RNN-based LSTM

Overview

A Long Short-Term Memory (LSTM) network, a type of recurrent neural network (RNN), is designed to efficiently process and analyze sequential data. This model was developed to predict the next emerging variant in viral genomes, leveraging its ability to capture complex dependencies in time-series data. By training an LSTM model on historical genomic sequences of a virus variants, and their patterns of emergence, the model can learn to identify subtle patterns and trends to predict potential future variants. Additionally, LSTM is capable of handling variable-length sequences, which is crucial given the variability in viral sequences across different strains and variants.

Features

This software processes and analyzes sequences using an LSTM (Long Short-Term Memory) model. It follows a structured pipeline for sequence alignment, padding, encoding, window generation, hyperparameter tuning, and model training.

• Loads and process DNA sequences from FASTA files.
• Performs sequence padding for uniform lengths.
• One-hot encode DNA sequences (A, C, T, G, -).
• Generates training windows for model input.
• Performs hyperparameter tuning for optimal LSTM training.
• Trains an LSTM model to predict future mutations.

Project Structure

📂 project_root/
├── 📂 modules/
│   ├── data_loader.py  # Load sequences from FASTA files
│   ├── sequence_padder.py  # Pad sequences to match reference length
│   ├── encoder_decoder.py  # One-hot encode DNA sequences
│   ├── window_generator.py  # Generate training windows
│   ├── encoder_decoder.py  # Encode sequences into numerical format
├── 📂 lstm_tuner/
│   ├── tuner.py  # Hyperparameter tuning
│   ├── model_builder.py  # Build LSTM model
├── 📂 model/  # model's main script
├── main.py  # Main script to run the pipeline
├── README.md  # Project documentation

Usage

python main.py --ref_path path/to/reference.fasta --mut_path path/to/mutated.fasta --num_mut_seqs --out_dir ./output

Arguments

Argument	Description
--ref_path	Path to reference sequence (FASTA format).
--mut_path	Path to mutated sequences (FASTA format).
--num_mut_seqs	(Optional) Number of mutated sequences to process.
--out_dir	Output directory for aligned sequences. Default: ./output

Pipeline Overview

• Load Sequences: Reads FASTA files and extracts sequences.
• Sequence Padding: Ensures all sequences are of uniform length.
• One-Hot Encoding: Converts sequences into numerical format.
• Reshaping Data: Prepares the encoded sequences for model input.
• Window Generation: Creates training windows for sequence prediction.
• Dataset Batching: Splits data into training and validation batches.
• Hyperparameter Tuning: Optimizes LSTM model settings.
• Model Training: Trains the LSTM model for sequence prediction.

Output

After successful execution, the output includes:

• Trained LSTM Model.
• Logs & Metrics