gpredomics v1.0.0 — Regression, 8 algorithms, code audit

What's New

Regression Support

• Data.y changed from Vec<u8> to Vec<f64> — continuous y values now supported
• 4 regression fitness functions: spearman, pearson, rmse, mutual_information
• Auto-detected from fit parameter — no config changes needed for classification
• Tested: Spearman=0.675 predicting gene count with 3 species (wetlab dataset)

New Algorithms (4)

• Simulated Annealing — single-solution, temperature-guided perturbation
• Iterated Local Search — fastest algorithm, greedy + perturbation
• LASSO / Elastic Net — coordinate descent mathematical baseline
• MCMC Gibbs — joint variable selection (replaces SBS as default)

Code Quality

• Full code audit: 12 critical + 14 medium issues fixed
• Individual.features: HashMap → BTreeMap for determinism
• ClassificationMetrics struct (Individual.cls refactor)
• Removed ~280 lines of dead code, fixed all clippy warnings

Documentation

• PDF documentation (17 pages)
• Vignette tutorial with Qin2014 examples
• Docs for all 7 algorithms with academic references
• Fully commented param.yaml (13 sections)

Benchmark (Qin2014)

Algorithm	Test AUC	k	Time
Beam	0.923	8	0.3s
SA	0.920	46	0.6s
LASSO	0.882	70	0.1s
ILS	0.880	38	0.3s
GA	0.855	28	12.5s
ACO	0.800	55	6.7s
MCMC-SBS	0.794	51	3.2s
MCMC-Gibbs	0.679	14	2.4s

Full Changelog: v0.9.0...v1.0.0

v0.8.3

New features

• Ant Colony Optimization (ACO) — Max-Min Ant System with parallel construction, local search, pheromone heatmap, evolution timeline, and entropy tracking
• Vote specialization — experts classified as PositiveSpecialist/NegativeSpecialist/Balanced with class-restricted voting and implicit veto
• Feature importance — prevalence and coefficient importance types alongside MDA
• use_fbm parameter for intuitive FBM toggle (replaces confusing fbm_ci_alpha out-of-range hack)
• k_max enforcement during mutations (previously only at initial population)

Bug fixes

• Full determinism — Individual.features converted from HashMap to BTreeMap. All algorithms (GA, ACO, Beam, MCMC) now produce identical results with the same seed (#57)
• MCMC crash on multi-value data_type string (raw,prev)
• Beam search crash on empty classes (data.classes[1] index out of bounds)
• Epoch age overflow when using initial population from previous run
• Security — replaced fast-float with fast-float2 (fixes unsafe empty string vulnerability)

Performance

• Clone removal O(n²) → O(n) with HashSet (28x faster for 5000 individuals)
• Jaccard dissimilarity — sorted vectors instead of HashSet allocation (10x faster)

Documentation

• Full ACO documentation (docs/aco.md)

v0.8.2

Bug fixes

• Fix panics when feature selection results in 0 or few features (#46, #48, #49)
  • Descriptive error messages reporting which criteria caused the failure
  • Lowered hard minimum population size from 50 to 2 (warns when <50)
  • Guard against empty feature sets with binary language
• Fix jury section missing from CSV despite non-empty FBM (#50)
  • Allow jury with 1 expert (was requiring >1)
  • Add info-level logs for each filter step (FBM CI, min_perf, min_diversity)
• Fix non-deterministic results in parallel execution (#40)
  • Replace unstable sort with stable sort + hash tie-breaking in diversity filter
  • Replace find_any with find_first

Enhancements

• Add n_features_total column to CSV report (total features before filtering)
• Suppress backtrace on errors, display clean error messages instead
• Add reproducibility test for GA with same seed

v0.8.0

What's New in v0.8.0

New Features

• FBM CI methods: Configurable binomial CI methods (Wilson, Clopper-Pearson, Agresti-Coull, Wald) for Family of Best Models selection
• Threshold CI bootstrap: Bootstrap-based confidence intervals for model thresholds
• Overfitting safeguards: Inner-fold overfitting penalty and resampling controls
• YAML config validator: Detects unknown parameters/sections and warns about typos
• GA k_min/k_max enforcement: Now limits model size during evolution
• GA documentation: New docs/ga.md
• MDA importance progress logging: Real-time progress during feature importance computation
• Jury expert count controls: min_experts and max_experts parameters

Bug Fixes

• Fix FBM metrics display when keep_trace=false and random sampling is disabled
• Fix rejection case display in jury consensus when individual models reject
• Fix display() panic when data.classes is empty
• Fix default values for threshold CI parameters

v0.7.8

Small update: fix panic in display() when data.classes is empty

v0.7.7

Hotfix: remove runtime git dependency (when saving the experiment)

v0.7.6

Features:

• feat(beam): improve ratio management and ter for k=1 with automatic conversion
• feat(ga): add run_pop_and_data() for training existing populations (GpredomicsR)
• feat(param): add save() method to export experiment parameters

Core improvements:

• refactor(ga): rename kmin/kmax to k_min/k_max for consistency
• refactor(beam): rename kmin/kmax to k_start/k_stop to distinguish from GA

Documentation:

• docs: add dev.md, individual.md, population.md, use.md
• docs: improve README and add badges
• docs: enforce and complete public function documentation

v0.7.5

fix(beam): rename best_models_ci_alpha to best_models_criterion & fix threshold logic

• Rename param.beam.best_models_ci_alpha to param.beam.best_models_criterion to support Predomics-legacy logical compatibility (if > 1).
• Fix critical bug in compute_roc_and_metrics_from_value: other metrics incorrectly used strict threshold (>) instead of inclusive (>=).
• Add related unit tests to validate the corrected behavior.

v0.7.4

Add support for feature and sample annotations (prior weights, penalties, stratified CV), implement native holdout train/test split with stratification, fix critical FBM population bug in CV mode, refactor Jaccard filtering using model groups, add --export-param CLI flag, improve threshold CI rendering, and add GPL-3.0 license.

v0.7.3

Add of the threshold confidence interval concept, add new fitting methods, add numerous unit and integration tests, add a GitHub workflow, add first documentation, add a p-value adjustment via BH, add pruning of individuals, refactor ga, beam, add a common runner for ga/beam/mcmc and other various improvements.