Summary
A bioRxiv preprint introduces LucaCell, a foundation model that represents genes through mRNA sequence embeddings instead of fixed gene identifiers. The authors report applications spanning cross-species cell annotation, microbial analysis, chromatin accessibility and influenza infection states.
A bioRxiv preprint introduces LucaCell, a sequence-centric foundation model designed to analyse single-cell data across species and biological measurement types. The model represents genes using pre-trained messenger RNA (mRNA) sequence embeddings rather than relying only on fixed gene identifiers, a design the authors say improves transfer between datasets that do not share the same gene vocabulary.
The researchers pre-trained LucaCell on 85 million human and mouse single cells. They then evaluated it on human, mouse and lemur gene-expression profiles, human chromatin-accessibility data, unaligned sequencing reads from more than 50 prokaryotic taxa and five influenza A virus genomes.
Contents
What LucaCell changes
Single-cell transcriptomic experiments measure which genes are active in individual cells. Many computational models process those measurements by assigning each gene a fixed identifier, such as a token in a predefined vocabulary. That approach works well when training and test data use compatible annotations, but it can make cross-species or cross-assay transfer more difficult.
LucaCell instead uses embeddings derived from mRNA sequences to represent genes. An embedding is a numerical representation that allows a machine-learning model to compare biological features. Because the representation is based on sequence rather than only on an annotation label, related genes or sequences can contribute information even when the datasets use different identifiers.
The model converts gene-expression values into discrete bins and processes them with a Transformer encoder. Transformers are neural-network architectures that model relationships among elements in a sequence. In LucaCell, the resulting representation is used to describe the state of a cell while retaining sequence-derived information about its genes.
The model's reported applications
The authors report that LucaCell enabled cross-species cell-type annotation without manual gene mapping. The evaluation included human, mouse and lemur expression profiles. In this setting, cell-type annotation means assigning biological identities to cells based on their molecular expression patterns.
The study also applies the model to microbial data. It describes an alignment-free embedding framework for unaligned reads from more than 50 prokaryotic taxa. Sequence alignment normally compares reads with reference sequences before downstream analysis. An alignment-free approach works directly from sequence-derived patterns. According to the preprint, LucaCell's microbial representations simultaneously distinguished bacterial species identity and physiological states within a species.
The model was also evaluated beyond ordinary gene-expression input. The authors report results on human chromatin-accessibility data, an assay that measures regions of DNA more accessible to molecular machinery involved in gene regulation. This tests whether sequence-informed representations can transfer between related but distinct single-cell data types.
A further experiment incorporated donor-specific exonic single-nucleotide polymorphisms into the mRNA sequence embeddings. Exonic variants occur in protein-coding portions of genes. The authors report that this information improved gene-expression reconstruction, meaning the model more effectively recreated expression measurements from its representation.
For influenza A virus, LucaCell was evaluated across five viral genomes. The preprint reports that it predicted cellular viral load and identified infection-like transcriptional states in cells labelled as mock-infected. Viral load refers to the amount of viral material associated with a cell or sample, while a transcriptional state describes the pattern of genes active in that cell.
Evidence status
The work is presented as a bioRxiv preprint and has not been described in the supplied source as peer-reviewed. Its evidence is computational: the model was trained and evaluated on biological sequencing and single-cell datasets rather than tested as a clinical diagnostic or treatment.
The abstract reports the datasets and application areas but does not provide the detailed performance values, comparator models, dataset sizes for each evaluation or error analyses needed to judge the size and consistency of the reported gains. The findings therefore support LucaCell as a research model for cross-species and cross-data-type analysis, while its usefulness in specific biological or clinical workflows will depend on further benchmarking and validation.