A machine learning approach for classifying gene families from raw DNA sequences using k-mer counting and natural language processing techniques.

This project demonstrates how to apply machine learning techniques to genomic data by treating DNA sequences as a "language" and using k-mer counting with bag-of-words methodology. The model predicts gene families from DNA sequences and validates its effectiveness across different species.

• Multiple Encoding Methods: Implements three different DNA sequence encoding approaches
  • Ordinal encoding
  • One-hot encoding
  • K-mer counting (bag-of-words)
• Cross-Species Validation: Tests model generalizability on human, chimpanzee, and dog genomes
• High Accuracy: Achieves 94% accuracy on human DNA classification
• Biological Sequence Processing: Uses Biopython for efficient DNA sequence handling

The project uses DNA sequence data from three species:

• Human: 4,380 sequences across 7 gene families
• Chimpanzee: 1,682 sequences for cross-species validation
• Dog: 820 sequences for divergent species testing

1. G protein-coupled receptors
2. Tyrosine kinases
3. Tyrosine phosphatases
4. Synthetases
5. Synthases
6. Ion channels
7. Transcription factors

The project treats DNA sequences as natural language by:

1. Breaking sequences into overlapping k-mers (default k=6)
2. Creating "sentences" from k-mer "words"
3. Applying bag-of-words vectorization
4. Training a Multinomial Naive Bayes classifier

DNA: "ATGCATGCA" → K-mers: ['ATGCAT', 'TGCATG', 'GCATGC', 'CATGCA']

biopython numpy pandas scikit-learn matplotlib

The model demonstrates strong performance on human data and maintains good generalizability to closely related species (chimpanzee), with expected performance degradation on more evolutionarily distant species (dog).

• Kmers_funct(): Generates overlapping k-mers from DNA sequences
• string_to_array(): Converts sequence strings to arrays
• ordinal_encoder(): Implements ordinal encoding (A=0.25, C=0.50, G=0.75, T=1.00)
• one_hot_encoder(): Creates one-hot encoded representations

• Algorithm: Multinomial Naive Bayes
• Features: 4-gram k-mer counts (vocabulary size varies by dataset)
• Alpha: 0.1 (Laplace smoothing parameter)
• Train/Test Split: 80/20 on human data

This project demonstrates that:

• DNA sequences can be effectively processed using NLP techniques
• K-mer counting captures important biological patterns
• Machine learning models can predict gene function from sequence alone
• Cross-species classification reveals evolutionary relationships