🧬 GLIMPS-GNN

Graph-based Liquid-biopsy Inductive Modeling for PreeclampSia

GNN Challenge: cfRNA → Placenta Inductive Prediction

This repository hosts a prediction-only challenge focused on maternal-fetal health modeling using graph learning. Participant code is run outside this repository. Submissions are scored in CI against hidden labels.

🏆 Click me to join competition

Scientific Focus

• Inductive graph learning across cfRNA and placental transcriptomics to detect maternal-fetal health issues.
• Learn transferable representations that generalize to unseen samples and domains rather than treating each dataset independently.

Alignment with BASIRA Lab's Mission

• Prioritizes robust generalization across heterogeneous datasets.
• Uses compute-efficient, non-data-hungry graph learning methods that can run on standard hardware.

Inspiration from GNN Literature

• Draws from studies on inductive learning, message passing, and representation transfer.
• Model design follows DGL Lectures 1.1-4.6, covering:
  • Graph construction from tabular data
  • Node feature encoding
  • Neighborhood aggregation (GraphSAGE-style inductive updates)
  • Mini-batch training via neighborhood sampling
  • Inductive inference on unseen nodes

Overview

• Task: Binary classification (0=Control, 1=Preeclampsia)
• Setting: Inductive transfer from cfRNA (train) to placenta (test)
• Primary metric: F1 Score
• Additional metrics: Accuracy, Precision, Recall
• Public leaderboard: Auto-updated after merged submissions

Dataset Source and Description

Source

• Public datasets from Gene Expression Omnibus (GEO, NIH)
• Maternal plasma cfRNA: GSE192902
• Placental RNA-seq: GSE234729

Data Splits

• Training set: cfRNA samples
• Test set: placenta samples (unseen during training)
• Labels: binary disease status

Purpose and Integration Goal

• Identify and validate cfRNA biomarkers for early prediction of preeclampsia, often before clinical symptoms appear.
• Support research in maternal-fetal health and early detection of preeclampsia.
• Integrate gene expression and clinical metadata to capture subtle risk patterns while handling noisy and imbalanced data for robust and equitable predictions.

🧩 Mandatory Graph Specification

This competition explicitly provides both required graph components:

• Adjacency matrix A: data/public/adjacency_matrix.csv
• Node feature matrix X: derived from data/public/train.csv and data/public/test.csv

Related graph files:

• data/public/graph_edges.csv
• data/public/node_types.csv
• data/public/graph_artifacts.pt

Interpretation:

• A[i, j] = 1 indicates an edge between nodes i and j, else 0
• X is node-by-feature and includes harmonized expression features and released covariates
• Node alignment is by node_id; use data/public/test_nodes.csv (and node files) as the ordering reference so rows in X correspond to the same nodes indexed in A.

🌍 Dataset Difficulty and Realism

The benchmark includes meaningful modeling difficulty:

• 🧪 Noisy and partially missing metadata
• ⚖️ Label imbalance pressure
• 🧬 High-dimensional features relative to sample size (sparsity pressure)
• 🔄 Cross-domain distribution shift (cfRNA -> placenta)
• 🕸️ Inductive generalization to unseen test nodes

⏱️ Computational Affordability

• Full training should not exceed 3 hours on CPU per competition.
• If needed, downsize graph complexity (for example by reducing node count, edge density, or neighborhood sampling size) while preserving task integrity.

Dataset Construction and Preprocessing

build_dataset.ipynb and Kaggle

Objective: Ensure structural compatibility for graph construction and inductive learning by handling expression data, parsing and cleaning metadata, and expression-metadata fusion.

Advanced GNN Implementation

advanced_GNN_model.py

Objective: Implement an advanced inductive GNN for cfRNA -> placenta prediction, ensuring generalizable node representations and inductive learning.

Key Components:

• Graph Construction: Build hetero-graphs using similarity and ancestry edges.
• Node Feature Encoding: Integrate gene expression and metadata into node-level features.
• Neighborhood Aggregation: GraphSAGE-style layers with BatchNorm and ReLU for neighbor information propagation.
• Mini-Batch Training: Use neighborhood sampling for efficient training on large graphs.
• Inductive Inference: Generate predictions for unseen placenta nodes without label leakage.

Starter Assets

• starter_code/advanced_GNN_model.py
• starter_code/baseline.py
• starter_code/build_adjacency_matrix.py
• starter_code/build_graph_artifacts.py

Submission Policy

Submission instructions are in CONTRIBUTING.md.

Key policy:

• Only one submission attempt per participant (enforced in CI)
• Submission files are public but participant predictions are encrypted at rest (predictions.csv.enc); only CI with organizer secrets decrypts for scoring.

Leaderboard

• Public page: https://mubarraqqq.github.io/gnn-challenge/leaderboard.html
• Source CSV: leaderboard/leaderboard.csv
• Rendered markdown: leaderboard.md
• Tie handling: equal scores share rank

Maintainer Regeneration Command

Use this command to regenerate all leaderboard outputs from the canonical pipeline:

python update_leaderboard.py && python competition/render_leaderboard.py

Citation

@dataset{gnn_challenge_2026,
  title={GNN Challenge: cfRNA -> Placenta Inductive GNN for Maternal-Fetal Health Prediction},
  author={Mubaraq Onipede},
  year={2026},
  url={https://github.com/Mubarraqqq/gnn-challenge}
}

License

See LICENSE.