brain-uce-pilot-mix-v3
brain cell embedding model
2 min readupdated
The brain-uce-pilot-mix-v3-d1280l20v2-s2048-bican model by researcher KuanP is a large UCE-like cell embedding model for single-cell brain RNA-seq data. The architecture features 20 layers with a hidden dimension of 1280, processing sentences of 2048 randomly selected genes (up to 3584 tokens) for each dataset. The model generates a 1280-dimensional L2-normalized cell vector as output. This new release is useful for bioinformatics and neuroscience researchers working on clustering and cell typing of human and mouse brain cells.
Key features
- Architecture: 20 layers, hidden dimension of 1280, processing 2048 genes per cell.
- Performance: About 41 ms per cell on an A100 GPU with a batch size of 32.
- Quality: Improved results on bioinformatics benchmarks compared to the previous UCE-Brain v0.5 version.
- Compatibility: Uses the standard gene vocabulary and the same calls as other models in the v2026-09 lineup.
What you can do
- Cell embedding: Run inference on single-cell expression data (
h5ad) to obtain high-quality vector representations. - Cell typing: Apply the resulting embeddings for classification tasks and nearest neighbor (kNN) search on human and mouse brain datasets.
- Model comparison: Test the integration of embeddings into gene expression analysis pipelines alongside other UCE versions.
Example
import numpy as np
import scanpy as sc
from uce_brain.data import load_gene_mapping
from uce_brain.inference import resolve_checkpoint, find_gene_mapping, load_model, load_cell_sentence_params, build_dataset, embed_adata
ckpt = resolve_checkpoint("KuanP/brain-uce-pilot-mix-v3-d1280l20v2-s2048-bican")
params = load_cell_sentence_params(ckpt)
model = load_model(ckpt, device="cuda")
gene_mapping = load_gene_mapping(find_gene_mapping(ckpt))
adata = sc.read_h5ad("cells.h5ad")
ds = build_dataset(adata, gene_mapping, params, species="homo_sapiens", mask_prop=0.0)
adata = adata[np.asarray((ds.adata.X[:, ds.valid_h5ad_indices] > 0).sum(axis=1)).ravel() > 0].copy()
adata.obsm["X_uce"] = embed_adata(model, adata, gene_mapping, params, species="homo_sapiens")
Read more
Details• Benchmarks: UCSC (human fine cell typing) — 0.850 F1; cross-collection transfer — 0.944; WMB-10X mouse supertype — 0.297.
• Hardware requirements: About 16–17 GB of VRAM and 14 GB of system RAM with a batch size of 32; a batch size of 64 requires about 24 GB.
• Limitations: A pilot research model with a single training seed; the 1280-d output size does not directly replace older 512-d or 768-d models.
• License: Other (must contact authors before use outside of evaluation).










Comments