Update README.md

README.md

@@ -1,166 +1,124 @@
 ---
 license: cc-by-4.0
 language:
-- en
+  - en
 tags:
-- mitochondrial-variants
-- pathogenicity-prediction
-- protein-language-model
-- ESM-2
-- machine-learning
-pretty_name: IDP Pathogenicity Model
+  - mitochondrial-variants
+  - pathogenicity-prediction
+  - protein-language-model
+  - ESM-2
+  - machine-learning
+  - bioinformatics
+pretty_name: IDP Pathogenicity Model – Mitochondrial Missense Variants
 size_categories:
-- 10K<n<100K
+  - 10K < n < 100K
 ---
-Mechanistically Informed Machine Learning for Pathogenicity Prediction of Mitochondrial Missense Variants
-
-Harrizi S., Nait Idraha I., Mostafa K., Arnoult D.
-Submitted to Bioinformatics (Oxford University Press)
-
-
-Overview
-This repository contains the full analysis pipeline for a hybrid machine learning framework that predicts the pathogenicity of missense variants in mitochondrial proteins. The model integrates ESM-2 protein language model embeddings with 45 interpretable biophysical features tailored to mitochondrial biology, trained and validated on 11,928 ClinVar-annotated variants spanning 639 mitochondrial proteins.
-Key results:
-ModelAUC–ROCAUC–PROur model (MLP + ESM-2)0.8990.923Random Forest + ESM-20.882—Logistic Regression + ESM-20.864—AlphaMissense0.9420.954PolyPhen-20.8450.799SIFT0.8260.749
-All models evaluated under leave-protein-out cross-validation on the same variant set.
-
-Repository Structure
-IDP/
-│
-├── data/
-│   ├── raw/                        # Downloaded source data
-│   │   ├── clinvar/                # ClinVar variant annotations (release 2024/12)
-│   │   ├── uniprot/                # UniProt mitochondrial protein sequences
-│   │   ├── disprot/                # DisProt intrinsic disorder annotations
-│   │   ├── mobidb/                 # MobiDB disorder predictions
-│   │   ├── alphafold/              # AlphaFold structure data
-│   │   ├── dms/                    # Deep mutational scanning data
-│   │   └── llpsdb/                 # Liquid-liquid phase separation database
-│   │
-│   ├── processed/                  # Cleaned and merged datasets
-│   │   ├── mutation_dataset.parquet
-│   │   ├── mutation_features.parquet
-│   │   ├── mutations_dataset_final.parquet / .tsv
-│   │   └── mutations_dataset_mito_strict.parquet / .tsv
-│   │
-│   └── frozen/                     # MD5-versioned frozen dataset for reproducibility
-│       ├── mutations_dataset_final_FROZEN.parquet
-│       ├── mutations_dataset_mito_strict_FROZEN.parquet
-│       └── FREEZE_METADATA.json
-│
-├── embeddings/                     # ESM-2 embeddings (precomputed)
-│   ├── embeddings_combined_full.npy      # Global + local combined (full dataset)
-│   ├── embeddings_global_full.npy        # Global protein embeddings
-│   ├── embeddings_local_full.npy         # Local ±15-residue window embeddings
-│   ├── embeddings_combined_strict.npy    # Strict mitochondrial subset
-│   └── embeddings_by_protein.pkl         # Per-protein embedding dictionary
-│
-├── features/                       # Extracted biophysical features
-│   ├── features_classical_full.parquet
-│   └── features_classical_mito_strict.parquet
-│
-├── models/                         # Trained model checkpoints
-│   ├── checkpoints/
-│   │   ├── pathogenicity_classifier.pkl
-│   │   └── pathogenicity_classifier_v2.pkl
-│   ├── model_classical_baseline.pkl
-│   └── model_classical_mito_strict.pkl
-│
-├── results/                        # Outputs and evaluations
-│   ├── evaluations/
-│   │   └── lpocv_results.json
-│   ├── lpocv_predictions_final.parquet
-│   ├── hierarchical_validation_final.pkl
-│   ├── full_analysis_with_ensemble.parquet
-│   └── feature_importances_classical.csv
-│
-├── scripts/                        # Python analysis scripts
-│   ├── data_download.py
-│   ├── build_clinvar_mito_dataset.py
-│   ├── build_mutation_dataset.py
-│   ├── phase1_freeze_and_classical_features.py
-│   ├── esm2_t33_650M_UR50D.py
-│   ├── train_baseline_classical_model.py
-│   ├── final_mlp_embedding_model.py
-│   ├── lpocv_validation.py
-│   ├── hierarchical_validation_no_leakage.py
-│   ├── model_comparison_features_vs_esm2.py
-│   └── analyze_bias_and_train_strict_mito.py
-│
-└── configs/                        # Configuration files
-
-Installation
-bashgit clone https://github.com/YOUR_USERNAME/IDP-Pathogenicity-Model.git
-cd IDP-Pathogenicity-Model
-
-pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
-pip install fair-esm
-pip install numpy scipy pandas scikit-learn statsmodels
-pip install biopython matplotlib seaborn plotly
-pip install einops tqdm joblib requests pyyaml networkx pyarrow
-Python version: 3.10
-Key dependencies: PyTorch 2.0 · scikit-learn 1.3 · NumPy 1.24 · pandas 2.0 · fair-esm
-GPU: NVIDIA A100 recommended (Google Colab Pro+ used for training)
-
-Reproducing the Analysis
-1. Download raw data
-bashpython scripts/data_download.py
-Downloads ClinVar (release 2024/12), UniProt mitochondrial proteins, DisProt, and MobiDB. Expected runtime: ~30 minutes.
-2. Build the variant dataset
-bashpython scripts/build_clinvar_mito_dataset.py
-python scripts/build_mutation_dataset.py
-Filters variants to mitochondrial proteins using MitoCarta 3.0 and OMIM annotations.
-3. Extract biophysical features
-bashpython scripts/phase1_freeze_and_classical_features.py
-Computes 45 biophysical features per variant.
-4. Extract ESM-2 embeddings
-bashpython scripts/esm2_t33_650M_UR50D.py
-Generates global and local (±15-residue window) embeddings using esm2_t33_650M_UR50D. For proteins longer than 1,022 residues, embeddings are computed on the N-terminal portion. Requires GPU. Expected runtime: ~2–4 hours.
-5. Train and validate the model
-bashpython scripts/final_mlp_embedding_model.py
-python scripts/lpocv_validation.py
-Trains the MLP classifier and runs leave-protein-out cross-validation. StandardScaler and PCA are fitted exclusively on the training fold — no test-set leakage.
-6. Baseline model comparison
-bashpython scripts/model_comparison_features_vs_esm2.py
-python scripts/train_baseline_classical_model.py
-Evaluates logistic regression and random forest baselines on the same feature set under the same LPOCV framework.
-
-Model Architecture
-ComponentDetailsInput features173 (128 PCA-reduced embedding dimensions + 45 biophysical)ArchitectureMLP: Input → 256 → 128 → 1 (sigmoid)ActivationReLU + Dropout (0.3, 0.2)OptimizerAdam (lr = 0.001, weight decay = 1×10⁻⁴)LossBinary cross-entropyEpochs30–50ValidationLeave-protein-out cross-validationPreprocessingStandardScaler + PCA fitted on training fold only
-Baseline comparison
-ModelAUC–ROCLogistic Regression + ESM-20.864Random Forest + ESM-20.882MLP + ESM-2 (proposed)0.899
-
-Dataset
-PropertyValueTotal variants11,928Pathogenic / likely pathogenic6,882 (57.7%)Benign / likely benign5,046 (42.3%)Proteins639SourceClinVar release 2024/12Protein sequencesUniProt release 2024_05Mitochondrial annotationMitoCarta 3.0 + OMIM
-The frozen dataset (data/frozen/) is MD5-checksummed for reproducibility.
-
-Benchmarking
-ToolN variantsAUC–ROCAUC–PROur model (MLP + ESM-2)11,7630.8990.923AlphaMissense11,2320.9420.954PolyPhen-27,3040.8450.799SIFT6,9970.8260.749
-AlphaMissense scores retrieved from the public bulk file (AlphaMissense_hg38.tsv.gz). PolyPhen-2 and SIFT scores retrieved via the UniProt Proteins variation API.
-
-Data Availability
-Raw data sources:
-
-ClinVar: https://www.ncbi.nlm.nih.gov/clinvar/
-UniProt: https://www.uniprot.org/
-MitoCarta 3.0: https://www.broadinstitute.org/mitocarta
-DisProt: https://disprot.org/
-MobiDB: https://mobidb.org/
-AlphaMissense scores: https://storage.googleapis.com/dm_alphamissense/AlphaMissense_hg38.tsv.gz
-
-Processed datasets and model checkpoints are available upon reasonable request to the corresponding author.
+
+# IDP Pathogenicity Model  
+**Mechanistically Informed Prediction of Pathogenicity for Mitochondrial Missense Variants**
+
+> Harrizi S., Nait Idraha I., Mostafa K., Arnoult D.  
+> Submitted to *Bioinformatics* (2025)
+
+This repository / model card describes a hybrid machine-learning classifier that predicts the pathogenicity of **missense variants** in **mitochondrial proteins**.  
+
+The model combines **ESM-2** (t33_650M_UR50D) protein language model embeddings with 45 carefully designed **biophysical and mitochondrial-specific features**, and is trained under strict **leave-protein-out cross-validation** to minimize data leakage.
+
+## Key Performance (Leave-Protein-Out CV)
+
+| Model                              | AUC–ROC | AUC–PR |
+|------------------------------------|---------|--------|
+| **Our model (MLP + ESM-2 + features)** | **0.899** | **0.923** |
+| Random Forest + ESM-2              | 0.882   | —      |
+| Logistic Regression + ESM-2        | 0.864   | —      |
+| AlphaMissense                      | 0.942   | 0.954  |
+| PolyPhen-2                         | 0.845   | 0.799  |
+| SIFT                               | 0.826   | 0.749  |
+
+*All tools evaluated on the same ClinVar mitochondrial missense set under leave-protein-out protocol.*
+
+## Model Inputs
+
+- **ESM-2 embeddings** — global protein embedding + local window (±15 residues)  
+- **45 biophysical features** — conservation, disorder (DisProt/MobiDB), AlphaFold pLDDT, charge changes, hydrophobicity, mitochondrial-specific annotations, etc.
+
+Final input dimensionality after PCA: **173 features** (128 PCA-reduced ESM dimensions + 45 classical features)
+
+## Architecture
+
+Multi-Layer Perceptron (MLP)
+
+| Layer          | Units | Activation | Regularization     |
+|----------------|-------|------------|--------------------|
+| Input          | 173   | —          | —                  |
+| Hidden 1       | 256   | ReLU       | Dropout 0.3        |
+| Hidden 2       | 128   | ReLU       | Dropout 0.2        |
+| Output         | 1     | Sigmoid    | —                  |
+
+- Optimizer: Adam (lr=0.001, weight decay=1e-4)  
+- Loss: Binary Cross-Entropy  
+- Preprocessing: StandardScaler + PCA (fitted only on train fold)  
+- Training epochs: 30–50 (early stopping used)
+
+## Dataset Summary
+
+- **Total variants**: 11,928 ClinVar missense variants (release 2024/12)  
+- **Pathogenic / likely pathogenic**: 6,882 (57.7%)  
+- **Benign / likely benign**: 5,046 (42.3%)  
+- **Mitochondrial proteins**: 639 (MitoCarta 3.0 + OMIM filtering)  
+- Sequence source: UniProt 2024_05  
+- Frozen & MD5-versioned dataset available in `data/frozen/`
+
+## Repository Structure (main folders)
+IDP-Pathogenicity-Model/
+├── data/               # raw, processed & frozen datasets
+├── embeddings/         # precomputed ESM-2 embeddings (.npy, .pkl)
+├── features/           # extracted biophysical features
+├── models/             # trained checkpoints (joblib/pkl)
+├── results/            # LPOCV predictions, feature importances, etc.
+├── scripts/            # full reproducible pipeline
+└── configs/            # YAML configuration files
+
+## Quick Start – Inference (once model is uploaded)
+
+```python
+import joblib
+import torch
+from fair_esm import pretrained
+
+model = joblib.load("pathogenicity_classifier_v2.pkl")
+esm_model, alphabet = pretrained.load_model_and_alphabet("esm2_t33_650M_UR50D")
+
+# Example: get prediction for one variant
+# (you need to prepare the 173-dim feature vector yourself)
+prob = model.predict_proba(features)[0, 1]   # pathogenic probability
+Full inference pipeline coming soon (feature extraction + ESM inference + prediction in one script).
+Reproducing the Experiments
+See detailed steps in the original README:
+
+data_download.py
+build_clinvar_mito_dataset.py + build_mutation_dataset.py
+phase1_freeze_and_classical_features.py
+esm2_t33_650M_UR50D.py (GPU heavy – ~2–4 h)
+final_mlp_embedding_model.py + lpocv_validation.py
+
+Limitations & Notes
+
+Model optimized specifically for mitochondrial proteins — performance on nuclear proteins is not guaranteed
+AlphaMissense remains stronger on general proteins (0.942 vs 0.899), but our model is competitive while using far fewer parameters and incorporating explicit mitochondrial biology
+Strict leave-protein-out CV was used → realistic generalization to unseen proteins
 
 Citation
-Harrizi S., Nait Idraha I., Mostafa K., Arnoult D. (2025).
-Mechanistically informed machine learning for pathogenicity prediction
-of mitochondrial missense variants. Bioinformatics (submitted).
-
-Authors
-Saad Harrizi — Laboratoire Santé, Environnement et Biotechnologie, Faculté Des Sciences Ain Chock, Université Hassan II de Casablanca, Maroc.
-Imane Nait Idraha — Laboratoire Santé, Environnement et Biotechnologie, Université Hassan II de Casablanca, Maroc.
-Kabine Mostafa — Laboratoire Santé, Environnement et Biotechnologie, Université Hassan II de Casablanca, Maroc.
-Damien Arnoult (Corresponding author) — INSERM UMR-S-MD 1193, Université Paris Saclay, Villejuif, France; INSERM-U1124, Université Paris Cité, Paris, France; CNRS Research Director (DR2).
-✉ damien.arnoult@inserm.fr
-
-License
-This repository is released for academic and research use. Please contact the corresponding author for commercial use inquiries.
+bibtex@article{harrizi2025mechanistically,
+  title   = {Mechanistically informed machine learning for pathogenicity prediction of mitochondrial missense variants},
+  author  = {Harrizi, Saad and Nait Idraha, Imane and Mostafa, Kabine and Arnoult, Damien},
+  journal = {Bioinformatics},
+  year    = {2025},
+  note    = {Submitted}
+}
+Authors & Contact
+
+Saad Harrizi (main developer) – Université Hassan II de Casablanca – saadharrizi0@gmail.com 
+Damien Arnoult (corresponding) – INSERM, Paris –  damien.arnoult@inserm.fr
+
+We welcome feedback, questions, and potential collaborations.