# MutationPredictor-CNN v4

## Overview

MutationPredictor-CNN v4 is a sequence-based convolutional neural network designed to predict pathogenicity of genomic variants with splice-awareness.

The model uses:

- 401 bp genomic window (GRCh38)
- One-hot encoded forward and reverse complement sequence
- Positional encoding
- Canonical splice motif features (GT/AG)
- Explicit mutation type encoding (12 SNV classes)
- Auxiliary region and splice feature embeddings

This model is trained on ClinVar-derived labeled splice-impact variants.

---

## Architecture

Input Features:
- 11-channel sequence encoding
- 12-dim mutation one-hot encoding
- 2-dim region embedding
- 3-dim splice embedding

Model:

Conv1d(11 → 64, kernel=7)  
Conv1d(64 → 128, kernel=5)  
Conv1d(128 → 256, kernel=3)  
AdaptiveAvgPool1d(1)

Fully connected layers:
- 312 → 128
- 128 → 64
- 64 → 1 (logit output)

Activation: ReLU  
Regularization: Dropout(0.3)  
Loss: BCEWithLogitsLoss  

---

## Training Details

- Genome build: GRCh38
- Window size: 401 bp
- Optimizer: Adam
- Learning rate: 0.001
- Epochs: 30
- Batch size: 256

Best model selected by highest training AUC.

---

## Evaluation

### Internal (ClinVar-derived)
AUC ≈ 0.97

### External Benchmark – SpliceAI Comparison (1000 variants)
AUC ≈ 0.91

### Strict Fair Benchmark (134 variants not seen in training context)
AUC ≈ 0.54

### Fair Benchmark by Consequence

| Variant Type        | Count | AUC  |
|---------------------|-------|------|
| Splice Donor        | 37    | 0.53 |
| Splice Acceptor     | 34    | 0.51 |
| Other Consequences  | 63    | 0.61 |

---

## Known Limitations

- Trained primarily on ClinVar-labeled variants
- Limited indel generalization
- 401 bp receptive field (short-range context)
- Not validated for clinical decision-making
- Research-use only

---

## Intended Use

This model is intended as:

- Research tool
- Variant prioritization support
- Experimental splice-aware classifier

Not intended for direct clinical decision-making.

---

## Future Directions

- Expand receptive field (>2kb)
- Dilated convolutions
- Transformer-based architecture
- Multi-task pathogenicity + splice scoring
- Larger independent external validation

---

## License

Research use only.

---

## Author

Nilesh Hanotia