NRPSTransformer

A Transformer-Based Predictor for Nonribosomal Peptide Synthetases (NRPS) Specificity

Welcome! You can try our web demo here: http://www.nrpstransformer.cn

Overview

NRPSTransformer is a deep learning model designed to predict the substrate specificity of adenylation (A) domains in Nonribosomal Peptide Synthetases (NRPS). It employs a transformer-based architecture that capitalizes on transfer learning. It generates high-quality protein embeddings from a large, pre-trained language model, which are subsequently fed into a fine-tuned classifier to achieve state-of-the-art prediction accuracy of 93%.

Publication

This work has been published in the Journal of the American Chemical Society (JACS).

DOI: 10.1021/jacs.5c08076

Prerequisites

Before you begin, ensure you have the following installed:

• Conda/Miniconda: To manage the project environment. You can find installation instructions here.
• Git LFS: For handling large model files. You can install it by following the instructions here.
• HMMER: For domain identification and scanning. Install it using:

  # On Ubuntu/Debian
  sudo apt-get install hmmer
  
  # On macOS with Homebrew
  brew install hmmer
  
  # On CentOS/RHEL
  sudo yum install hmmer

Installation

Please follow these steps carefully to set up the project environment and download the necessary models.

1. Clone the Repository First, clone this repository to your local machine.

git clone https://github.com/westlake-repl/NRPSTransformer.git
cd NRPSTransformer

2. Set Up the Conda Environment Create and activate a dedicated conda environment using the provided requirement.txt file.

# Create a new conda environment named 'NRPS' with Python 3.10
conda create -n NRPS python=3.10 -y
# Activate the newly created environment
conda activate NRPS
# Install the required Python packages
pip install -r requirements.txt

3. Download the ESM-2 Backbone Model This project uses the esm2_t33_650M_UR50D model as a backbone. You need Git LFS to clone it from Hugging Face.

# Initialize Git LFS
git lfs install
# Clone the ESM-2 model repository
git clone https://huggingface.co/facebook/esm2_t33_650M_UR50D

4. Download Project Checkpoints and Labels The model checkpoints and class labels are hosted on Zenodo. Download them and place them in the correct directories.

# Create the necessary directories
mkdir -p checkpoints
mkdir -p model/class_label

# Download checkpoint files into the 'checkpoints' directory
wget -O checkpoints/all.ckpt https://zenodo.org/records/15771488/files/all.ckpt
wget -O checkpoints/benchmark.ckpt https://zenodo.org/records/15771488/files/benchmark.ckpt
wget -O checkpoints/clade.ckpt https://zenodo.org/records/15771488/files/clade.ckpt

# Download label mapping files into the 'model/class_label' directory
wget -O model/class_label/labelid2label-17.pt https://zenodo.org/records/15771488/files/labelid2label-17.pt
wget -O model/class_label/labelid2label-43.pt https://zenodo.org/records/15771488/files/labelid2label-43.pt

After completing these steps, your project directory should look like this:

NRPSTransformer/
├── checkpoints/
│   ├── all.ckpt
│   ├── benchmark.ckpt
│   └── clade.ckpt

├── model/
│   ├── class_label/
│   │   ├── labelid2label-17.pt
│   │   └── labelid2label-43.pt
│   ├── esm2_t33_650M_UR50D/
│   └── ... (Hugging Face model files)
├── requirements.txt
├── run.sh
└── ... (other project files)

Usage

The primary way to run predictions is through the run.sh script. It provides an interactive command-line interface to specify your input and output files.

1. Execute the Script Make sure you are in the NRPSTransformer root directory and the NRPS conda environment is active.

bash run.sh

2. Interactive Prompts The script will prompt you for two pieces of information:

• Input File Path: The path to your FASTA file containing the protein sequences of the A-domains you want to predict.
• Result File Path: The path where the prediction results will be saved in CSV format. You can either type a new path or press Enter to accept the default values shown in the parentheses.

Here is an example of the terminal interaction:

Please enter the input sequence file path (Default: ./sequence/sequence.fasta):
# User can type a path like './my_sequences.fasta' or just press Enter

setting input file to ./sequence/sequence.fasta

Please enter the result file path (Default: ./result/result.csv): 
# User can type a path like './my_results.csv' or just press Enter

setting result path to ./result/result.csv

# The prediction process will start now...
# Running HMM scan...
# Parsing domain results...
# Generating domain sequences...
# Running AI prediction...
# Time: 45.2sec
# Successfully finished all steps!
# The result is saved to ./result/result.csv

Example Input File Format: Your input FASTA file should contain protein sequences with A-domains, like this:

>pepstatin-pepB(leu)
MSADTALSSAQQRLWFLWQLRPDGNEYNVPRATRLRGALDSDALRRALDLLVARHDALRATFPTVAGAPVLRIAAEAEVPLPLTDLSALPTAERDAALDEAVTRAALAPFDLAHGPVLRAELIRLAEDDHALVLTFHHIAVDGWSMGIIDRDLAALYAATVAGDDAPSAAPARSYRDCVAAEQRLLAGPRRHEALDHWRTELAGAPLELALPTDRPHPTAPSFAGDSRGFPVPPELAERLDAVAVRHRVTRFIALLSGYAALLSRMTSTPDLVIGVPVSGRTEMEVEGTVGLFVNMMPVRIRCTADTTFGALLHQVRDTVLAGHEYQDLPFQLLVEEVQPDRSTARHPLFQTVVTYEDLPTEGTALPGLDASPLPIPTRTAKYELALHLAGNRQRTEAWVGYQTDVFDDRSAELIGDRYLRLLSAALDDPDTPVAQLPVLGSEEERLLLTDWAGAPERAAPCPARERVDQLVERQARAHPDAVAVRSDEGALSYAALDRAADRIAAGLRAAGAGPGTVVATCLPRGADLVTAQLGVLKSGAAYLSLDPAHPAPRLTALLAEARPLLTLTGPEHRAGLEYGRVETLHSLRDAGDAPDAPDVPDAPDGPPAASAGPDDLAYVMYTSGSTGRPKGVLVEHHALANLVAWHRAEFGIGPGDRCTLVAAPGFDASVWETWSALTAGATVEVPSAETVLSPSELRSWLIERGITSAFLPTPLLERMLQEPWPAASALRSVLTGGDRLHGTGRNPLPFRLVNNYGPTENTVVATSGTVPADGGERDTLPGIGRPVTGTEAYVLDAELRPVPVGVPGELYLGGEGLARGYLGQPALTADRFVPHPFSRTPGARLYRTGDLVRWRTDAGLDFLGRNDHQVKIRGIRVEPGEIENVLRAHPGVHDAVVAAGSPDGTAEPELTAYLVPADASGAPDLAALLDHTTRRLPRHMHPRRYLLLPSVPLTANGKVDRTALAGAARELAPPPAAAHRRARTPLERLISDVWALTLGHDSFGTDDNFFDAGGHSLLLASVRDRLTAELGTPVRIADLYAYPTIATLARQLSGSPGSTAEERPAGRTGGDGVAERRRRGAARLTAMRARTSHGRAERQ

>pepstatin-pepD(ala)
MTDLTATAPGPTGTADADPSHPASFAQRRLWFLDQLDSGNAAYNLMAALRLRGPLDTEALHWSLNRIVDRHEVLRTVFRADGEGEPRQRVTPHRPLDLPVTDLTGQPREERDERARTLIETETERAFDLATGPLIRTRVVRLDQREHVLAVICHHTVCDGWSMDRLFAELSELYAARTAGRDPALEPLPLQFGEYAARQRGELAGAEARAALAHWRTRLQGAPTLLELPTSFPRPAVQGFDGATHTVPLSPGTWQEVRAAAARHKATPFMLLLSVFALQLSRLSGADDLVIGSPSAGRGRPELAPLIGMFVNTLPLRIDLSGEVTFEELLRRVRRSALDAFPRQDVPLERIVTELGPERARSHDPLFQTMFALQQPLSAPELAGLATDIFPASPRTTFTDLWLEIRPHAAGSDGTGGADCCFRYRTELFDAETVGRLARQYQHLLRTALDAPGTRLSRFSLTGEQETARLLRQGNRSAQAHPWDGPVHEAVDRQARRTPGATAVVFADERLSYAQLAGRTGQLARHLAAHGAAADTPVAVCLPRGTGLVTSVLAVMASGSAYLPLSPEDPPARLVRMLRAAGAAHILTTRELAPVLADAGVPVTAVDTFPWEEGGWPDRPMAPGRVGPDHLAYVIHTSGSTGAPKAVGVPHRGLANRIRGIQDTHGLDTTDRVLHKTPVTFDVSLWELLWPLTVGATLVVAAPGGHRDPTYLVELIERERVTTVHFVPPLLAAFLEEPGLHRCASLRRVLCSGQELPRATRDRCLERLPARLFNAYGPTEASIEVTEEECAPKATGGGAPDPRVSIGRPIAGAEVYVLDAELRPVPAGVPGELYLGGVALARGYLGQPALTADRFVPHPFSRTPGARLYRTGDLVRWRSDAALDFLGRNDHQVKIRGIRVEPGEIENVLRAHPDVRDAVVVATRPDGATGIQLTAYLVPYATDGSRAAPDADALVAAVDDRLRASLPGFMVPSRTVVLPRLPLNASGKVDRAALPAPAAPAPAPATARVAPADHVESTLARIWSQVLDRPGTDVTEDFFALGGDSLKSIQLVHRAREAGLSLRVGDIFHHPTVRDLAAHVRRTAAPAEPREDR

The program will then run the prediction pipeline, and upon completion, a message will confirm that the results have been saved to your specified output file.

Workflow Overview

NRPSTransformer follows a comprehensive pipeline to predict NRPS substrate specificity:

1. HMM Domain Scanning

• Uses HMMER to scan input protein sequences for A-domain regions
• Identifies conserved motifs and domain boundaries
• Generates domain-specific sequences for further analysis

2. Domain Extraction and Processing

• Parses HMM scan results to extract A-domain sequences
• Validates domain quality and filters out low-confidence predictions
• Prepares sequences in the format required by the transformer model

3. AI-Based Prediction

• Loads the pre-trained ESM-2 transformer model
• Applies fine-tuned classification layers for substrate specificity
• Generates confidence scores for multiple substrate predictions

4. Result Compilation

• Combines predictions with original sequence information
• Provides top-3 substrate predictions with confidence scores
• Outputs results in CSV format for easy analysis

Output Format

The prediction results are saved as a CSV file with the following columns:

Column	Description
ID	Original sequence identifier from input FASTA
Domain	Extracted A-domain sequence
Top-1(score)	Highest confidence substrate prediction with score
Top-2(score)	Second highest confidence substrate prediction with score
Top-3(score)	Third highest confidence substrate prediction with score

Example Output:

ID,Domain,Top-1(score),Top-2(score),Top-3(score)
pepstatin-pepB(leu),LAYVMYTSGSTGRPKGVLVEHHALANLVAWHRAEFGIGPGDRCTLVAAPGFDASVWETWSALTAGATVEVPSAETVLSPSELRSWLIERGITSAFLPTPLLERMLQEPWPAASALRSVLTGGDRLHGTGRNPLPFRLVNNYGPTENTVVATSGTVPADGGERDTLPGIGRPVTGTEAYVLD,gln(0.9978),ser(0.0011),hty(0.000151)
pepstatin-pepD(ala),LAYVIHTSGSTGAPKAVGVPHRGLANRIRGIQDTHGLDTTDRVLHKTPVTFDVSLWELLWPLTVGATLVVAAPGGHRDPTYLVELIERERVTTVHFVPPLLAAFLEEPGLHRCASLRRVLCSGQELPRATRDRCLERLPARLFNAYGPTEASIEVTEEECAPKATGGGAPDPRVSIGRPIAGAEVYVLD,asp(0.5783),asn(0.3875),salicylate(0.0146)
pepstatin-pepG-SA(val),PAYVIYTSGSSGDPKGVEVSHRNVTALLAACDRVFALRGDDVWTLFHSPCFDFSVWEMWGALAHGAKLVVVPAEVARSPEATLDLVVSEGVTVLNQVPSVFRYLSRSAVARAEDTADRPATALRYVIFGGEPVDVDAVRAWRALHGTRTEFFNMYGITETTVFATCRRLPESEIDPRPTTAPDAPAPTGTTADPELNIGRPLDGFEVVLLD,tyr(0.9705),ser(0.007952),cys(0.006917)
syrilipamide-nrps(leu-ser-hse-val-bala-ser),LAYVIYTSGSTGQPKGVMIEHRNLVNLVAWHCEAFGLTHRKRVSSVAGVGFDACVWELWPALCVGASLSLLPGQALGNDVDALLGWWRRQDLDVSFLPTPIAEIAFAQGIEPASLQTLLIGGDRLRQFPNPDSRVALINNYGPTETTVVATSGLIDATQSVLHIGRPIANTQVYLLD,gln(0.9996),arg(8.013e-05),ser(6.634e-05)

Advanced Usage

Direct Python Script Execution

For users who prefer more control over the prediction process, you can run individual components directly:

1. HMM Domain Scanning

# Run HMM scan on your sequences
./hmm/scan.sh your_sequences.fasta

# Parse the HMM results
python ./hmm/parse_dbtl.py

# Generate domain sequences
python ./hmm/gen_domains.py --fasta_path your_sequences.fasta

2. Direct AI Prediction

# Run prediction with custom parameters
python inference.py --inference_dataset hmm/result/domains.csv --result_path your_results.csv

Custom Model Checkpoints

You can use different pre-trained model checkpoints for specific use cases:

• all.ckpt: General-purpose model trained on all available data (recommended)
• benchmark.ckpt: Model optimized for benchmark datasets
• clade.ckpt: Model specialized for specific bacterial clades

To use a different checkpoint, modify the CHECKPOINT_PATH variable in inference.py:

CHECKPOINT_PATH = "checkpoints/benchmark.ckpt"  # or clade.ckpt

Batch Processing

For processing multiple FASTA files, you can create a simple batch script:

#!/bin/bash
# batch_process.sh

for fasta_file in /path/to/your/fasta/files/*.fasta; do
    echo "Processing $fasta_file"
    python inference.py --inference_dataset "$fasta_file" --result_path "${fasta_file%.fasta}_results.csv"
done

Troubleshooting

Common Issues and Solutions

1. HMMER Not Found

Error: The input file 'sequence.fasta' does not exist.

Solution: Ensure HMMER is properly installed and accessible in your PATH:

which hmmsearch  # Should return the path to hmmsearch
hmmsearch -h     # Should show help information

2. CUDA/GPU Issues

RuntimeError: CUDA out of memory

Solutions:

• Reduce batch size in inference.py (change BATCH_SIZE = 8 to BATCH_SIZE = 4 or BATCH_SIZE = 2)
• Use CPU-only mode by modifying the trainer configuration:

trainer = pl.Trainer(
    accelerator="cpu",  # Change from "gpu" to "cpu"
    devices=1,
    logger=False,
    enable_progress_bar=False,
    enable_model_summary=False,
)

3. Model Download Issues

Error: Failed to download model files

Solutions:

• Ensure Git LFS is properly installed: git lfs install
• Check internet connection and try downloading manually:

wget -O checkpoints/all.ckpt https://zenodo.org/records/15771488/files/all.ckpt

4. Memory Issues

MemoryError: Unable to allocate array

Solutions:

• Close other applications to free up RAM
• Process sequences in smaller batches
• Use a machine with more available memory

5. No A-domains Found

Warning: No A-domains detected in input sequences

Solutions:

• Verify your input sequences contain NRPS A-domains
• Check that sequences are in proper FASTA format
• Ensure sequences are long enough (>200 amino acids)
• Try with known NRPS sequences first

6. Permission Issues

Permission denied: cannot execute script

Solutions:

chmod +x run.sh
chmod +x hmm/scan.sh

Getting Help

If you encounter issues not covered here:

1. Check the GitHub Issues page
2. Verify your environment matches the requirements
3. Try running with the provided example sequences first
4. Check the log files in the project directory for detailed error messages

Citation

If you use NRPSTransformer in your research, please cite our paper:

@article{nrpstransformer2024,
  title={NRPSTransformer: A Transformer-Based Predictor for Nonribosomal Peptide Synthetases Specificity},
  author={[Authors]},
  journal={Journal of the American Chemical Society},
  year={2024},
  doi={10.1021/jacs.5c08076}
}

Contributing

We welcome contributions to NRPSTransformer! Here's how you can help:

Reporting Issues

• Use the GitHub Issues page to report bugs
• Include detailed information about your environment and the error
• Provide example sequences that reproduce the issue (if applicable)

Contributing Code

1. Fork the repository
2. Create a feature branch: git checkout -b feature-name
3. Make your changes and test thoroughly
4. Commit your changes: git commit -m "Add feature"
5. Push to the branch: git push origin feature-name
6. Submit a pull request