🧬 SinCAA

Learning representations for peptides containing non-canonical amino acids (ncAAs).
This repository provides the implementation, pretrained models, and training pipeline for the paper:

"Similarity-Enhanced Representation Learning of Non-Canonical Amino Acids for Therapeutic Peptide Modeling"
Chencheng Xu et al., 2025

---

🚀 Overview

Standard protein language models are typically trained only on canonical amino acids, which limits their capacity to model peptides containing ncAAs. This project introduces SinCAA, a framework that integrates a 3D conformational similarity metric into a graph transformer trained with dual objectives—contrastive learning and masked-node reconstruction—to generate transferable molecular embeddings capable of generalizing from individual ncAAs to complete peptides. SinCAA demonstrates strong performance across multiple downstream tasks, including peptide binding affinity prediction, cell-penetrating ability estimation, and protein–peptide binding site prediction. Notably, it achieves substantial improvements over existing methods and exhibits remarkable zero-shot generalization, underscoring its potential to accelerate therapeutic peptide discovery.

🔥 Quick start

To ensure reproducibility, all dependencies required for SinCAA can be installed via the provided environment.yml file. Follow the steps below to create and activate the environment.

conda env create -f environment.yml ; conda activate sincaa

SinCAA relies on OpenFold for feature extraction. Please follow the official OpenFold installation instructions to install it properly.

The pretrained weights for SinCAA are available at: data/results/n1_weight0.1_innl2_both/

To generate embeddings for peptides or amino acids from a CSV file containing molecular representations in the SMILES column, execute the following command:

python -m Tools.generate_emb_from_smiles \
    --csv_path path_to_file.csv \
    --pretrained_dir data/results/n1_weight0.1_innl2_both/ \
    --save_path save_path.pt

Here:

• --csv_path specifies the path to the input CSV file (e.g., data/examples/exam_csv.csv).

• --pretrained_dir provides the directory containing the pretrained SinCAA model weights.

• --save_path defines the output path for saving the generated embeddings. The file is a PyTorch serialized object saved with torch.save and can be loaded in Python using torch.load.

📂 Repository Structure

The structure of the SinCAA repository is as follows:

SinCAA/
├── data/                       # Datasets and pretrained model weights
│   ├── AAList/                 # Non-canonical amino acids for model pretrianing
│   ├── examples/               # Examples to show model inputs
│   └── results/                # Model outputs and pretrained weights
│       └── n1_weight0.1_innl2_both/
│
├── Tools/                      # Utility scripts and helper functions
│   ├── generate_emb_from_smiles.py   # Generate embeddings from SMILES strings
│   └── build_sim_aa.py         # Generate similar amino acids
├── Experiments/                      # Code for downstream experiments
│   ├── 0_benchmark   # Cell penetration ability prediction
│   └── 1_ppb   # Protein-peptide binding prediction
│
├── utils/                      # Utility functions for similarity calculations and data preparation
│   ├── data_utils.py           # Data loader utilities
│   ├── data_constants.py       # Constants for data preparation
│   ├── align_utils.py          # Utilities for protein chain alignment
│   ├── amino_acid.py           # Amino acid feature extraction utilities
│   ├── feats_acid.py           # Protein feature preprocessing utilities
│   ├── rigid_utils.py          # Utilities for rigid structure handling
│   ├── similarity_utils.py     # Utilities for conformational similarity calculation
│   ├── train_utils.py          # Utilities for model training
│   └── build_sim_aa.py         # Generate similar amino acids
│
├── models/                     # Model architectures
│   └── sincaa.py
│
├── figs/                  # Figures for README
│
├── scripts/                    # Shell scripts for training
│   ├── split_data.py          # Utilities for model training
│   └── train.sh
│
├── environment.yml             # Dependencies
├── README.md                   # Project overview and usage instructions
└── LICENSE                     # License information

Usage

Model Training

Download the ZINC15 10M dataset from the official website. Split the training and validation datasets:

python scripts/split_data.py path_to_zinc_data

Start model training:

bash scripts/train.sh

Similarity Calculation

To calculate the similarity between a pair of non-canonical amino acids A and B:

from utils.similarity_utils import get_space_distance

grid_size = 1        # Recommended values: 0.5, 1, or 2
num_samples = 20     # Recommended: 20 or higher

similarity = get_space_distance(A, B, grid_size, num_samples=num_samples)
print(similarity)

Downstream Applications

For experiments related to downstream applications, please refer to the README in each folder under experiments/.