PSEN1 Mutation Risk Prediction Pipeline

A lightweight computational pipeline for predicting pathogenic mutations in Presenilin-1 (PSEN1) and APP using ESM-2 protein language model, evolutionary features (BLOSUM62), and PolyPhen-2 labels.

Overview

This project addresses the challenge of predicting pathogenic mutations with limited labeled data. The pipeline combines:

• ESM-2: Pre-trained protein language model for mutation likelihood estimation (saturation mutagenesis)
• BLOSUM62: Evolutionary substitution matrix
• PolyPhen-2: Weak supervision labels for training
• Logistic Regression: Lightweight, interpretable risk prediction model

Project Structure

GenAIProject_PSEN1/
├── src/
│   ├── data/              # Data processing modules
│   │   ├── convert_polyphen.py   # Mutation format conversion for PolyPhen-2
│   │   └── tidy.py               # PolyPhen output tidying
│   ├── models/            # Model training and scoring
│   │   └── risk_model.py         # Risk prediction model (Logistic Regression)
│   └── visualization/     # Visualization tools
│       └── visualize_polyphen.py  # PolyPhen distribution plots
├── scripts/               # Main execution scripts
│   ├── run_esm2.py              # ESM-2 saturation mutagenesis
│   └── run_risk_models.py        # Train model and score variants
├── data/
│   ├── raw/               # Raw input data
│   │   ├── psen1_mutations_raw.csv
│   │   ├── app_mutations_raw.csv
│   │   ├── P49768.fasta (PSEN1)
│   │   └── P05067.fasta (APP)
│   ├── processed/         # Processed intermediate data
│   │   ├── polyphen2/     # PolyPhen-2 outputs
│   │   └── tidy/          # Tidy PolyPhen data
│   └── outputs/          # Results and model outputs
│       ├── esm2/          # ESM-2 generated mutations
│       ├── psen1/         # PSEN1 results
│       └── app/           # APP results
└── data_crawler.py        # Alzforum data crawler

Workflow

1. Data Collection (Optional)

Crawl mutation data from Alzforum:

cd GenAIProject_PSEN1
python data/data_crawler.py --protein psen1
python data/data_crawler.py --protein app

2. Convert Mutations for PolyPhen-2

Convert mutations to PolyPhen-2 batch input format:

python src/data/convert_polyphen.py \
    --protein psen1 \
    --input data/raw/psen1_mutations_raw.csv \
    --output data/processed/psen1_known_mutations_batch.txt

3. Generate Mutations with ESM-2

Run ESM-2 saturation mutagenesis (all possible mutations):

python scripts/run_esm2.py \
    --fasta data/raw/P49768.fasta \
    --protein PSEN1 \
    --output data/outputs/esm2/psen1_esm2_all.csv

This generates all possible mutations at every position with:

• delta_loglik: Change in log-likelihood (mut_logp - wt_logp)
• entropy: Position uncertainty
• wt_logp, mut_logp: Wild-type and mutant log-probabilities

4. Train Model and Score Variants

Train risk model and score all variants:

python scripts/run_risk_models.py

This script:

1. Loads ESM-2 generated mutations (PSEN1 and APP)
2. Adds BLOSUM62 scores
3. Loads PolyPhen-2 labels
4. Trains logistic regression model on PSEN1 data
5. Scores all PSEN1 variants
6. Applies PSEN1 model to APP variants (zero-shot transfer)
7. Generates risk distribution plots

Outputs:

• data/outputs/psen1/psen1_risk_scorer.joblib: Trained model
• data/outputs/psen1/psen1_scored_variants.csv: All PSEN1 variants with risk scores
• data/outputs/app/app_scored_variants_psen1_model.csv: APP variants scored with PSEN1 model
• data/outputs/*/psen1_risk_dist.png: Risk score distributions

5. Tidy PolyPhen Data (Optional)

Convert PolyPhen output to tidy format:

python src/data/tidy.py \
    --input data/processed/polyphen2/psen1_polyphen_scores.txt \
    --output data/processed/tidy/psen1_polyphen_tidy.csv

6. Visualization

Visualize PolyPhen-2 prediction distribution:

python src/visualization/visualize_polyphen.py \
    --input data/processed/polyphen2/psen1_polyphen_scores.txt \
    --output data/outputs/psen1/psen1_polyphen_distribution.png \
    --protein PSEN1

Key Features

Mutation Generation

• Saturation Mutagenesis: Generates all 19 possible mutations per position
• Uses ESM-2 (650M parameters) to compute mutation likelihoods
• Extracts position-level entropy for uncertainty quantification

Risk Scoring

• Features: delta_loglik, entropy, blosum62
• Model: Logistic Regression (lightweight, interpretable)
• Training: Weak supervision from PolyPhen-2 labels
• Output: Risk scores (0-1) for all candidate mutations
• Zero-shot Transfer: PSEN1-trained model applied to APP

Validation

• PolyPhen-2 predictions as weak supervision
• Known vs. novel variant separation
• Risk distribution visualization

Requirements

See requirements.txt for full dependencies. Key packages:

• pandas, numpy: Data manipulation
• scikit-learn: Machine learning
• biopython: Sequence analysis and BLOSUM62
• torch, fair-esm: ESM-2 model
• matplotlib, seaborn: Visualization
• requests, beautifulsoup4: Web scraping (for data crawler)

Install with:

pip install -r requirements.txt

Data Requirements

1. Raw mutation data:
   • data/raw/psen1_mutations_raw.csv
   • data/raw/app_mutations_raw.csv
2. Protein sequences:
   • data/raw/P49768.fasta (PSEN1, UniProt format)
   • data/raw/P05067.fasta (APP, UniProt format)
3. PolyPhen-2 outputs:
   • data/processed/polyphen2/psen1_polyphen_scores.txt
   • data/processed/polyphen2/app_polyphen_scores.txt

Results

The pipeline generates:

• Risk scores for all candidate mutations (saturation mutagenesis)
• Separation of known vs. novel variants
• High-risk variant prioritization
• Model performance metrics (AUC, accuracy)
• Zero-shot transfer results (PSEN1 model → APP)

Supported Proteins

• PSEN1 (Presenilin-1): Primary training target
• APP (Amyloid Precursor Protein): Zero-shot transfer target

Future Extensions

• Support for additional proteins
• AlphaFold structure-based filtering
• Ensemble methods for improved predictions
• Web interface for variant prioritization
• Cross-protein model transfer analysis

Citation

If you use this pipeline, please cite:

• ESM-2: Lin et al., 2022
• PolyPhen-2: Adzhubei et al., 2010