Skip to content

Models

Browse all releases in the PoreFM collection on Hugging Face. Choose the model for your sequencing chemistry and analysis method.

RNA

Chemistry: RNA004. Backbone embedding dimension: 512.

Model on Hugging Face Version Checkpoint files Description
PoreFM-RNA 0.1.0 backbone.pt, pretraining.pt Pretrained RNA backbone for embeddings, pairwise comparison, and downstream training; also available with pretraining heads.
PoreFM-RNA_zero-shot 0.1.0 RNA004_zero-shot.pt Canonical-null model with whitening and calibration for detecting deviations from canonical RNA signals.
PoreFM-RNA_prototype 0.1.0 RNA004_prototype.pt Prototypes learned from small labelled datasets, with joint footprinting for m6A, inosine, m5C, pseudouridine, and 2′-O-methylated A/C/G/U.
PoreFM-RNA_m6A-Inosine 0.1.0 model.pt Supervised classifier distinguishing canonical A, m6A, and inosine.
PoreFM-RNA_pseU 0.1.0 model.pt Supervised classifier distinguishing canonical U and pseudouridine (pseU).

DNA

Chemistry: R10.4.1 (E8.2, 400 bps). Backbone embedding dimension: 384.

Model on Hugging Face Version Checkpoint files Description
PoreFM-DNA 0.1.0 backbone.pt, pretraining.pt Pretrained DNA backbone for embeddings, pairwise comparison, and downstream training; also available with pretraining heads.
PoreFM-DNA_zero-shot 0.1.0 R10.4.1_zero-shot.pt Canonical-null model with whitening and calibration for detecting deviations from canonical DNA signals.
PoreFM-DNA_prototype 0.1.0 R10.4.1_prototype.pt Prototypes learned from small labelled datasets, with joint footprinting for 5mC, 5hmC, and 6mA.
PoreFM-DNA_BrdU-EdU 0.1.0 model.pt Supervised classifier distinguishing canonical T, BrdU, and EdU.

Choosing a checkpoint

For modification identities, choose a supervised classifier or a prototype model that includes your target classes. Supervised models use classifiers trained on labelled data. Prototype models learn prototypes from small labelled datasets and use footprinting to account for neighboring effects. Both use contextual PoreFM representations. The released prototype models cover more modification types, so a modification can have a prototype model even when no supervised model is available.

  • backbone.pt contains the pretrained backbone for embeddings, pairwise comparison, and downstream tasks.
  • pretraining.pt contains the same backbone plus pretraining objective heads and the EMA target encoder. Optimizer and scheduler states are excluded.
  • Zero-shot and prototype bundles contain their required backbone and fitted components; separate model-component paths are unnecessary.
  • Supervised model.pt files contain the backbone, classifier, preprocessing, class labels, and base/modification-code defaults.

Derived models use the corresponding RNA or DNA pretrained backbone and include the weights needed for inference. A separate base-model download is unnecessary when using a derived model.

Downloading and using models

Use --model OWNER/REPOSITORY@vVERSION to download a published release automatically. The version is part of the same argument; there is no separate version option. An explicit tag such as @v0.1.0 is required.

porefm prototype classify \
  --model WGLab/PoreFM-RNA_prototype@v0.1.0 \
  --bam_path basecalls.bam --pod5 pod5/ --ref reference.fa \
  --chrom chr1 --motif 'N 0' \
  --effect_penalty 50 \
  --output_dir outputs/prototype

The four inference commands select the appropriate artifact from the release:

Command Example model argument Artifact selected
detect-modification --model WGLab/PoreFM-RNA_m6A-Inosine@v0.1.0 Complete supervised classifier.
zero-shot detect --model WGLab/PoreFM-RNA_zero-shot@v0.1.0 Complete canonical-null bundle.
prototype classify --model WGLab/PoreFM-RNA_prototype@v0.1.0 Complete prototype bundle.
pairwise-comparison --model WGLab/PoreFM-RNA@v0.1.0 Pretrained backbone; supervised classifier releases can also supply embeddings and class filtering.

For supervised pseudouridine detection, use --model WGLab/PoreFM-RNA_pseU@v0.1.0 with --read_based --motif 'T 0'. Its class labels are U and pseU; RNA uridine is represented as T in BAM sequences and modification tags. The checkpoint supplies SAM modification code 17802. The HeLa example runs both RNA supervised models.

Downloads use public HTTPS file URLs with Python's standard library. No Hugging Face package, CLI, account, or login is required for these public releases. PoreFM reads manifest.json, pins the checkpoint download to the resolved repository commit, and verifies its size and SHA-256 before use. It never silently substitutes the latest release for a missing version.

Cache directory

The default cache is ~/.cache/porefm/models, or $XDG_CACHE_HOME/porefm/models when XDG_CACHE_HOME is set. Configure a shared cache through the environment:

export POREFM_MODEL_CACHE=/path/to/shared/porefm-models

Alternatively, add --model-cache-dir /path/to/cache to an inference or download command. This option takes precedence over the environment variable. PoreFM verifies cached checkpoint checksums and reuses valid files without contacting the server. Concurrent processes sharing a cache coordinate downloads. The cache is independent of Hugging Face's SDK cache.

Prefetching and offline use

Download on a machine with internet access before launching a GPU job:

porefm models download \
  --model WGLab/PoreFM-RNA_prototype@v0.1.0 \
  --model-cache-dir /path/to/shared/porefm-models

The command prints the verified local checkpoint path to stdout; download progress goes to stderr. On the compute node, use the same --model and cache directory, adding --offline to prohibit network access. A missing or corrupt cached model then produces an error.

Base-model repositories default to backbone.pt. Request the pretraining checkpoint explicitly when needed:

porefm models download --model WGLab/PoreFM-DNA@v0.1.0 --file pretraining.pt

Local files and reproducibility

Load local checkpoints with --model-path or --model_path. Use either a local path or --model, not both. Local paths never trigger a download, and can point to files obtained manually or through another cache. For training, pass the prefetched checkpoint path to the training command's checkpoint-loading option.

For --model runs, run_manifest.json records the requested model reference, version, resolved repository commit, selected filename, checkpoint SHA-256, and release-manifest SHA-256 under model_source. The local checkpoint path is also recorded. This provenance is retained when using cached models offline.

For setup and examples, see inputs and models, supervised detection, zero-shot detection, pairwise comparison, and prototype-based joint footprinting.

Checkpoint checksums

These SHA-256 checksums identify the exact checkpoint files for version 0.1.0. Each Hugging Face repository also includes SHA256SUMS and manifest.json. After downloading the checkpoints listed in that repository's SHA256SUMS, run this command from their download directory:

sha256sum -c SHA256SUMS

RNA checksums

PoreFM-RNA — 0.1.0 
0ed7a78abb862e9cf7fdbc4d167209014753c440226f5871bc06f8f3e72e6fae  backbone.pt
895db92ecb38d4f505185e7e5c5f55f60caba8dd75ef6ec04b781a7ec3440795  pretraining.pt
PoreFM-RNA_zero-shot — 0.1.0 
5200a9333f507f3c26b13fdb32a7a08da416021b1646640406b6284fa44b372c  RNA004_zero-shot.pt
PoreFM-RNA_prototype — 0.1.0 
79bec4aa0d52bcadf433f97ece376936a70f0cde08420b522addfd698aca2043  RNA004_prototype.pt
PoreFM-RNA_m6A-Inosine — 0.1.0 
828123f81624cf257d0f700c5ad38fe1833c41ac96c2e3d03bb267e1ebb95902  model.pt
PoreFM-RNA_pseU — 0.1.0 
91e249d9d4fb11ae9c029bdd63ea5dd64e1e8cfde680bb017b32c7b68c4b0d03  model.pt

DNA checksums

PoreFM-DNA — 0.1.0 
4367c9a7ef291108a4bfcf6ae8452723c06bb68aa89f715ad4b8115b560986ac  backbone.pt
d120e0e9abba56fe8ec62cf8549e5fb969c84321982a0d55d4113fe670fdcf2f  pretraining.pt
PoreFM-DNA_zero-shot — 0.1.0 
e3e8198166464c3b84a3786abca00087b4b893a7a5e80298087906917b2c26a2  R10.4.1_zero-shot.pt
PoreFM-DNA_prototype — 0.1.0 
83b0fc7a56edd24280f768da367f5e267bc7435b674786184147f8da8087cd26  R10.4.1_prototype.pt
PoreFM-DNA_BrdU-EdU — 0.1.0 
0b5a351fc4fcfd03333ba3ddfc412af8e7d91dc53b51b0ead7b673cbf4ee54a3  model.pt