RCSB Protein Structure Embedding Search

This repository contains the scripts to train a neural network model for Protein Structure Search based on protein-level embeddings.

Protein Encoding

Our model ingests protein residue-level embeddings computed with a structural aware protein language model available in this repo.

Training and Testing Datasets

Training and testing the model was achieved using the same datasets described in this publication. The training and testing datasets were carefully curated from CATH and ECOD domain classifications, ensuring independence and relevance. Ready-to-use residue-level embeddings for the training and testing domains in the datasets can be downloaded from zenodo. Domain classification identifiers for the different CATH and ECOD domain structures can be found in the resource folder.

Training and Testing

The main script for training and testing is lightning_structure_embedding.py. Two params are required

• --class_path Should point to the path were CATH and ECOD classification Ids files are located. The files are expected to be named as cath.tsv and ecod.tsv
• --embedding_path Should point to the path where the training and testing residue-level embeddings are stored. The script expect CATH domains embeddings to be stored in <embedding_path>/cath/embedding and ECOD <embedding_path>/ecod/embedding