Protein Sequence Generator

https://www.bakerlab.org/2023/05/08/deep-learning-improves-protein-binder-design-tenfold/

Description

This project is a Protein Sequence Generator that utilizes a Long Short-Term Memory (LSTM) neural network to generate new protein sequences based on existing ones from the UniProt Swiss-Prot database. The model learns patterns in protein sequences and generates sequences that mimic real proteins.

Project Initialization

Prerequisites

• Python 3.10 >=
• pip package manager.

Install Dependencies

Navigate to the project directory and install the required packages using:

pip install -r requirements.txt.

Download the Dataset.

You need to download the protein sequence dataset from UniProt:

Visit the UniProtKB website: https://www.uniprot.org
Select the FASTA format.
Save the file as data/uniprot_sprot.fasta in the project directory.

Training the Model

Run the following command to train the model:

python train_model.py

This script will:
Load and preprocess the protein sequences.
Train the ProteinLSTM model.
Save the trained model models/protein_lstm_model.pth and the amino acid mappings (models/aa_to_int.pth and models/int_to_aa.pth).

Generating Protein Sequences

After training, generate new protein sequences by running:

python generate_sequence.py

This script will:
Load the trained model and amino acid mappings.
Generate a new protein sequence based on a random seed sequence.

Sample Output

Below is an example of a generated protein sequence:

Generated Protein Sequence:

The sequence consists of single-letter codes for amino acids, following the standard IUPAC notation.
Each letter corresponds to one of the 20 standard amino acids found in proteins.
FHGKDCTRRPMMDKHDITCDPTDPMTKMEIHCMHPMIKDHTDCQNPNISPISDDPPKCPDPKPRRRRSSSSMMMHHHKMHIQIQDSSSSNRNSNSNNSPNRRMTKPMMRRRPRPPPPPPPPPKKQRKQQRNQQRNRNRQ

Notes

• Customization: You can adjust parameters like num_records, sequence_length, num_epochs, and num_generated in the code files to suit your needs.
• Dataset Size: By default, the code uses 100 protein sequences for training. For better results, consider increasing num_records to include more sequences.
• Model Parameters: Feel free to experiment with the model's hyperparameters such as embedding_dim, hidden_dim, and num_layers.