Stanford RNA 3D Folding Part 2

Kaggle Competition: stanford-rna-3d-folding-2

🧬 A complete RNA 3D structure prediction dashboard with reusable design system, interactive visualizations, and production-ready Python pipeline. Built for the $75,000 Stanford Kaggle competition.

🌟 Features

🎨 Reusable Design System

• 26 UI components (Cards, Buttons, Badges, Progress, Tables, Inputs)
• 400+ design tokens (colors, gradients, layouts, effects)
• Glassmorphism UI with Kaggle-inspired aesthetics
• 100% reusable for any Kaggle competition
• Fully documented (DESIGN_SYSTEM.md)

📊 Interactive Dashboard

• Homepage: Competition stats, Part 1 winners analysis, approach comparison
• 3D Visualizer: RNA structure viewer (2D linear + 3D helix)
• Metrics Dashboard: Training progress, comparison charts (Recharts)
• Experiment Tracker: Sortable data tables, hyperparameters, status tracking

🐍 Python Pipeline

• Metrics: TM-score, RMSD, GDT-TS, lDDT implementations
• Dataset: PyTorch Dataset with FASTA parsing
• Features: One-hot encoding, GC content, positional encoding
• Visualization: 3D structure plots, contact maps
• Training: Experiment tracking, checkpointing

Competition Overview

• Goal: Predict the 3D structure of RNA molecules from their sequences
• Prize Pool: $75,000
• Dataset Size: 23.4 GB (competition data)
• Entry Deadline: March 18, 2026
• Hosts: Stanford Medicine + HHMI Janelia
• Challenge: "Solve RNA structure prediction - one of biology's remaining grand challenges"

Background & Motivation

Understanding RNA three-dimensional structure is essential for advancing research in:

• Medicine and drug discovery
• Molecular biology
• Synthetic biology design
• RNA-targeting therapeutics

The structural flexibility of RNA leads to scarcity of experimentally determined data, making computational prediction critical but challenging.

Part 1 Competition Insights (Key Learnings)

Competition Results

• 1,700+ teams participated
• 43 previously unreleased structures as test set
• Top 3 winners: "john", "odat", and team "Eigen"
• Performance: Mean TM-align scores of 0.671, 0.653, and 0.615 on Public leaderboard

Winning Approaches

🏆 Template-Based Modeling (Surprise Winner!)

The most unexpected finding: the top strategy used template-based modeling WITHOUT deep learning.

Key insights:

• Template discovery pipeline outperformed deep learning approaches
• Winners achieved scores within statistical error of CASP16 competition winners
• All three top teams significantly outperformed AlphaFold 3
• Post-competition, the organizers developed RNAPro by integrating Kaggle strategies

Deep Learning Approaches (Also Competitive)

Teams used diverse strategies including:

• RNA foundation models: Aido.RNA, RNet
• Language model-based methods: RhoFold+
• Multi-modal approaches: Combining sequence, structure, and MSA features

Key Takeaway

"Template-based modeling shows growing importance in RNA structure prediction"

State-of-the-Art Methods (2024)

1. RhoFold+ (Leading Deep Learning Method)

• Type: RNA language model-based deep learning
• Training: Pretrained on ~23.7 million RNA sequences
• Performance: Superior on RNA-Puzzles and CASP15
• Advantage: Fully automated end-to-end pipeline

2. AlphaFold 3

• Publisher: DeepMind (Nature, 2024)
• Capability: Biomolecular interactions including RNA
• Limitation: Comparable to ML methods, challenges with ligand binding

3. RoseTTAFoldNA

• Extension of: RoseTTAFold (protein structure prediction)
• Specialty: Protein-DNA and protein-RNA complexes
• Output: 3D structures with confidence estimates

4. Other Notable Methods

• DRfold, DeepFoldRNA, trRosettaRNA
• Template-based methods (validated by Part 1 winners)

Evaluation Metrics

TM-score (Template Modeling score)

• Range: [0, 1] (1 = identical structures)
• Scale-independent structural similarity measure
• Primary metric for Part 1 competition

Other Important Metrics

• RMSD: Root Mean Square Deviation
• GDT-TS: Global Distance Test - Total Score
• lDDT: Local Distance Difference Test
• Pairwise Distance Accuracy: Correctly predicted distances

Project Structure

stanford-rna-3d-folding-2/
├── data/
│   ├── raw/              # Original competition data (23.4 GB)
│   ├── processed/        # Cleaned/transformed data
│   └── external/         # External datasets (PDB, RNA-Puzzles)
├── notebooks/
│   ├── 01-data-exploration.ipynb
│   └── 02-baseline-model.ipynb
├── src/
│   ├── data/
│   │   └── rna_dataset.py            # PyTorch Dataset classes
│   ├── models/
│   │   └── metrics.py                # TM-score, RMSD, GDT-TS, lDDT
│   ├── features/
│   │   └── rna_features.py           # Feature extraction
│   └── utils/
│       └── visualization.py          # 3D plots, contact maps
├── submissions/          # Competition submissions
├── configs/
│   └── default.yaml      # Experiment settings
├── models/               # Saved checkpoints
└── requirements.txt

Setup

1. Install Dependencies

conda create -n rna3d python=3.11
conda activate rna3d
pip install -r requirements.txt

2. Download Competition Data

Data is automatically downloading in background (currently at 36% - 8.5GB/23.4GB).

Once complete, extract:

cd data/raw
unzip stanford-rna-3d-folding-2.zip

Quick Start

Explore Data

jupyter notebook notebooks/01-data-exploration.ipynb

Train Baseline Model

jupyter notebook notebooks/02-baseline-model.ipynb

Custom Training

from src.data.rna_dataset import RNADataset
from src.models.metrics import StructureMetrics

# Create dataset
dataset = RNADataset(sequences=sequences)

# Evaluate predictions
metrics = StructureMetrics()
results = metrics.compute_all(pred_coords, true_coords)
print(f"TM-score: {results['tm_score']:.3f}")

Utility Files Created

✅ Data Processing - src/data/rna_dataset.py

• PyTorch Dataset, FASTA parsing, feature computation

✅ Evaluation - src/models/metrics.py

• TM-score, RMSD, GDT-TS, lDDT, clash detection

✅ Features - src/features/rna_features.py

• One-hot encoding, GC content, secondary structure

✅ Visualization - src/utils/visualization.py

• 3D structure plots, contact maps, distance matrices

Strategy & Approach

Phase 1: Baseline (Weeks 1-2)

• Complete data exploration
• Implement LSTM/Transformer baseline
• Establish evaluation pipeline with TM-score

Phase 2: Advanced Models (Weeks 3-6)

• Implement template-based search (Part 1 winner approach)
• Experiment with RNA language models
• Ensemble multiple approaches

Phase 3: Optimization (Weeks 7-10)

• Hyperparameter tuning
• Model ensembling
• Final submission

Key Resources

Competition

• Competition Page
• Part 1 Discussion

Research Papers

• Template-based RNA prediction (bioRxiv 2025)
• RhoFold+ (Nature Methods 2024)
• AlphaFold 3 (Nature 2024)
• RoseTTAFoldNA (Nature Methods 2023)

External Resources

• DAS Lab - Stanford
• RNA-Puzzles
• CASP RNA

---

Sources:

• Stanford RNA 3D Folding Part 2 - Kaggle
• Template-based RNA prediction
• RhoFold+ (Nature Methods)
• DAS Lab Stanford

🤝 Contributing

Contributions welcome! Please read CONTRIBUTING.md for guidelines.

📄 License

This project is licensed under the MIT License - see LICENSE for details.

🌟 Star History

If you find this useful, please ⭐ star this repository!

🙏 Acknowledgments

• Stanford Medicine & HHMI Janelia for hosting the competition
• Part 1 winners ("john", "odat", "Eigen") for inspiring the template-based approach
• The Kaggle community for discussions and insights

---

Built with 🧬 for RNA 3D structure prediction | Designed for 🏆 Kaggle competitions