🧬 GENOME MINIMIZER 2

VAE-powered minimal genome generation pipeline for E. coli.

Pipeline Overview

Data Files → [Preprocess] → [Explore] → [Training] → [Sample] → [Minimize]

1. Preprocess: Extract essential gene positions from literature
2. Explore: Analyze dataset distributions and generate visualizations
3. Training: Train VAE models with different configurations
4. Sample: Generate synthetic genomes from trained models
5. Minimize: Create actual minimized genome sequences

Setup

Prerequisites

• Python >=3.9,<3.11
• pip (usually comes with Python)

Installation

1. Clone the repository:

   git clone <repository-url>
   cd genome-minimizer-2

2. Create virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install project:

   pip install -e .

Quick Start

python main.py --mode preprocess
python main.py --mode training --preset v0 --epochs 1
python main.py --mode sample --model-path models/trained_models/v0_model/saved_VAE_v0.pt --genes-path data/essential_genes/essential_gene_positions.pkl --num-samples 100
python main.py --mode minimizer --genes-path models/v0_model/sampling_results/binary_samples_default.npy --single-file --output-file results.fasta

Data Setup

Place these files in data/:

data/
├── F4_complete_presence_absence.csv    # Gene presence/absence matrix
├── accessionID_phylogroup_BD.csv       # Phylogroup classifications
├── essential_genes.csv                 # Essential genes from literature
└── wild_type_sequence.gb               # E. coli reference genome

Commands

Mode	Purpose	Example
preprocess	Extract essential gene positions	python main.py --mode preprocess
explore	Generate data analysis plots	python main.py --mode explore
training	Train VAE models	python main.py --mode training --preset v0
sample	Generate synthetic genomes	python main.py --mode sample --model-path MODEL.pt --genes-path GENES.pkl
minimizer	Create FASTA sequences	python main.py --mode minimizer --genes-path SAMPLES.npy --single-file

Parameters by Mode

Preprocess

python main.py --mode preprocess [--force-reprocess]

• --force-reprocess: Regenerate even if files exist

Training

python main.py --mode training --preset PRESET [--epochs N]

• --preset v0/v1/v2/v3: Model architecture (required)
• --epochs N: Training epochs (default: 10000)

Experiment

python main.py --mode experiment [--interactive]

• --interactive: Prompt for custom parameters

Sample

python main.py --mode sample --model-path PATH --genes-path PATH [OPTIONS]

Required:

• --model-path: Trained model (.pt file)
• --genes-path: Essential gene positions (.pkl file)

Optional:

• --num-samples N: Number of genomes (default: 1)
• --sampling-mode default/focused: Strategy (default: default)
• --noise-level N: Noise for focused sampling (default: 0.1)
• --genome-path: Reference genome (.gb file)

Minimizer

python main.py --mode minimizer --genes-path PATH [OPTIONS]

Required:

• --genes-path: Binary samples (.npy file)

Optional:

• --genome-path: Reference genome (.gb file)
• --single-file: Output single FASTA vs multiple files
• --output-file: Specific output filename
• --output-dir: Directory for outputs (default: ./minimized_genomes)
• --model-name: Label for file naming (default: "default")

Model Architectures

Preset	Architecture	Features
v0	1024→64	Linear KL annealing
v1	512→32	+ Gene abundance + L1 regularization
v2	512→32	+ Cosine annealing
v3	512→32	+ Weighted abundance

Output Structure

├── data/essential_genes/           # Preprocessing results
├── models/
│   ├── trained_models/v0_model/    # Saved model weights
│   └── v0_model/
│       ├── figures/                # Training plots
│       └── sampling_results/       # Generated samples
└── [output-dir]/                   # Final FASTA files

Troubleshooting

• Missing files: Pipeline automatically checks and shows missing files
• Import errors: Ensure virtual environment is activated
• GPU issues: Auto-detects GPU/CPU availability
• Path errors: Use absolute paths for model/data files

Pipeline flow: preprocess → training → sample → minimizer