RecoverProject

🧬 RECOVER Project — Long COVID Genomic Analysis

ML analysis of metabolomic and proteomic data to predict Long COVID — conducted at Laboratoire Jacques Corbeil / François Laviolette, Université Laval.

Sami Bahig, MD MSc

---

🎯 Research Context

Long COVID remains poorly understood — risk factors, biomarkers, and treatment options are largely unknown. The only current medical consensus: symptoms persisting more than 12 weeks after initial infection.

This project applies machine learning to metabolomic and proteomic datasets to:

• Identify molecular signatures of Long COVID
• Predict Long COVID status from biological data
• Assess the impact of initial symptoms (fever, fatigue, anosmia) on prediction accuracy

---

🗂️ Datasets & Analyses

Analysis 1 — Metabolomic (Multi-timepoint)

• Multiple ML models benchmarked on metabolomic data
• 80/20 bootstrapping + cross-validation

Analysis 2 — Metabolomic (D0 — 100 patients)

• Timepoint D0: 100 patients
• Focus on early-stage biomarker prediction

Datasets analyzed

Dataset	Description
Metabolomic	Metabolite profiles — Long COVID vs control
Proteomic	Protein expression profiles
Proteomic.cyst	Cyst-specific proteomic subset
Proteomic.Merge	Combined proteomic dataset

---

📊 Results

Full Benchmark — 7 Models × 4 Datasets (80/20 Bootstrap + CV)

Dataset	Decision Tree	Random Forest	Random SCM	Ridge Classifier	SVM (Poly)	PLS-DA	SVM (rbf)
Proteomic — Train	90.9% ±7.4	88.2% ±8.5	91.2% ±14.6	99.9% ±0.4	86.5% ±14.0	58.0% ±2.6	67.5% ±22.1
Proteomic — Test	50.3% ±10.9	55.1% ±10.0	51.6% ±9.4	52.8% ±11.1	49.3% ±10.3	50.8% ±8.8	43.4% ±7.8
Metabolomic — Train	100% ±0.0	100% ±0.2	100% ±0.0	100% ±0.0	99.2% ±1.6	66.0% ±2.6	99.6% ±0.9
Metabolomic — Test	97.7% ±3.7	99.0% ±2.6	97.3% ±3.5	67.3% ±10.6	99.2% ±1.7	65.8% ±11.6	72.8% ±9.0
Proteomic.cyst — Train	87.1% ±8.1	93.8% ±8.1	76.1% ±10.0	88.3% ±2.9	85.3% ±7.3	53.9% ±2.8	61.8% ±19.1
Proteomic.cyst — Test	48.9% ±8.7	53.1% ±9.5	51.0% ±10.2	58.7% ±10.7	58.1% ±10.5	50.0% ±9.6	42.9% ±5.4

Effect of Initial Symptoms on Accuracy (Random Forest)

Symptom	Train Accuracy	Test Accuracy
Fever	89.8% ±5.3	64.0% ±11.9
Fatigue	90.7% ±4.5	63.3% ±12.3
Anosmia	90.1% ±4.6	63.8% ±12.1
Fever + Fatigue + Anosmia	90.9% ±3.5	63.6% ±12.7

Key Insight

Metabolomic data significantly outperforms proteomic data for Long COVID prediction — Random Forest and SVM(Poly) achieve ~99% test accuracy on metabolomic data, while proteomic models struggle to generalize (test accuracy ~50%). This suggests metabolomic signatures are stronger biomarkers for Long COVID status.

---

🛠️ Models Benchmarked

Model	Notes
Decision Tree	Baseline interpretable model
Random Forest	Best overall on metabolomic data
Random SCM	Set Covering Machine — rule-based
Ridge Classifier	Linear baseline
SVM (Poly + RBF)	Strong on metabolomic data
PLS-DA	Partial Least Squares Discriminant Analysis
ElasticNet	Regularized linear model

---

🗂️ Repository Structure

RecoverProject/
│
├── 📓 Copy_of_PLS_DA_Metabolomique.ipynb
├── 📓 DecisionTreeMetabolomic*.ipynb
├── 📓 DecisionTreeProteomic.ipynb
├── 📓 ElasticNet.ipynb
├── 📓 MergingDataset.ipynb
├── 📓 MergingDataset_Visualisation.ipynb
├── 📓 PLS_DA_Metabolomique.ipynb
├── 📓 RandomForest_Metabolomic.ipynb
├── 📓 RandomForest_Proteomic*.ipynb
├── 📓 RandomSCM*.ipynb
├── 📓 RidgeClassifier*.ipynb
├── 📓 svm*.ipynb
├── 📓 Recover_CleaningData_code.ipynb  ← Data preprocessing
└── 📖 README.md

---

🔮 Next Steps (as of presentation)

• AlphaFold 2 integration for protein structure analysis
• Graph Neural Networks for molecular interaction modeling
• Biobank data integration — larger cohort incoming

---

👤 Author

Sami Bahig, MD MSc — Data Scientist & AI Engineer Laboratoire Jacques Corbeil / François Laviolette — Université Laval

---

🔗 Related Projects

Project	Description	Link
FAERS 2025	FDA pharmacovigilance — 28M adverse event records	GitHub
Protocol TDM	OCR + CamemBERT — Radiology protocol classification	GitHub
RAG Chatbot	AI chatbot grounded in documents — LangChain · ChromaDB	GitHub

---

MIT License · Sami Bahig · Université Laval · 2022