analysis
• Learns optimal distance metrics to improve statistical power in detecting group differences
• Comprehensive validation against standard methods (Bray-Curtis, Euclidean, Jaccard)
• Robust ensemble learning approach with conservative pre-filtering
• Validated on real microbiome datasets with proper Type I error control
• Provides feature importance weights for biological interpretability
• Includes helper functions for CLR transformation and visualization (VIP plots, PCoA)
• Full integration with Bioconductor ecosystem (phyloseq, microbiome packages)