gMet

gMet is an R package for coverage-aware global and repeat-family DNA methylation metrics from Illumina 450K, EPIC v1 and EPIC v2 arrays.

The package implements the algorithms developed and frozen in the accMetLoss project. It is research software and is not validated for diagnosis, prognosis or treatment selection.

What is included

• Complete literature-defined Alu_global (595 probes), LINE1_global (9,817 probes) and LINE1_NatComm12 metrics.
• Locked DMLA47 scoring, requiring at least 43 of 47 probes; chronological age is mandatory before a value can be labelled DMLA47.
• Observed ALU, LINE1 and locked WGM = 11/28 * ALU + 17/28 * LINE1.
• Genuine end-to-end REMP reconstruction for fullbetaALU, fullbetaLINE1 and fullWGM, including dependency diagnostics, parcel caching and official prediction-quality filtering; reconstructed values are never labelled observed.
• Frozen BroadFamily, YoungFunctional, AncientComparator and CrossPlatformCommonCore ALU/LINE1 manifests for 450K, EPIC v1 and EPIC v2.
• Mean, median and 10-percent trimmed-mean aggregation with cohort- and sample-level coverage gates.
• Healthy-reference fitting and expected-minus-observed methylation-loss acceleration.
• An experimental fit/project interface for frozen Eigen-Transposon PC1 models. It is not a promoted primary metric.

The discontinued 450K historical measurable-subset metrics are intentionally absent from the package.

Installation

install.packages(
  "https://github.com/hyu020/gMet/releases/download/v0.10.0/gMet_0.10.0.tar.gz",
  repos = NULL,
  type = "source"
)
library(gMet)

After the repository is public, development versions can also be installed with remotes::install_github("hyu020/gMet").

Unified scoring

library(gMet)

# beta: probes in rows, samples in columns, values in [0, 1]
gm_methods()
result <- gm_score(
  beta,
  methods = c("dmla47", "alu_global", "line1_global",
              "cross_platform_common_core_alu"),
  platform = "450K"
)

as.data.frame(result)
gm_qc(result)
gm_report(result)

gm_score(methods = "all") evaluates every direct method and preserves ineligible methods as explicit INELIGIBLE_METHOD records instead of silently relaxing a probe definition.

DMLA47

probes <- dmla47_probes()
beta47 <- matrix(
  runif(length(probes) * 20, 0.4, 0.9),
  nrow = length(probes),
  dimnames = list(probes, paste0("sample", seq_len(20)))
)
age <- seq(25, 75, length.out = 20)

raw <- score_methylation_loss(beta47)
dmla47 <- compute_dmla(raw$raw_score47, age, slope = "fixed")

Without age, raw_score47 remains a raw methylation score and is never called DMLA47. DMLA47 is a within-cohort standardized quantity and its cohort means must not be compared as an absolute cross-cohort scale.

Six-metric workflow

scores6 <- assemble_six_metric_scores(
  observed_alu,
  observed_line1,
  remp_alu,
  remp_line1,
  sample_id
)

reference <- gm_fit_reference(
  score = scores6$ALU,
  age = healthy_age,
  cohort = healthy_cohort,
  reference_cohort = "reference",
  metric = "ALU",
  tissue = "whole blood"
)

dmla <- gm_dmla(tumour_alu, tumour_age, reference)

Positive DMLA means more methylation loss than expected for chronological age. Reference transport should be tissue matched and its cohort provenance retained.

End-to-end REMP reconstruction

Install the Bioconductor engine once, then run both repeat families through one interface. The first run creates and persistently caches the complete Alu and LINE-1 parcels; later runs reuse them.

gm_install_remp()
gm_remp_status()

remp_result <- gm_remp_end_to_end(
  beta,
  platform = "auto",  # 450K or EPIC v1
  genome = "hg19",
  ncore = 4,
  seed = 777
)

remp_result$six_metrics
remp_result$reconstructed
remp_result$families$ALU$product
remp_result$families$LINE1$product

Primary fullBeta summaries are calculated after the official REMP::rempTrim() quality filter. Raw products and summaries remain available under each family result. gm_remp_prepare() can build the cache in advance, while gm_remp() remains available for a single family. EPIC v2 is rejected because upstream REMP does not natively support it. gm_score() remains the direct-observation scorer and therefore does not silently reconstruct data.

Scientific safeguards

• ALU and LINE1 remain separate primary families.
• YoungFunctional-LINE1 may be a limited panel on 450K and is not promoted when fewer than 30 cohort-measurable loci remain.
• CrossPlatformCommonCore supports robustness analyses but does not establish absolute calibration across platforms.
• REMP outputs are reconstructed estimates.
• Eigen-Transposon loadings must be fitted in an outcome-free healthy resource, frozen, and only projected into new cohorts.
• Probe manifests and thresholds are versioned scientific assets.

Development

Rscript tests/testthat.R
R CMD build .
R CMD check gMet_0.10.0.tar.gz --as-cran

Bug reports and contributions are welcome at https://github.com/hyu020/gMet/issues.

License

MIT. See LICENSE.md.