PanMET

PanMET

PanMET is a genome-aware foundation-model framework for RNA sequencing and DNA methylation. It combines independently pretrained modality towers, leakage-free cross-modal alignment, and masked bidirectional cross-modal reconstruction.

This repository contains the retained canonical model weights in the safe, tensor-only safetensors format:

The original training checkpoints contained optimizer and scheduler state. Those training-only objects are intentionally excluded. Patient split manifests are also excluded. manifest.json records SHA-256 hashes for every source checkpoint and exported artifact, and every exported tensor was checked for exact equality after serialization.

Loading weights

PanMET uses a custom PyTorch architecture from the PanMET source repository. Download a component and load its state dictionary with safetensors:

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download(
    repo_id="alexalexak/PanMET",
    filename="stage_c/model.safetensors",
)
state_dict = load_file(path, device="cpu")
model.load_state_dict(state_dict, strict=True)

Stage C reconstruction uses stage_c/rna_decoder.safetensors and stage_c/methylation_decoder.safetensors in addition to the Stage C model. Architecture construction, preprocessing, and inference adapters are provided in the source repository.

Model inputs

• RNA sequencing: 9,721 genomically ordered genes
• DNA methylation: 27,183 ordered CpG islands in 65 chromosome-aware chunks

Input feature identity, order, and preprocessing must match the training caches. The model is intended for research use and is not a clinical device.

Citation

The PanMET manuscript is in preparation. Citation information will be updated when it becomes available.