This tutorial demonstrates how to load .npy embedding files from the DeepDrug Protein Embeddings Bank (DPEB), perform mean pooling, and apply clustering and t-SNE visualization to analyze protein families.

This script will perform the following actions:

• Load AlphaFold2 .npy embedding files stored in the dataset
• Apply mean pooling across residue-level embeddings to produce fixed-length vectors
• Aggregate per-residue embeddings into fixed-length vectors using mean pooling
• Merge these embeddings with protein family labels from protein_families23k.csv
• Save the aggregated results to eppi_alphafold_aggregated_embeddings.csv
• Generate t-SNE plots and save them as:
  • raw_embeddings_tsne_Alphafold.png
  • raw_embeddings_kmeans_tsne_Alphafold.png
• Print clustering evaluation metrics (Accuracy, Precision, Recall, F1-Score)

• AlphaFold2 embeddings: .npy files inside a .rar archive
• Protein family annotations: protein_families23k.csv
• Output aggregated embeddings: eppi_alphafold_aggregated_embeddings.csv

• tutorial_clustering.py: Main script that performs loading, aggregation, clustering, and visualization.
• protein_families23k.csv: CSV file mapping each protein to its family (used for supervised evaluation).
• Output:
  • eppi_alphafold_aggregated_embeddings.csv (generated), we have also provided aggregated embeddings of each type in csv format inside each embedding folder.
  • generated t-SNE plots/raw_embeddings_tsne_Alphafold.png (generated)
  • generated t-SNE plots/raw_embeddings_kmeans_tsne_Alphafold.png (generated)

• numpy, pandas
• matplotlib, scikit-learn
• torch, TSNE, KMeans
• ast, LabelEncoder

Activate your conda environment and install any missing dependencies.

conda activate DPEB
pip install scikit-learn matplotlib pandas torch

Before running the tutorial script, make sure to complete the following steps:

Download the embeddings folder the Box link. For example:

• All_ePPI_Alphafold2_Embeddings_np_v1.3.rar`
• esm2_dict_embeddings.rar
• protvec_dict_embeddings.rar
• bioemb_dict_embeddings.rar

Extract them using the following commands:

# Install unrar if needed
sudo apt install unrar       # For Ubuntu/Debian
# or
sudo yum install unrar       # For RHEL/CentOS

# Extract the .rar file
unrar x All_ePPI_Alphafold2_Embeddings_np_v1.3.rar

Download protein_families23k.csv from the GitHub tutorial folder.

In your script (e.g., tutorial_clustering.py), update the following paths:

# Example: Set this to your local extracted embedding directory
embedding_folder = "/your/local/path/to/All_ePPI_Alphafold2_Embeddings_np_v1.3/"

# Example: Set this to your local path to the metadata CSV file
protein_file = "/your/local/path/to/protein_families23k.csv"

You can execute the tutorial by running the Python script inside the tutorial/ folder:

python tutorial_clustering.py

The t-SNE visualizations generated by the tutorial script can be found at:

tutorial/generated t-SNE plots