README.md

---
license: cc-by-4.0
language:
  - en
tags:
  - 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
---

# IDP Pathogenicity Model
**Mechanistically Informed Prediction of Pathogenicity for Mitochondrial Missense Variants**

> Harrizi S., Nait Irahal 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)

```text
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 you have the model checkpoint and either precomputed embeddings or an ESM-2 inference setup, you can obtain pathogenicity predictions like this:

```python
import joblib
import torch
import numpy as np
from fair_esm import pretrained

# Load the trained classifier
model = joblib.load("models/pathogenicity_classifier_v2.pkl")            # adjust path if necessary

# (Optional) Load ESM-2 if you need to compute embeddings yourself
esm_model, alphabet = pretrained.load_model_and_alphabet("esm2_t33_650M_UR50D")
esm_model = esm_model.eval()
if torch.cuda.is_available():
    esm_model = esm_model.cuda()

# ───────────────────────────────────────────────────────────────
# Example: predict for one variant
# ───────────────────────────────────────────────────────────────
# You must provide a (1, 173) array containing:
#   • 128 PCA-reduced ESM-2 embedding dimensions (global + local ±15 residues)
#   • 45 biophysical / mitochondrial-specific features
#
# (Full feature-extraction + inference pipeline coming soon)

features = np.random.rand(1, 173).astype(np.float32)   # ← REPLACE with your real features!

prob_pathogenic = model.predict_proba(features)[0, 1]
print(f"Predicted probability of being pathogenic: {prob_pathogenic:.4f}")
```

*A complete end-to-end inference script (from protein sequence + variant → ESM embedding → features → final prediction) is in preparation and will be added shortly.*

## Reproducing the Experiments

Run the scripts in this order (all located in the `scripts/` folder):

1. `data_download.py`
2. `build_clinvar_mito_dataset.py` + `build_mutation_dataset.py`
3. `phase1_freeze_and_classical_features.py`
4. `esm2_t33_650M_UR50D.py` (GPU recommended – ~2–4 hours runtime)
5. `final_mlp_embedding_model.py` + `lpocv_validation.py`

## Important Notes & Limitations

- The model has been specifically developed and optimized for missense variants in mitochondrial proteins — performance on nuclear-encoded proteins has not been assessed.
- AlphaMissense achieves higher AUC on broader variant sets (0.942 vs our 0.899), but our model remains very competitive while using far fewer parameters and explicitly incorporating mitochondrial biology knowledge.
- All performance metrics are derived from strict leave-protein-out cross-validation, giving a realistic estimate of generalization to completely unseen proteins.

## Citation

```bibtex
@article{harrizi2025mechanistically,
  title   = {Mechanistically informed machine learning for pathogenicity prediction of mitochondrial missense variants},
  author  = {Harrizi, Saad and Nait Irahal, Imane and Mostafa, Kabine and Arnoult, Damien},
  journal = {Bioinformatics},
  year    = {2025},
  note    = {Submitted}
}
```

## Authors & Contact

- **Damien Arnoult** (corresponding author) — INSERM UMR-S-MD 1193 & UMR-S 1124, Paris  
  ✉ damien.arnoult@inserm.fr
- **Saad Harrizi** (main developer) — Université Hassan II de Casablanca  
  ✉ saadharrizi0@gmail.com

*We welcome any feedback, questions, or collaboration proposals!*