JLinAlg v0.3.6

JLinAlg 0.3.6

Differential, regional, multiplicity, and imputation inference

• Added continuous empirical-Bayes and count-aware voom differential models,
  plus a separate median-ratio normalized negative-binomial estimator with
  cross-feature dispersion shrinkage and fixed-dispersion Wald inference.
• Added coordinate/build-aware EWAS regions using signed Stouffer aggregation
  with an explicit exponential spatial-correlation matrix, coverage metadata,
  duplicate-coordinate rejection, and complete-region BH adjustment.
• Added cross-fitted IHW-style bin weights with weighted BH and a separate
  hierarchy-respecting leaf-weighted Bonferroni procedure. Failed hypotheses
  remain in the prespecified family with p=1; hierarchical output controls
  FWER rather than being mislabeled as hierarchical FDR.
• Added mixed-type chained-equations imputation with reproducible independent
  streams, predictive mean matching, binary/categorical probability draws,
  chain diagnostics, Rubin pooling, and Barnard-Rubin finite-sample degrees of
  freedom.
• Added differential, ewas-regions, multiple-test,
  multiple-impute, and mi-pool commands, a worked vignette, a dedicated
  validation report, and website/citation navigation.
• Frozen R 4.6.1/Bioconductor 3.23 fixtures validate effects against limma
  3.68.5, voom, edgeR 4.10.5, and DESeq2 1.52.0. Independent checks cover the
  NB mean ratio, spatial covariance, fold isolation, hierarchy accounting,
  reproducible imputations, and Rubin variance arithmetic.
• Updated the generated scientific references page with a dedicated v0.3.6
  index linking the primary sources for differential, regional EWAS,
  multiple-testing, and multiple-imputation inference.

Latent confounders, batch adjustment, and CLI completion

• Added feature-by-sample PCA, Bioconductor-style iteratively reweighted SVA,
  deterministic AutoSVA, official-source dense PEER VBFA, and parametric or
  nonparametric continuous-data ComBat APIs.
• Added confounders and batch-adjust commands with sample-ID intersection,
  protected/null designs, factor/loadings/weight outputs, adjusted matrices,
  factor-selection paths, convergence metadata, and batch-confounding checks.
• Pinned Bioconductor sva 3.60.0 and PMBio PEER 1.3 source revisions. Frozen
  R fixtures cover PCA, standard SVA, AutoSVA and ComBat; PEER retains an
  explicit native-runtime validation boundary on the Windows host.
• Added a tall-matrix sample-Gram/backend path so global factor extraction never
  allocates a feature-by-feature covariance matrix.
• Wired the preceding feature tranche through glm-predict, iv-regression,
  existing conditional-score, arima-regression --smooth, and the existing
  general --family probit path, with end-to-end CLI tests.
• Added a practical latent-confounder vignette and source/validation audit.

Probit, predictions, IV, conditional scores, and scalable ARIMA

• Added a stable binomial probit GLM through GlmFamilies.probit() and the
  --family probit / binomial-probit CLI aliases. Predictor-aware normal-tail
  likelihood, deviance, IRLS weights, and inference are checked against base R.
• Added ModelPredictions, a shared GLM/GEE API for expected responses,
  scenario differences, risk ratios, and average marginal effects. It carries
  joint scenario covariance through analytic delta-method inference, separates
  population averages from average-covariate predictions, and reports mean
  uncertainty rather than future-outcome intervals.
• Added individual-level linear IV/2SLS with QR projections, SVD rank and
  conditioning checks, homoskedastic/HC0/HC1/cluster covariance, structural-
  residual inference, first-stage partial R-squared/F diagnostics, and base-R
  fixtures. Weak-identification-robust intervals and exogeneity tests remain
  outside this API.
• Added conditional-score to import complete version-1 conditional-GWAS score
  blocks, align exact alleles (and REF/ALT swaps only for Cox or models with an
  intercept), validate model contracts and matrix coverage, condition each
  cohort with a Schur complement, and pool independent cohorts. Unknown cross-
  block covariance is rejected; nonlinear
  cohort refits and the existing mr-estimate path remain distinct.
• Replaced the historical ARIMA smoother's dense observation conditioning and
  1024-date cap with exact diffuse backward information recursions. ARIMA
  regression now exposes joint regression/dynamic parameter covariance,
  explicitly conditional forecasts, and delta-method forecasts that add fitted-
  parameter uncertainty while excluding innovation-variance estimation.
• Added three focused vignettes and updated the GLM, conditional-GWAS, time-
  series, estimator-extension, README, TODO, and website feature guidance.

xWAS genetic architecture, molecular prediction and scores

• Added ldsc for unpartitioned observed-scale heritability, genetic covariance
  and correlation with free intercepts, shared block-jackknife uncertainty,
  and reusable S/V matrix exports.
• Added allele-aware twas / pwas summary association from raw-dosage model
  weights and reference LD, with optional full-rank joint model inference.
• Added genomic-factor for a full-WLS single genetic factor and optional SNP
  GLS effects/heterogeneity conditional on fitted loadings.
• Added score-train / score-apply for Gaussian penalized training, portable
  coefficients, genetic dosage alignment, and held-out score evaluation.
• Added independent R fixtures, executable examples, CLI regression checks,
  and four web/Markdown vignettes. Recorded the remaining five analysis
  families in TODO.md. See validation and limitations.

Rare-variant model metadata and Raremetal2 assessment

• Added version-1 quantitative score metadata, rare-score --trait-id and
  --trait-units, and rare-meta --model-metadata strict. Legacy quantitative
  imports retain logged assumptions; declared unsupported models/calibrations
  and incompatible cohort or score/covariance metadata fail before results.
• Validate covariance column headers explicitly, including rejection of
  unsupported Raremetal2 compressed/allele-aware formats. Quantitative score
  arithmetic and historical supported RMW/rvtests layouts are unchanged.
• Completed the Raremetal2 assessment and trait-model contract,
  with a deterministic R rare-case counterexample and separate implementation
  prerequisites for multiallelic groups, --useExact, and additional traits.

Conditional GWAS aggregate exports and local refits

• Added --conditional-gwas-summary with required --score-genome-build to
  genotype Gaussian, binary logistic, Poisson and model-based Cox scans. It
  appends efficient scores, variances, counts, one-step estimates and explicit
  normal-only calibration status, and writes linked covariance/metadata files.
• Added --condition-on for cohort-side null refitting with requested lead
  genotypes, without exporting participant records. --score-block-size bounds
  covariance blocks; unavailable cross-block covariance is never assumed zero.
• Added streaming Cox genotype scans with offsets, reference-coded categorical
  covariates, right censoring/left truncation and Efron/Breslow ties. Related,
  repeated-subject and other unsupported covariance models fail explicitly.
• Non-Gaussian genotype scans print and log the conditional-GWAS limitation;
  the ordinary output schema is unchanged unless the switch is used. Existing
  mr-estimate remains the Gaussian summary approximation. Rare-event tail
  calibration remains a separate extension; the summary-only importer above
  consumes only complete, compatible score/covariance blocks.
• See the schema, CLI workflow and independent R validation.

Quantile covariance, missing ordinal SEM, and diffuse ARIMA inference

• Added QuantileRegressionInference for exact quantile fits with supplied
  conditional-density sandwich covariance or a pooled iid Gaussian residual
  kernel at a caller-set bandwidth. Inference requires continuous responses,
  independent observations, regular densities, and a converged exact fit.
• Added SemOrdinal.fitPairwiseMissing with explicit -1 missing categories,
  available-pair likelihood and case/cluster covariance. This requires MCAR or
  a justified pair-specific mechanism; general MAR ordinal FIML is not implied.
  Ordinal nonconvergence now suppresses covariance and all parameter inference.
• Added diffuse ARIMA coefficient covariance and standard errors, including
  joint drift/dynamic cross-covariance and seasonal drift units. Unresolved
  information, transform-bound solutions, and nonconvergence suppress inference.
• Independent R fixtures, analytic checks, and reproduction commands are in
  the validation report. Other estimator
  extensions and mathematical limits remain tracked in TODO.

GRM construction CLI and simpler getting started

• Added grm --genotypes FILE --out matrix.tsv for VCF/BCF, BGEN, and dosage
  tables, with MAF/call-rate filters, bounded genotype blocks, backend selection,
  labeled CSV/TSV output compatible with --grm, and timestamped run logs.
• Added GenomicRelationshipMatrix.fromSource, sharing the existing array
  estimator's normalization. Dense sample-matrix storage remains quadratic.
• Simplified the README to requirements, build/run instructions, website links,
  numerical evidence, and license; detailed runtime setup is in the documentation.

Rare-variant score summaries and meta-analysis

• Added rare-score for unrelated-sample quantitative-trait Gaussian scores,
  original-unit score covariance, and BGZF/...

JLinAlg v0.3.5

JLinAlg 0.3.5

Executable mediation and fine-mapping workflows

• The executable JAR now exposes Gaussian mediation workflows for ordinary,
  grouped-REML, and pedigree-REML models, with common complete-case alignment
  and explicit component-model convergence output.
• New susie and coloc commands run summary-statistic fine mapping from
  labeled LD matrices and multi-signal colocalization from the resulting
  effect tables.
• Progressive CLI tutorials now cover association, Mendelian randomization,
  mediation, SuSiE, and colocalization from input files through interpretation.

Association preflight and option validation

• --max-maf and --max-mac complement the existing lower-bound variant
  filters and are applied after sample alignment.
• Phenotype-only --dry-run and --explain now compile and validate the
  selected route before fitting. Genotype transforms, unsupported family/link
  combinations, and the common --explains typo fail with specific guidance.

Pedigree identity and hierarchy

• Family columns now disambiguate duplicate member IDs without partitioning
  the ancestry graph. Globally unique raw observation and parent IDs resolve
  automatically across family labels, while exact family:individual IDs
  remain accepted and ambiguous raw IDs fail with a qualification request.
• Parent-child links recursively define the full pedigree hierarchy, including
  shared ancestors and consanguineous matings. A direct R pedigreemm 0.3.5
  fixture now verifies the additive relationship matrix, inbreeding
  coefficients, and Henderson sparse inverse through a cousin-mating example.
• Runs with zero aligned source-pedigree matches now fail rather than silently
  constructing an all-singleton model. Logs and manifests distinguish matched
  source observations, automatically resolved aliases, and true singletons.

Release downloads: jlinalg-0.3.5.jar is the self-contained executable;
JLinAlg-0.3.5-library.jar is the thin library. Sources and Javadoc JARs and
SHA256SUMS.txt are also included.

JLinAlg v0.3.4

JLinAlg 0.3.4

High-core adaptive omics scheduling

• Automatic feature blocks are sized from both current JVM heap headroom and
  the requested thread count.
• When memory permits, each block queues at least two complete 256-feature
  chunks per scan worker, giving work stealing enough queued work to absorb
  uneven feature-fit times without leaving cores idle at block barriers.
• Under tighter memory, JLinAlg selects the largest complete worker wave that
  fits or safely caps worker capacity. An explicit --block-size remains
  authoritative.

Pedigree singleton handling

• Phenotype individuals absent from the pedigree file are retained as
  unrelated, noninbred singleton founders instead of stopping the analysis.
• Repeated observations with the same missing pedigree ID remain grouped in
  one singleton family. The console, log, and manifest report singleton
  families and observations.

CLI documentation and transform semantics

• Added a comprehensive CLI-only guide and copy-pasteable command-line examples
  to the model vignettes, including grouped REML, pedigrees, GRMs, GLMMs,
  genotype scans, Cox models, and specialized subcommands.
• Omics row transforms intentionally run after phenotype complete-case
  filtering, so winsor_mad and other row-statistic transforms use the final
  analysis sample set.

Release downloads: jlinalg-0.3.4.jar is the self-contained executable;
JLinAlg-0.3.4-library.jar is the thin library. Sources and Javadoc JARs and
SHA256SUMS.txt are also included.

JLinAlg v0.3.3

JLinAlg 0.3.3

Complete-case sample handling

• Phenotype rows with missing values in any model column are omitted by
  default, and the corresponding samples are removed from streamed omics
  inputs before fitting.
• Console, log, and manifest output distinguish aligned samples from the final
  analysis sample count and report how many incomplete phenotype rows were
  omitted.

Exact streamed mixed-model scans

• Numeric Gaussian omics scans now refit variance components by REML for every
  feature through the lmer-like pathway instead of using P3D or EMMAX.
• Non-Gaussian numeric omics scans now refit a first-order Laplace GLMM for
  every feature through the glmer-like pathway instead of PQL.
• Pedigree-enabled scans use direct sparse Henderson relationship precision
  with ancestry-graph inbreeding, matching the intended pedigreemm model
  structure without materializing a dense relationship matrix.
• The selected mixed-fit strategy is recorded in the console preflight, run
  log, and manifest. Genotype LMM scans retain their documented null-model
  P3D pathway.

Delimited output and compact schemas

• A .csv output path now produces correctly quoted comma-separated output;
  .tsv and other existing outputs remain tab-separated.
• Generic numeric omics results omit genotype-only allele, frequency, call,
  imputation, Hardy-Weinberg, and filter columns.
• Per-row error fields are consolidated as failure_reason. Invariant
  omics_type, statistic_type, df_method, partial_r2_method, and
  output_format metadata now live in the run log and manifest instead of
  being repeated in every result row.

Release downloads: jlinalg-0.3.3.jar is the self-contained executable;
JLinAlg-0.3.3-library.jar is the thin library. Sources and Javadoc JARs and
SHA256SUMS.txt are also included.