EON Ensemble

EON (Ecological Oncology Networks) maps histopathology whole-slide images to per-patch cell-state program (CP) predictions by training a neighborhood-aware model (EONv) on scRNA-seq-derived meta-programs. Those predictions are then used to build spatial patch graphs and detect recurrent tissue motifs across cancer types. The core model, EONv, combines a vision encoder with a cross-attention neighborhood aggregation head that is trained end-to-end with a MSE objective against NMF-derived CP soft labels.

This repository hosts two ensembles: the EON CP ensemble (eon_ensemble.safetensors) and the downstream OptiPRISM microbial ensemble (optiprism_ensemble.safetensors).

Usage

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

Outputs are z-scores per CP relative to the training cohort mean. Use model.target_means and model.target_stds to invert to original NMF-usage scale.

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

OptiPRISM Ensemble

OptiPRISM is a slide-graph regressor that predicts the intratumoral microbial community of a slide from its tissue architecture. Each slide is represented as a spatial patch graph whose nodes carry UNI2-h patch embeddings and whose edges are derived from the EON CP predictions above. A 199-head ensemble of GATv2 message-passing networks with attention pooling (OptiPRISM) maps the graph to a per-genus log relative-abundance vector, trained against PRISM-derived microbial abundances with an MSE objective.

Artifacts:

• optiprism_ensemble.safetensors — flattened weights for all heads.
• optiprism_config.json — shared head architecture and the ordered target_genera.
• optiprism_metadata.json — per-head config, normalization, and validation correlations.

OptiPRISM inference and ensembling utilities (OptiPRISMEnsemble, OptiPRISMi) ship with the EON package; predictions are consensus genus vectors with per-genus ensemble uncertainty.

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