Changes in version 2.0.0                        

(In Bioconductor devel as version 1.99.0, following the convention for a
major rewrite; it becomes 2.0.0 at the next release.)

A ground-up rewrite. The package now computes 44 epigenetic clocks
behind a declarative registry, runs out-of-core on HDF5 through
BigDataStatMeth, and ships a toolbox to compare, interpret and
quality-control the results.

Architecture

  - Declarative clock registry (clock_register(), clock_list(),
    clock_info(), and the shipped clock_catalog table): one entry per
    clock — target, engine, native platforms, training tissue, licence
    and citation. Adding a clock is adding a row, not editing the
    engine.
  - Manifest-backed resource resolver (mcd_resource(), mcd_backends()):
    coefficients are resolved through a configurable backend chain — a
    local mirror, ExperimentHub, or a direct Zenodo download — and
    normalized to one canonical shape on entry (normalize_coef()).
  - Three predictor engines in this release: linear (including models
    with squared-beta terms), counter (mitotic clocks, including the
    parametric epiTOC2), and neural networks (AltumAge with weights from
    HDF5; a compiled Bayesian network).
  - A unified methylclock() entry point; DNAmAge() and DNAmGA() remain
    as compatible wrappers.

Clocks

  - 44 clocks: chronological (including the cross-platform Garma 2024
    model, the only one here native to EPICv2), biological (PhenoAge),
    pace of ageing (DunedinPACE), telomere length, gestational (cord
    blood and placenta), neonatal (the four NEOage clocks), mitotic
    counters, causal clocks, and nine trait/exposure EpiScores (opt-in
    via target = "trait").
  - Every clock declares the tissue it was trained on and its native
    arrays; restrictively licensed clocks are implemented and clearly
    marked.

Scale

  - Input can be an in-memory matrix or an HDF5-backed matrix;
    whole-array clocks (BLUP, EN) stream from disk without materializing
    the array, and the heavy algebra runs through BigDataStatMeth.
    mc_to_hdf5() converts a matrix or a text file once, to work on disk
    from then on.
  - Results follow the same policy: HDF5 as the durable store with an
    in-memory cache, and persist() to keep a session result.

Missing data

  - Configurable imputation: impute = "mean", "reference", "none", or
    "knn" — a C++ k-nearest-neighbours imputation that runs out-of-core
    on HDF5, exposed also as imputeKNN(); plus a per-sample coverage
    floor (min.perc.sample) so heavily imputed samples are not passed
    off as measured ones.

Interpretation toolbox

  - Cell composition and acceleration: cellCounts(), ageAcceleration()
    (ageAcc, residual, and the cell-adjusted residualCells), plus the
    canonical named measures: IEAA() (Horvath residual adjusted for the
    seven immune covariates of Chen et al. 2016, the three rare subsets
    computed from published CpG estimators) and EEAA() (the fixed-weight
    Hannum blend of Chen et al. 2016, with the parameters published in
    European patent EP 3 494 210 B1). ageAccelerationChen() computes a
    related Chen-style blend whose Klemera-Doubal weights are
    re-estimated from the data at hand; its output column is chenAcc, to
    keep the EEAA name for the canonical measure.
  - Plots: predicted-vs-age (per group), clock agreement, distributions,
    densities by group (overlay, mirror and ridge layouts), acceleration
    by group, forest/dumbbell group differences, per-sample discordance,
    sample PCA, Bland–Altman — one consistent theme, colour-blind-safe
    palettes.
  - Reference ranges (plotReferenceRange()): percentile bands of a named
    reference cohort with the user's samples on top; nine shipped
    reference cohorts spanning four continents, each described by what
    it is.
  - Longitudinal tools: clockTrajectories(), trajectoryRates() (each
    subject's pace, with its uncertainty caveats) and
    plotTrajectories(), plus a real repeated-measures example dataset.
  - Quality control: clockAccuracy(), sampleQC() (outlier score and
    clock discordance) and qcReport() — an HTML report where every
    section carries a written reading of its own numbers.

Data

  - Only estimates and phenotypes ship with the package; coefficient
    tables live in methylclockData/Zenodo, and validation used frozen
    public cohorts. Bundled datasets: clock_catalog, methylclock_demo,
    methylclock_validation, methylclock_references,
    methylclock_longitudinal, methylclock_smoking, methylclock_cells,
    methylclock_betas.

Compatibility notes

  - Estimates were verified against the previous implementations; the
    one deliberate change: Wu uses adult.age = 48, the value of the
    clock's publication.
  - EPICv2 arrays: the pre-2023 clocks do not apply to EPICv2 probe
    names and say so rather than guessing; the Garma model runs on
    EPICv2 natively. A proper probe mapping for the older clocks is
    planned.