NephroGrad is a multimodal deep learning framework for kidney disease staging, adapted from the OncoGrad architecture originally developed for oncology applications.
NephroGrad predicts two independent clinical staging outcomes from single-cell/single-nucleus RNA-seq data:
- AKI (Acute Kidney Injury) — KDIGO stage classification
- CKD (Chronic Kidney Disease) — G-stage classification
The model is trained on the KPMP (Kidney Precision Medicine Project) merged single-cell/single-nucleus RNA-seq dataset.
NephroGrad retains the core encoders from OncoGrad, adapted for transcriptomic-only input:
- Gene expression MLP encoder — processes single-cell gene expression profiles
- BioKG encoder — graph neural network over a biological knowledge graph (GO, KEGG, TRRUST regulatory relationships)
- Two independent ordinal prediction heads — one for AKI staging, one for CKD staging
Spatial transcriptomics and histology encoders used in OncoGrad are not included in this application, as the KPMP dataset does not include matched spatial or imaging data.
| File | Description |
|---|---|
nephrograd_mvp.ipynb |
The pipeline actually run on KPMP data. End-to-end: h5ad construction, AKI/CKD label derivation, gene panel selection, BioKG (TRRUST + GO) construction, model training, evaluation, and Integrated Gradients interpretation. |
oncograd_reference_architecture.ipynb |
Reference only — the original OncoGrad prostate/PDAC notebook, kept for architectural lineage. Not run on kidney data. |
NephroGrad is a disease-specific adaptation of the OncoGrad framework, demonstrating the architecture's extensibility beyond its original oncology use case. See OncoGrad for the base framework and prostate cancer/PDAC applications.
This repository is under active development as part of an ongoing grant application.