Pretrained models & bundles

Overview

Pretrained MethScope models are distributed from the methscope_data repository as self-contained bundles — a single file that carries both the model and the MRMP feature definition it needs, so a query .cg can be run without supplying a separate .cm reference.

There are three bundle kinds:

Using a pretrained bundle

Because a bundle carries its own MRMP, pass it directly — no unbundling needed:

methscope predict query.cg   hg38_celltype.ubjx     > labels.tsv
methscope deconv  mixture.cg  hg38_65celltypes.refx > props.tsv
methscope upscale -o out.cg   hg38_10k1.updecx query.cg

Inspect a bundle — its framework mark, on-disk layout, and model summary — without running it:

methscope inspect hg38_celltype.ubjx

Bundling and unbundling

Wrap any model together with the MRMP it needs. By convention the bundle gets an x suffix (.ubj → .ubjx, .ref → .refx, .updec → .updecx):

# classifier: booster + MRMP (+ class labels, + framework mark)
methscope bundle -m ref.mrmp -k xgboost -l labels.tsv -o model.ubjx booster.ubj

# unwrap a bundle back into its parts (names derived from the bundle path)
methscope unbundle model.ubjx          #  -> model.ubj + model.mrmp

Bundle format (technical)

A bundle (magic MSBNDL1) uses an MRMP-first layout: the MRMP .cm is the file prefix (offset 0), so YAME tools read it directly (yame summary model.ubjx works), and any subcommand that expects a <ref.mrmp> also accepts a bundle in that slot. After the MRMP come the container sections:

• mrmp — the MRMP feature definition (a fmt2 YAME .cm).
• kind — the framework mark: xgboost / threshold / logistic for a classifier (predict requires it), or refx for a deconvolution reference. Upscale decoders need no mark.
• outcpg (upscale only) — a genome-wide mask of the imputed CpG locations, letting upscale emit a whole-genome .cg.
• model — the raw inner model bytes: an XGBoost UBJ booster, a methscope-linear text spec, an .updec decoder, or a .refx signature TSV.

Because the model and its MRMP travel together, predictions are reproducible and you never have to hand-match a loose .mrmp to the right model.

Models in R

The MethScope R package ships built-in models — e.g. Zhou2025_HumanAtlas_P1000() and Liu2021_MouseBrain_P1000() — carrying their MRMP metadata, used the same way through PredictCellType(). See the Get started tutorial and the methscope-cli command-line guide.