TACO 🌮: TabPFN-Augmented Causal Outcomes for Early Detection of Long COVID

This repository contains all code, data processing pipelines, and evaluation scripts to reproduce the results from our paper:

TACO: TabPFN-Augmented Causal Outcomes for Early Detection of Long COVID

Key Innovation: TACO integrates foundation models with causal inference to achieve presymptomatic Long COVID detection, demonstrating that mechanistic understanding through causal gene selection substantially outperforms traditional statistical approaches.

---

🎯 Project Overview

Long COVID affects 10-40% of COVID-19 survivors, yet current approaches rely on symptom-based diagnosis, preventing early intervention when it could be most effective. TACO transforms Long COVID management from reactive diagnosis to proactive prevention.

🔬 The TACO Framework

TACO (TabPFN-Augmented Causal Outcomes) synergistically combines:

• Foundation Models: TabPFN, a pre-trained transformer for tabular data
• Causal Inference: Differential Causal Effects (DCE) analysis to identify mechanistic drivers
• Presymptomatic Detection: Risk stratification during acute infection, before symptom manifestation

🏆 Key Achievements

• Superior Performance: 69.2% precision, 0.663 ROC-AUC using only 411 causal genes
• Outperforms Benchmarks: 4.4% precision improvement, 10.1% ROC-AUC improvement over best benchmark using 500 genes
• Biological Validation: 23.6% of causal genes confirmed in Long COVID literature
• Clinical Relevance: High precision minimizes false positives for clinical deployment

Figure 1: TACO integrates DCE-identified causal genes with TabPFN's foundation model capabilities for early Long COVID detection

---

📊 Main Results

Performance Comparison

Model	Genes	Precision	ROC-AUC
TACO (Ours)	411 CG	69.2% ± 2.1%	0.663 ± 0.023
Bagging	500 MVG	66.7% ± 2.8%	0.651 ± 0.038
RandomForest	500 MVG	68.6% ± 2.8%	0.641 ± 0.059
SVC	500 MVG	67.8% ± 1.5%	0.639 ± 0.063

CG: Causal Genes

MVG: Most Variable Genes

TACO achieves superior performance using fewer, causally-selected genes (p < 0.01, Wilcoxon signed-rank test)

Some Key Validated Causal Genes

• Viral Entry: AR (androgen receptor), TMPRSS2, ACE2
• Immune Response: TP53, CDKN1A, RB1
• Epigenetic Regulation: CREBBP, EP300, HDAC1
• Tissue Remodeling: SMAD2, SMAD3 (TGF-β pathway)

---

🔧 Repository Structure

---

🚀 Quick Start

Prerequisites

# Python 3.11 environment
conda create -n taco python=3.11.11
conda activate taco

# Install dependencies
pip install -r requirements.txt

# Specifically for TACO
pip install tabpfn==2.0.9  # Pin version for reproducibility

▶️ Usage from within TACO.ipynb

The DCE and ML Models Pipelines with all the functions live in the notebook:

jupyter notebook notebooks/TACO.ipynb

---

📝 Notes

• All file paths are relative to your notebook’s working directory.
• Precision & ROC-AUC are the primary metrics for clinical relevance.

📊 Data Requirements

RNA-seq Data

• Download from GEO: GSE215865
• 1,392 samples (303 Long COVID cases + 186 controls)

KEGG Pathways

• Automatically downloaded via keggrest R package (see TACO.ipynb file)
• 348 human pathways for DCE analysis

---

📋 Citation

If you use TACO in your research, please cite:

@inproceedings{taco2025,
  title={TACO: TabPFN-Augmented Causal Outcomes for Early Detection of Long COVID},
  author={Authors},
  booktitle={Australasian Data Science and Machine Learning Conference},
  year={2025}
}

---

🤝 Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Areas for Extension

• Integration with additional Outcomes modalities
• Application to other post-viral syndromes
• Real-time clinical deployment pipeline
• Multi-cohort validation studies

---

📄 License

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

---

🙏 Acknowledgments

• DCE developers (Jablonski et al., 2022)
• TabPFN (Hollmann et al., 2023)
• Mount Sinai COVID-19 Biobank for RNA-seq data (Thompson et al., 2023)
• KEGG for pathway annotations (Kanehisa et al., 2000)

---

📧 Contact

For questions or collaborations:

• Open an issue on GitHub
• Email:

---

🔗 Links

• Paper preprint (Coming soon)
• Supplementary Materials
• DCE Documentation
• TabPFN Documentation
• Mount Sinai COVID-19 Biobank

---

Last Updated: August 2025 Version: 1.0.0