CLASP: Contrastive Learning of Amino acid, Structure, and Protein description

CLASP is a tri-modal contrastive learning framework for unified representation of protein structure, sequence, and description. It enables downstream applications such as cross-modal retrieval, similarity scoring, and zero-shot classification by learning a shared embedding space across all three modalities.

The implementation of CLASP is described in "Multi-Modal Protein Representation Learning with CLASP".

Table of contents

1. Overview of the CLASP framework
2. Repository structure
3. Environment setup
4. Documentation and usage
5. Versions tested
6. Typical install and run time
7. BibTex Citation
8. License

Overview of the CLASP framework

CLASP (Contrastive Learning of Amino acid, Structure, and Protein description) is a tri-modal contrastive learning framework designed to unify protein structure (as PDB files), amino acid sequence, and natural language description into a shared embedding space. By learning aligned representations across these biologically complementary views, CLASP supports downstream applications such as cross-modal retrieval, similarity scoring, and zero-shot classification without requiring task-specific supervision.

Each modality is first encoded independently: protein structures are converted into residue-level graphs using Graphein and embedded via an E(3)-invariant graph neural network (EGNN), amino acid sequences are embedded using pretrained language models like ProtT5, and textual descriptions are embedded using biomedical large language models such as BioGPT. Lightweight projection heads then map all three modalities into a common vector space. CLASP is trained using a tri-modal contrastive loss that simultaneously encourages alignment between matching structure-sequence, structure-description, and sequence-description pairs.

Repository structure

CLASP/
├── assets/                # directory for asset files
├── docs/                 
│   ├── data_preparation.md
│   ├── evaluation.md
│   ├── inference_utilities.md
│   ├── training_clasp.md
│   └── README.md
├── src/                   
|   ├── models/            # model definitions for CLASP
│   ├── utils/             # utility functions for data processing and model training
│   ├── compute_similarity_matrices.py
│   ├── eval_zero_shot_classification.py
│   ├── get_quick_similarity_scores.py
│   ├── preprocess_pdb_graphs.py
│   ├── rank_amino_acids_by_description.py
│   ├── train_clasp.py     
├── .gitignore             
├── environment.yml       
└── README.md      

Environment setup

conda env create -f environment.yml
conda activate claspenv

After activating the environment, you need to manually install torch-scatter to match your device. For example, we use the following command for a CUDA 12.4 compatible setup with PyTorch 2.6.0 (this command will very if you use a CPU-only setup or a different CUDA version):

pip install torch-scatter==2.1.2+pt26cu124 -f https://data.pyg.org/whl/torch-2.6.0+cu124.html

On a CPU-only setup, you can use:

pip install torch-scatter==2.1.2 -f https://data.pyg.org/whl/torch-2.6.0+cpu.html

Documentation and usage

This repository provides a set of utilities and scripts for preparing data, training the CLASP model, evaluating its performance, and running inference. Below is a summary of key components, their purposes, and where to find detailed instructions. More information can be found in the documentation.

Component	Purpose	Documentation
preprocess_pdb_graphs.py	Converts .pdb files into residue-level graphs compatible with CLASP	data_preparation.md
compute_similarity_matrices.py	Computes projected embeddings and similarity matrices between structure, sequence, and description	inference_utilities.md
get_quick_similarity_scores.py	Prints similarity scores for a single triplet (structure, sequence, description)	inference_utilities.md
rank_amino_acids_by_description.py	Ranks amino acid sequences based on a natural language query	inference_utilities.md
train_clasp.py	Main training script for CLASP	training_clasp.md
eval_zero_shot_classification.py	Evaluates a trained CLASP model on zero-shot classification tasks (PDB–AAS, PDB–DESC, AAS–DESC)	evaluation.md

Versions tested

CLASP has been trained and tested with the following versions:

• Python 3.10
• PyTorch 2.6.0
• Graphein 1.7

For a full list of dependencies, please refer to the environment.yml file.

Typical install and run time

The typical installation time for the CLASP environment is approximately 5-10 minutes, depending on your internet connection and system performance. The training time for the CLASP model on a standard GPU (e.g., NVIDIA RTX 3090 with 32 CPU cores) is approximately 13.08 hours (∼2 minutes per epoch) when using the datasets and hyperparameters specified in the documentation. The training time may vary based on the specific hardware configuration, dataset size, and hyperparameters used.

BibTex Citation

@article{bolouri2025clasp,
  title     = {Multi-Modal Protein Representation Learning with CLASP},
  author    = {Bolouri, Nicolas and Szymborski, Joseph and Emad, Amin},
  journal   = {bioRxiv},
  year      = {2025},
  doi       = {10.1101/2025.08.10.669533},
  url       = {https://www.biorxiv.org/content/early/2025/08/12/2025.08.10.669533}
}

License

Copyright (C) 2025 Emad's COMBINE Lab: Nicolas Bolouri, Joseph Szymborski, and Amin Emad.

CLASP is free software: you can redistribute it and/or modify it under the terms of the
GNU General Public License as published by the Free Software Foundation, either version 3
of the License, or (at your option) any later version.

CLASP is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without
even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with CLASP.
If not, see https://www.gnu.org/licenses/.