ProtVAE

This project implements a Variational Autoencoder (VAE) to encode and generate protein sequences using a one-hot encoded representation of amino acids.

Features

• Supports multi-sequence FASTA files for input.
• Uses a one-hot encoding scheme for amino acids, including padding.
• Trains a VAE to learn a latent representation of protein sequences.
• Generates new protein-like sequences by sampling from the latent space.

Installation

1. Clone this repository:

   git clone https://github.com/annadiarov/ProtVAE
   cd ProtVAE

2. Install the required dependencies (we recommend using a virtual environment and install pytorch using the instructions from the official website):

   pip install -r requirements.txt

Usage

Training

To train the VAE, run the following command:

python src/training.py --fasta_file data/example.fasta --epochs 50 --batch_size 32

This will train the VAE on the sequences in data/example.fasta for 50 epochs using a batch size of 32. The trained weights will be saved as vae_weights.pth by default but can be changed using the --output-weights argument.

Generating Sequences

After training the VAE, you can generate new sequences by running:

python src/sampling.py --weights vae_weights.pth --num_samples 10

This will generate 10 new sequences using the trained VAE weights in generated_sequences.fasta by default, but this can be changed using the --output-file argument.

License

This project is licensed under the MIT License - see the LICENSE file for details.