Parkinson's Disease Protein Sequence Classifier

Overview

A machine learning pipeline for classifying protein sequences associated with Parkinson's disease using various sequence features and multiple classification models. The project achieves 80.3% accuracy using LSTM architecture with comprehensive sequence feature analysis.

Table of Contents

• Installation
• Dataset
• Pipeline Architecture
• Results
• Technical Details
• Future Work

Installation and Dependencies

Prerequisites

• Python 3.11+
• CUDA capable GPU (optional, for faster training)

Quick Start

git clone https://github.com/Spritan/ParkinsonDiseaseClassifier
cd ParkinsonDiseaseClassifier
pip install uv
uv run main.py

Dependencies

dependencies = [
    "biopython>=1.84",
    "matplotlib>=3.9.2",
    "pandas>=2.2.3",
    "rich>=13.9.4",
    "scikit-learn>=1.5.2",
    "seaborn>=0.13.2",
    "torch>=2.5.1",
    "xgboost>=2.1.2",
]

Dataset

Data Format (FASTA)

FASTA files contain protein sequences in a text-based format:

>sp|Q9Y6M9|NDUB9_HUMAN NADH dehydrogenase [ubiquinone] 1 beta subcomplex subunit 9
MAFLASGPYLTHQQKVLRLYKRALRHLESWCVQRDKYRYFACLMRARFEEHKNEKDMAKA
TQLLKEAEEEFWYRQHPQPYIFPDSPGGTSYERYDCYKVPEWCLDDWHPSEKAMYPDYFA
KREQWKKLRRESWEREVKQLQEETPPGGPLTEALPPARKEGDLPPLWWYIVTRPRERPM

• Header line (starts with '>'): Contains sequence identifier and metadata
• Subsequent lines: Amino acid sequence in single-letter code

Data Sources

1. Control Sequences (129 samples)

   • Source: data/cleaned_output (1).fasta
   • Average length: 482 amino acids
   • Contains various human proteins without PD association

2. Parkinson's Disease Sequences (165 samples)

   • Source: data/uniprotkb_parkinson_disease_protein_AND_2024_11_15.fasta
   • Average length: 523 amino acids
   • Proteins associated with PD pathology

Sequence Analysis

1. Sequence Properties Distribution

• Length Distribution: Right-skewed, ranging from 100 to 2000+ amino acids
• GC Content: Normal distribution centered around 0.52
• Hydrophobic Ratio: Peaks at 0.40, indicating typical protein composition
• Charged Ratio: Centered at 0.25, showing consistent charge distribution

2. Amino Acid Composition

• Most Abundant:
  • Leucine (L): ~10%
  • Alanine (A): ~7.8%
  • Serine (S): ~7.2%
• Least Abundant:
  • Tryptophan (W): ~1.4%
  • Cysteine (C): ~2.2%
  • Histidine (H): ~2.4%
• Class Differences:
  • Slight variations in Serine (S) content between healthy and PD
  • Minor differences in charged amino acids (R, K, D, E)

3. K-mer Patterns

Bi-mer Analysis:

Top 5 Most Frequent:
1. LE (7.2%): Leucine-Glutamic acid
2. AL (6.8%): Alanine-Leucine
3. LA (6.5%): Leucine-Alanine
4. EL (5.9%): Glutamic acid-Leucine
5. SE (5.7%): Serine-Glutamic acid

Tri-mer Analysis:

Top 5 Most Frequent:
1. LEE (3.1%): Leucine-Glutamic acid-Glutamic acid
2. ALA (2.9%): Alanine-Leucine-Alanine
3. LAA (2.8%): Leucine-Alanine-Alanine
4. ELL (2.7%): Glutamic acid-Leucine-Leucine
5. AAL (2.6%): Alanine-Alanine-Leucine

4. Feature Correlations

• Strong correlations between related k-mers
• Moderate correlations between physicochemical properties
• Weak correlations between different feature types

Data Quality

• No missing values
• Valid amino acid sequences (20 standard amino acids)
• Balanced class distribution (43.88% healthy, 56.12% PD)
• Consistent sequence format and annotation

Preprocessing Steps

1. Sequence validation
2. Length normalization
3. Feature extraction
4. Feature scaling
5. Train-test split (80-20)

Pipeline Architecture

1. Data Processing

def load_fasta_data(healthy_path, parkinsons_path):
    """
    Load and process FASTA sequences
    Returns DataFrame with sequences and labels
    """

• FASTA file parsing using BioPython
• Sequence validation and preprocessing
• Data frame construction with sequence metadata

2. Feature Engineering (More details later on ...)

Features extracted include:

• K-mer frequencies (k=2,3)
• Basic sequence properties:
  • Length
  • GC content
  • Hydrophobic ratio
  • Charged amino acid ratio
• Total features extracted: 7,950
• Features selected for modeling: 50

3. Model Training

Multiple models evaluated:

• Deep Learning:
  • LSTM
  • Neural Network
• Traditional ML:
  • SVM
  • Random Forest
  • Gradient Boosting
  • XGBoost
  • KNN

Feature Engineering and Selection

Feature Extraction Pipeline

The feature extraction process is handled by the SequenceFeatureExtractor class (reference: src/feature_extraction.py, lines 9-48), which implements a comprehensive approach to protein sequence analysis:

class SequenceFeatureExtractor:
    def __init__(self, k=3):
        self.k = k
        self.vectorizer = CountVectorizer(analyzer='char', ngram_range=(k, k))
        
    def compute_basic_features(self, sequence):
        return {
            'length': len(sequence),
            'gc_content': (sequence.count('G') + sequence.count('C')) / len(sequence),
            'hydrophobic_ratio': sum(aa in 'AILMFWYV' for aa in sequence) / len(sequence),
            'charged_ratio': sum(aa in 'DEKR' for aa in sequence) / len(sequence)
        }

Feature Categories

1. Sequence-based Features

• Length: Raw sequence length
• GC Content: Proportion of Glycine and Cytosine
• Hydrophobic Ratio: Proportion of hydrophobic amino acids (A, I, L, M, F, W, Y, V)
• Charged Ratio: Proportion of charged amino acids (D, E, K, R)

2. K-mer Features

The system uses CountVectorizer to generate:

• Bi-mers (k=2): Captures local sequence patterns
• Tri-mers (k=3): Captures extended sequence motifs

Feature Selection Process

Feature selection is implemented in SequenceDataAnalyzer (reference: src/data_analysis.py, lines 81-97) using a multi-method approach:

def select_features(self, n_features=50):
    """
    Select top features using ensemble of methods:
    1. Random Forest importance
    2. ANOVA F-scores
    3. Mutual Information
    """

Selection Methods

1. Random Forest Importance

   • Evaluates feature importance through decision trees
   • Particularly effective for capturing non-linear relationships
   • Robust to feature scaling

2. ANOVA F-scores

   • Statistical test for feature relevance
   • Identifies features with significant class separation
   • Assumes normal distribution

3. Mutual Information

   • Information theory-based approach
   • Captures non-linear relationships
   • Scale-invariant feature selection

Feature Selection Workflow

1. Initial Feature Space

   • Total features extracted: 7,950
   • Includes all k-mers and basic properties
   • Sparse matrix representation for efficiency

2. Feature Ranking

   • Each method ranks features independently
   • Scores are normalized to [0,1] range
   • Combined ranking through weighted voting

3. Feature Reduction

   • Final selection: Top 50 features
   • Balanced representation across feature types
   • Optimized for model performance

Feature Importance Analysis

The system provides comprehensive feature importance visualization through multiple plots:

1. Correlation Analysis

startLine: 103
endLine: 127

2. Feature Importance Plotting

startLine: 128
endLine: 159

Performance Impact

Feature selection significantly improves model performance:

1. Computational Efficiency

   • 99.4% reduction in feature space
   • Faster training times
   • Reduced memory requirements

2. Model Performance

   • Improved generalization
   • Reduced overfitting
   • More interpretable models

3. Cross-validation Results

   • Selected features maintain 80.3% accuracy
   • Stable performance across folds
   • Robust to different model architectures

Future Enhancements

1. Advanced Feature Engineering

   • Position-specific scoring matrices
   • Secondary structure predictions
   • Evolutionary conservation scores
   • Domain-specific features

2. Selection Methods

   • Deep learning-based feature selection
   • Ensemble selection strategies
   • Dynamic feature importance tracking

3. Optimization

   • Feature selection hyperparameter tuning
   • Custom feature importance metrics
   • Real-time feature importance updates

Results

Detailed Model Performance Analysis

1. LSTM (Best Model)

┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┓
┃ Metric    ┃ Score (mean ± std) ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━┩
│ Accuracy  │      0.803 ± 0.046 │
│ Precision │      0.816 ± 0.078 │
│ Recall    │      0.845 ± 0.095 │
│ F1        │      0.825 ± 0.054 │
└───────────┴────────────────────┘

• Best overall performance
• Balanced precision and recall
• Consistent performance across classes

2. Neural Network

┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━┓
┃ Metric    ┃ Score (mean ± std) ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━┩
│ Accuracy  │      0.718 ± 0.028 │
│ Precision │      0.718 ± 0.065 │
│ Recall    │      0.830 ± 0.094 │
│ F1        │      0.764 ± 0.042 │
└───────────┴────────────────────┘

• Second-best performer
• High recall but lower precision
• More stable metrics (lower std)

3. Traditional ML Models Comparison

Model              F1-Score        Accuracy
----------------------------------------
Gradient Boosting  0.714 ± 0.048  0.656 ± 0.040
Random Forest      0.690 ± 0.043  0.639 ± 0.036
XGBoost           0.686 ± 0.040  0.636 ± 0.044
KNN               0.683 ± 0.034  0.595 ± 0.039

Feature Analysis Results

1. Amino Acid Distribution

• Leucine (L): ~10% frequency in both classes
• Key differences in Serine (S) and Alanine (A)
• Rare amino acids: Cysteine (C), Tryptophan (W)

2. K-mer Analysis

Most significant patterns:

Bi-mers:  LE (7.2%), AL (6.8%), LA (6.5%)
Tri-mers: LEE (3.1%), ALA (2.9%), LAA (2.8%)

3. Feature Importance Rankings

Top features by method:

1. Random Forest:
   • Hydrophobic ratio
   • Length
   • GC content
2. ANOVA F-score:
   • K-mer patterns
   • Charged ratio
3. Mutual Information:
   • Sequence length
   • K-mer frequencies

Performance Visualization

• LSTM consistently outperforms other models
• Deep learning models show superior performance
• Traditional ML models cluster around 65-70% accuracy

Technical Implementation Details

Core Components

1. Data Processing Pipeline:

def process_sequences(sequences):
    """
    Process raw sequences:
    1. Validate amino acid content
    2. Extract features
    3. Normalize values
    """

2. Feature Extraction:

class SequenceFeatureExtractor:
    """
    Extract features from protein sequences:
    - K-mer frequencies
    - Sequence properties
    - Physicochemical properties
    """

3. Model Training:

class SequenceClassifier:
    """
    Train and evaluate multiple models:
    - Deep learning (LSTM, NN)
    - Traditional ML
    """

Performance Optimization

1. Feature Selection Process:

• Initial features: 7,950
• Selected features: 50
• Selection methods:
  • Random Forest importance
  • ANOVA F-scores
  • Mutual Information

2. Cross-validation Strategy:

• 5-fold stratified CV
• Consistent random seed
• Performance metrics with std

Future Work

1. Dataset Enhancement

• Target: 1000+ sequences per class
• Include:
  • More control sequences
  • Various PD subtypes
  • Related neurodegenerative diseases

2. Feature Engineering

• Position-specific scoring matrices
• Secondary structure predictions
• Domain-specific features
• Evolutionary conservation scores

3. Model Improvements

• Attention mechanisms
• Transformer architectures
• Ensemble strategies
• Model interpretability

4. Pipeline Optimization

• Model persistence
• Parallel processing
• Configuration management
• API development

Acknowledgments

• UniProt database for protein sequences
• BioPython community
• Rich library developers
• PyTorch team