EVA

EVA is a foundation model for immunology and inflammation developed by Scienta Lab, a Paris-based biotech, and introduced in February 2026. The model addresses a persistent bottleneck in immune-disease drug development: signals measured in preclinical lab models frequently fail to translate to human biology, and immunology data is fragmented across species, experimental platforms, and resolutions. EVA is positioned by its authors as the first cross-species, multimodal foundation model of the immune system, designed to learn a shared representation of immune state that generalizes across these sources of heterogeneity.

The model harmonizes transcriptomics data spanning different species, sequencing platforms, and resolutions (bulk, microarray, and pseudobulked single-cell), and integrates histology to produce unified patient-level representations. Rather than targeting a single benchmark, the work frames immunology as a continuum across the drug development pipeline and introduces a 39-task evaluation suite that spans discovery, preclinical development, and clinical applications, including target prediction, molecular perturbation modeling, and patient stratification.

Scienta Lab released an open transcriptomics-only version of EVA on Hugging Face. This public checkpoint (48.6M parameters) produces sample-level and gene-level embeddings from RNA-seq profiles in human and mouse; the full multimodal model that additionally incorporates histology is described in the paper but is not the released checkpoint.

#Key Features

• Cross-species harmonization: EVA learns representations that align transcriptomics across species (notably human and mouse), aiming to make signals measured in animal models more directly comparable to human biology.
• Multimodal patient representations: The full model integrates transcriptomics with histology to construct unified patient-level embeddings rather than modality-specific features.
• Platform- and resolution-agnostic inputs: The released model accepts bulk RNA-seq, microarray, and pseudobulked single-cell profiles, harmonizing data collected under different protocols.
• Zero-shot capabilities: The authors report zero-shot results for target efficacy and gene function prediction, allowing the model to be applied without task-specific fine-tuning.
• 39-task evaluation suite: A benchmark spanning discovery, preclinical development, and clinical tasks provides a standardized way to assess immune foundation models across the drug development pipeline.
• Empirical scaling laws: The paper establishes scaling laws relating model size and compute to downstream performance for immune transcriptomics.

#Technical Details

EVA is a transformer-based foundation model. The publicly released transcriptomics version (ScientaLab/eva-rna) has 48.6M parameters and generates both sample-level and gene-level embeddings from RNA-seq inputs in human and mouse, with utilities for gene symbol conversion, batch processing, and an expression decoder. The paper reports state-of-the-art results across the 39-task suite and provides mechanistic interpretability analyses indicating that the learned representations capture biologically meaningful structure across species and technologies. The preprint does not publicly specify the parameter count of the largest multimodal model, the full training-corpus size, or the complete list of species; users requiring those details should consult the full paper. The open checkpoint is distributed under a custom Scienta Lab EVA Model License, so users should review its terms before downstream or commercial use.

#Applications

EVA is aimed at pharmaceutical and translational researchers working in immunology and inflammation. Its embeddings support target identification and efficacy prediction during discovery, modeling of molecular perturbations in preclinical development, and patient stratification for clinical decision-making. Because the released model harmonizes bulk, microarray, and pseudobulked single-cell data across human and mouse, it is well suited to repurposing existing public and proprietary immune transcriptomics datasets, and to comparing signals between animal models and human cohorts within a single representation space.

#Impact

EVA contributes to a growing class of biological foundation models that move beyond single-modality, single-species training toward integrative representations of disease biology. By packaging a cross-species, multimodal approach with a standardized 39-task benchmark and empirical scaling laws, the work offers both a usable model and an evaluation framework that the immunology AI community can build on. The open release of the transcriptomics version lowers the barrier for academic and industry groups to generate immune-state embeddings, while the histology-integrated full model points toward unified patient representations for translational research. As a recent preprint, its real-world adoption and the durability of its benchmark results remain to be established through independent validation.