Skip to content

Latest commit

 

History

History

protvec

readme.md

ProtVec Embedding Generator

This repository provides a simple and reproducible way to generate ProtVec embeddings for protein sequences using a pretrained Word2Vec model.
Each protein is converted into raw k-mer–level embeddings (3-mers) and saved as a Python dictionary in .npy format.

This script is suitable for downstream tasks such as protein–protein interaction (PPI) prediction, graph neural networks (GNNs), Transformers, or other deep learning pipelines.

What this script does

For each protein sequence, the script:

  • Splits the sequence into overlapping 3-mers
  • Uses three reading-frame offsets (0, 1, 2)
  • Looks up embeddings from a pretrained ProtVec model
  • Stores raw k-mer embeddings (shape: N × 100)
  • Saves one .npy file per protein

Each saved file contains a dictionary with the following keys:

{
  "protein_id": "<protein_id>",
  "fasta": "<protein_sequence>",
  "embedding": numpy.ndarray  # shape: (num_kmers, 100)
}

Requirements

1. Create and activate a conda environment (recommended)

conda create -n protvec_env python=3.8
conda activate protvec_env

2. Install required Python packages

pip install numpy pandas gensim

Pretrained ProtVec model

You need a pretrained ProtVec Word2Vec model (trained on protein 3-mers, e.g., SwissProt). The pretrained model can be downloaded from https://github.com/kyu999/biovec/tree/master/trained_models link

In the script, the model is loaded as:

model = Word2Vec.load(
    "/home/saiful/ePPI_dgl/protvec/swissprot-reviewed-protvec.model"
)

You may change this path to wherever your ProtVec model is stored.

Important assumptions about the model:

  • Tokenization uses 3-mers

  • Embedding dimension is 100

  • Model is compatible with gensim.models.Word2Vec

Input format

The script expects a CSV file with two columns:

Column name Description protein_id Unique protein identifier fasta Protein sequence (amino acids)

Example CSV (prot_to_fasta.csv)

protein_id,fasta
P12345,MKVLYNLKDGKVT...
Q9XYZ1,MSDTQLERK...

How to run the script

  1. Clone the repository
git clone https://github.com/<your-username>/<repository-name>.git
cd <repository-name>
  1. Edit paths inside the script
Open get_protvec_embeddings1.2.py and update the following paths:

# Path to ProtVec model
model = Word2Vec.load("PATH_TO_PROTVEC_MODEL")

# Path to input CSV
file_path = "PATH_TO_PROT_TO_FASTA_CSV"

# Output directory
output_dir = "PATH_TO_OUTPUT_DIRECTORY"
  1. Run the script
python get_protvec_embeddings1.2.py

Output

  • One .npy file is generated per protein

  • File naming format:

<protein_id>_protvec.npy

Example:
A0A024R0L9_protvec.npy

Each file stores:

  • Protein ID, FASTA sequence, and Raw ProtVec k-mer embeddings (N × 100)

Notes

This script does NOT pool or average embeddings; it preserves raw k-mer–level information