Eva

Eva ("Encoder of visual atlas") is a foundation model for tissue imaging that learns multi-scale spatial representations linking molecular, cellular, and sample-level structure. It addresses a persistent gap in computational pathology: histology (hematoxylin and eosin, H&E) images are cheap and ubiquitous but lack molecular detail, while spatial proteomics platforms such as CODEX/PhenoCycler resolve dozens to hundreds of protein markers but are costly and not routinely collected. By pretraining on matched H&E and spatial proteomics from the same tissue, Eva learns a shared representation that bridges the two modalities from a single fixed checkpoint.

The model was developed by Enable Medicine in collaboration with Stanford University, MD Anderson Cancer Center, the University of Tübingen, and the University of Hong Kong, and posted to bioRxiv in December 2025 (co-senior authors Alexandro E. Trevino and Zhenqin Wu). Because it ties imaging morphology directly to spatially resolved protein expression, Eva sits at the intersection of pathology foundation models (such as UNI and Virchow) and spatial-proteomics encoders (such as KRONOS), rather than belonging cleanly to either group.

Once trained, the same Eva checkpoint supports a range of downstream tasks without task-specific fine-tuning, including imputation of protein expression from morphology, zero-shot retrieval across tissues, image quality control and annotation, cell and microenvironment classification, survival modeling, and patient stratification.

#Key Features

• Matched multimodal pretraining: Eva is trained on paired H&E and spatial proteomics from the same tissue regions, letting it align morphological features with spatially resolved protein expression.
• Two-stage hierarchical transformer: A channel stage learns relationships across protein/imaging channels and a spatial stage models relationships across spatial domains, producing representations at molecular, cellular, and sample scales.
• Cross-modal inference: From a fixed checkpoint, Eva can impute protein-marker expression directly from routine H&E morphology, extending molecular readouts to tissues that were never assayed by spatial proteomics.
• Zero-shot retrieval: Eva embeddings support similarity search and annotation across cohorts without retraining, easing data curation and quality control.
• Clinically oriented outputs: The learned representations drive patient stratification and clinical-outcome prediction, connecting tissue-level features to survival and other endpoints.

#Technical Details

Eva is a vision transformer built around a two-stage hierarchical design: a channel stage that learns dependencies across imaging/protein channels and a spatial stage that captures relationships across spatial neighborhoods, yielding multi-scale tissue embeddings. It is pretrained by masked reconstruction of matched spatial proteomics and histopathology images drawn from over 4,000 tissue regions, roughly 64 million cells, and approximately 200 protein biomarkers. The authors evaluate Eva on an external validation cohort of more than 8,000 regions and roughly 50 million cells, reporting that it outperforms both general-purpose and domain-specific baselines — including pathology foundation models such as UNI and Virchow and spatial-proteomics models such as KRONOS — across imputation, classification, retrieval, and outcome-prediction tasks. The preprint releases under a CC BY-NC-ND 4.0 license; as of the December 2025 preprint, model weights and code are not publicly released, and no standalone model card or dataset card is available.

#Applications

Eva is aimed at researchers and clinical teams working with multiplexed tissue data in oncology and tissue biology. Because it can impute spatial-proteomic signal from inexpensive H&E, it offers a route to molecular-resolution analysis for archival or routine slides that were never run on a spatial-proteomics platform. Its zero-shot retrieval and annotation capabilities streamline curation of large imaging cohorts, while its sample-level representations support biomarker discovery, patient stratification, and survival modeling — tasks that bridge research pathology and translational/clinical decision support.

#Impact

Eva extends the foundation-model paradigm from pure histopathology into matched multimodal tissue modeling, demonstrating that a single self-supervised encoder can reason jointly over morphology and spatially resolved protein expression. By reporting gains over established pathology and spatial-proteomics models on a large external cohort, it points toward more general "search engine for biology" workflows in which molecular readouts can be inferred from routine imaging. Its main limitations are practical: as a commercially developed model released without open weights or code, it is not yet independently reproducible, and its training and validation cohorts — though large — concentrate on CODEX/PhenoCycler-style spatial proteomics, so generalization to other platforms and tissue types remains to be established.