New model · Alibaba Group
LucaCell
A sequence-centric single-cell foundation model that swaps fixed gene IDs for mRNA embeddings
· 6 claims checked against 1 source
What it is
LucaCell is a single-cell foundation model developed by researchers at Alibaba Group. It is a sequence-centric model that represents genes through pre-trained mRNA sequence embeddings rather than static gene annotations. The model was pre-trained on 85 million human and mouse single cells.
How it works
Rather than anchoring gene expression to a fixed gene-ID vocabulary, LucaCell assigns each gene a pre-trained mRNA sequence embedding. Expression values are discretized into bins, and the resulting representation is modeled with a Transformer encoder. Because gene identity is derived from sequence rather than a curated ID list, the same representation can be applied across species, assay types, and organisms lacking curated annotations.
What is new
- Cross-species transfer without manual gene mapping, evaluated on human, mouse and lemur gene expression profiles and human chromatin accessibility data.
- An alignment-free microbial embedding framework built on the same sequence representations, distinguishing bacterial species identity and intra-species physiological states across more than 50 prokaryotic taxa from unaligned reads.
- SNP-aware reconstruction: incorporating donor-specific exonic SNP information into the mRNA embeddings improves gene expression reconstruction accuracy.
- Viral load prediction across five influenza A virus strains, including flagging infection-like transcriptional states in mock-infected cells.
Why it matters
For laboratories working across organisms, tissues, or data modalities, a gene representation grounded in sequence rather than a fixed vocabulary means less manual remapping between datasets. It suggests a path toward applying one model where separate models were previously needed for each species or assay type.
Availability
LucaCell is described in a bioRxiv preprint. The authors have not stated a code or weights release.
The evidence
Each statement above was written with the sentence from the source that supports it, and a program (not a model) checked that every quote and every figure appears in the pages it read. Here they are.
- 01
LucaCell represents genes through pre-trained mRNA sequence embeddings rather than static gene annotations
“a sequence-centric foundation model that represents genes through pre-trained mRNA sequence embeddings rather than static gene annotations” S1
- 02
LucaCell was pre-trained on 85 million human and mouse single cells
“Pre-training on 85 million human and mouse single cells” S1
- 03
It was evaluated on human, mouse and lemur expression profiles, human chromatin accessibility data, unaligned reads from more than 50 prokaryotic taxa, and five influenza A virus genomes
“evaluated on human, mouse and lemur gene expression profiles, human chromatin accessibility data, unaligned reads from more than 50 prokaryotic taxa, and five influenza A virus genomes” S1
- 04
It enables manual-mapping-free cross-species cell type annotation and an alignment-free microbial embedding framework distinguishing bacterial species and physiological states
“enables manual-mapping-free cross-species cell type annotation and an alignment-free microbial embedding framework that simultaneously distinguishes bacterial species identity and intra-species physiological states” S1
- 05
Incorporating donor-specific exonic SNP information into mRNA sequence embeddings improves gene expression reconstruction
“improves gene expression reconstruction by incorporating donor-specific exonic SNP information into mRNA sequence embeddings” S1
- 06
The model predicts cellular viral load across influenza A virus strains and highlights infection-like transcriptional states in mock-infected cells
“predicts cellular viral load across influenza A virus strains while highlighting infection-like transcriptional states in mock-infected cells” S1
