KuanP/uce-multispecies-2025-11-08

UCE multi-species (2025-11-08)

Universal Cell Embedding model trained across 10 species (12 dataset views) in one shared protein-embedding space. Architecture: 8-layer transformer, d_model=512, 4 heads, sequence length 2048. Vocab size 178,331 (legacy 145,469 ESM2 tokens + 32,862 marmoset/chimpanzee additions).

Species covered: human, mouse, macaca_mulatta, callithrix_jacchus, pan_troglodytes, danio_rerio, microcebus_murinus, sus_scrofa, sus_scrofa_domesticus, plus brain-snapshot variants for human / macaca_mulatta / callithrix_jacchus.

Usage

from uce_brain.model import UCEForExpressionPrediction
from uce_brain.data import H5ADDataset, UCEDataCollator, load_gene_mapping
from huggingface_hub import hf_hub_download
import scanpy as sc

model = UCEForExpressionPrediction.from_pretrained("KuanP/uce-multispecies-2025-11-08").eval()
gene_mapping = load_gene_mapping(hf_hub_download("KuanP/uce-multispecies-2025-11-08", "gene_mapping.json"))

adata = sc.read_h5ad("path/to/your.h5ad")
dataset = H5ADDataset(adata, gene_mapping=gene_mapping, species="callithrix_jacchus")

See UCE-brain (multi-species branch) for the full inference pipeline and a notebook running on marmoset BICAN data.