MUSS checkpoints

Official checkpoints for MUSS: A Multi-scale and Sequence-based Model for Single-cell Gene Regulation (NeurIPS 2026).

Code and reproduction guide · Data and reference files

Contents

Path	Use
checkpoints/pretrain/muss_pretrain.pt	Initialize fine-tuning or run zero-shot cell clustering
checkpoints/finetune/muss_pbmc10k.pt	PBMC10K gene-expression prediction
checkpoints/finetune/muss_bmmc.pt	BMMC gene-expression prediction
checkpoints/finetune/muss_openproblems.pt	Open Problems 2022 gene-expression prediction
checkpoints/finetune/muss_hca_bone.pt	HCA bone gene-expression prediction

Download and use

Run these commands in the environment from the code repository. Both downloads use the same MUSS_ASSETS directory, preserving the paths expected by the scripts.

export MUSS_ASSETS="$PWD/../MUSS_assets"
hf download kangjiezheng/MUSS-data --repo-type dataset --local-dir "$MUSS_ASSETS"
hf download kangjiezheng/MUSS --include "checkpoints/*" --local-dir "$MUSS_ASSETS"

Preprocess the datasets before using any checkpoint. Follow the guide for preprocessing, fine-tuning and evaluation. These PyTorch checkpoints are loaded through the MUSS scripts. For new pre-training, see the pre-training instructions.

The code is MIT-licensed. AlphaGenome-related materials remain subject to the applicable model terms and output terms.