🧬 GNN Challenge: cfRNA → Placenta Inductive GNN for Maternal-Fetal Health Prediction

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

---

Dataset Source and Description

Source

• Publicly available on Gene Expression Omnibus (GEO), maintained by the NIH.

Datasets Used

• Maternal plasma cfRNA data: GSE192902
• Placental RNA-seq data: GSE234729
• Features: 6,000 harmonized gene expression features across two cell types

Training and Test Data

• Training Data: 209–210 cfRNA samples (balanced)
• Test Data: 123–124 placenta samples (inductive, unseen during training)
• Classes: 0 = Control, 1 = Preeclampsia

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.

---

Dataset Construction and Preprocessing (build_dataset.ipynb) and Kaggle

Objective: Ensure structural compatibility for graph construction and inductive learning by Hnadling Expression Data, Parsing and Cleaning Metedata, and Expression-Metadata Fusion

Steps Implemented:

• Expression Data Handling: Load and align expression matrices (sample × gene).
• Metadata Parsing and Cleaning: Normalize clinical and biological attributes; clean string-based metadata.
• Expression–Metadata Fusion: Merge expression and metadata tables using sample IDs to form node-level feature matrices.

Dataset Properties and Complexity:

• Small enough for local download yet challenging: high-dimensional features, rich but noisy metadata, biological heterogeneity.

Constraints:

• No external data
• Fixed feature space
• Inductive setting: test samples unseen during training
• Feasible on standard hardware

Deliberate Complexity:

• Noise and missingness in metadata
• Unbalanced disease labels
• Predictive patterns emerge only through neighborhood aggregation
• Large feature space vs. sample size requiring inductive bias
• Cross-dataset domain shift (cfRNA vs. placenta) requiring generalizable representations

---

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.

📝 Citation

If you use this challenge or dataset in your research, please cite:

@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 file for details.

Challenge Status: ✅ Active
Leaderboard: Live & Auto-updating
Submissions: Open via GitHub PRs
Last Updated: January 7, 2026

Good luck! 🚀 We look forward to your submissions!