Multi-Modal Foundation · ScientaLab
eva-rna
A transformer foundation model generating sample- and gene-level embeddings from human and mouse transcriptomic profiles.
What eva-rna is
EVA-RNA is a transcriptomics foundation model developed by ScientaLab. Built using a transformer architecture, it generates both sample-level and gene-level representations from RNA-seq expression profiles in human and mouse biology.
The model handles multiple transcriptomic measurement technologies, accommodating bulk RNA sequencing, microarrays, and pseudobulked single-cell RNA-seq. By mapping expression profiles into latent embeddings, it serves as a general-purpose feature extractor for computational biology and immunology workflows.
How it works
EVA-RNA takes aligned gene identifiers and normalised expression values as input. Inputs are tokenised using NCBI Gene IDs, with built-in mapping utilities to convert human gene symbols, human Ensembl IDs, or mouse gene symbols.
The transformer processes input tokens to output 256-dimensional sample-level embeddings alongside 256-dimensional embeddings for each gene. EVA-RNA also includes a pre-trained, deterministic expression decoder designed to reconstruct predicted continuous expression values directly from the generated gene embeddings.
Key facts
- Architecture
- Transformer
- Embedding dimension
- 256
- Supported species
- Human, mouse
- Supported inputs
- Bulk RNA-seq, microarrays, pseudobulked single-cell RNA-seq
- Gene identifier standard
- NCBI Gene ID
What you can use it for
- Sample Embedding Extraction
- Derive 256-dimensional vector representations of whole samples from bulk RNA-seq, microarray, or pseudobulked single-cell experiments for downstream phenotypic or clinical classification.
- Gene-Level Representation Learning
- Extract per-gene contextual embeddings across diverse cellular profiles to study gene behaviour and regulatory patterns.
- Expression Decoding and Prediction
- Reconstruct and predict gene expression levels directly from gene-level latent embeddings using the pre-trained deterministic decoder.
- Cross-Species Transcriptomic Analysis
- Encode expression matrices across both human and mouse cohorts using integrated gene symbol and Ensembl mapping utilities.
Run it here
We list this model but cannot run it yet: excluded (gated): Gated; non-commercial research only; no derivatives..
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Accommodates bulk RNA-seq, microarray, and pseudobulked single-cell profiles across both human and mouse datasets.
- Produces both sample-wide classification embeddings and fine-grained per-gene embeddings.
- Provides a deterministic decoder module to translate latent gene embeddings back into expression values.
- Features automated mixed-precision allocation across Ampere and older GPU architectures, with optional Flash Attention support.
Watch out for
- Expects gene inputs as NCBI Gene IDs, necessitating conversion steps for datasets indexed by Ensembl IDs or gene symbols.
- Flash Attention functionality is restricted to post-Ampere CUDA GPUs running in half precision, and cannot be used on CPU.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · other
This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
Questions
Can I use eva-rna commercially?
Check the licence before commercial use. This model is gated by its authors — its licence has to be accepted on the source repository before it will run.
What transcriptomic data types are supported by EVA-RNA?
EVA-RNA accepts bulk RNA-seq, microarrays, and pseudobulked single-cell RNA-seq profiles from both human and mouse samples.
What are the dimensions of the output embeddings?
The model generates a 256-dimensional sample-level embedding vector and a 256-dimensional embedding vector for each input gene.
Which gene identifiers does the tokenizer use?
The vocabulary relies on NCBI Gene IDs as strings, though built-in mapper dictionaries are provided to convert human symbols, human Ensembl IDs, and mouse symbols.
Can the model decode embeddings back into gene expression values?
Yes, EVA-RNA includes a pre-trained deterministic expression decoder that maps gene embeddings back to predicted expression levels.
How to cite it
Scienta Team, "EVA: Towards a universal model of the immune system", arXiv, 2026.
Written from eva-rna's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.