EON

EON (Ecological Oncology Networks) maps histopathology whole-slide images to per-patch cell-state program (CP) predictions. The contextualizer uses frozen UNI2-h embeddings and a neighborhood-aware EONv ensemble trained against scRNA-seq-derived meta-programs. Its cross-attention head combines each center patch with its eight spatial neighbors and is optimized with BCE for program presence plus MSE for active-program magnitude.

This repository currently hosts the contextualizer as contextualizer_ensemble.safetensors. The former eon_ensemble.safetensors name is retired. A synthesizer for per-patch microbial species presence is under development and is not published yet; the former OptiPRISM artifact names are also retired.

Usage

from EON.models import ContextualizerEnsemble
model = ContextualizerEnsemble.from_hub("ViggyVenkat/EON")

Outputs are raw pyUCell-scale CP scores in [0, 1].

from_hub loads the contextualizer_ensemble.safetensors weights together with config.json (raw UCell target metadata, uncertainty temperature, OOD calibration) and metadata.json (per-head architecture and validation correlations used for ensemble weighting).

Synthesizer (coming soon)

The synthesizer is a spatial bacterial classifier: for each patch and each target species it predicts presence/absence from tissue architecture alone. It is the same EONv head as the contextualizer with n_outputs = n_taxa and score_mode="presence", trained on per-spot Visium targets derived by running PRISM on the matched FASTQs. A 199-head bagged ensemble is weighted by per-taxon validation AUROC.

Once released, outputs will be per-species presence probabilities in (0, 1). Per-taxon operating points fitted on validation by Youden's J will ship in synthesizer_metadata.json; AUPRC should always be read against the reported positive rate.

Planned artifacts:

• synthesizer_ensemble.safetensors — flattened weights for all heads.
• synthesizer_config.json — shared head architecture and the ordered target_cols.
• synthesizer_metadata.json — per-head config, per-taxon AUROC, and operating points.

Package

https://github.com/Viggyvenkat/EON

Contact

EON was developed in the De Laboratory at the Rutgers Cancer Institute.
Contact: Vignesh V. Venkat; vvv11@scarletmail.rutgers.edu