PlaszymeAlpha

Overview

PlaszymeAlpha is a machine learning framework designed for the classification and analysis of plastic-degrading enzymes (Plaszymes). It leverages protein sequence embeddings and supervised learning models to predict degradation potential across different plastic types.

The repository provides utilities for data preprocessing, feature embedding generation, model training, evaluation, and prediction on new protein sequences.

---

Repository Structure

PlaszymeAlpha/
├── data/               # (Optional) Dataset folder for input sequences and labels
├── models/             # Saved models (e.g., classifier.pkl, label_encoder.pkl)
├── results/            # Output predictions and evaluation metrics
├── utils.py            # Helper functions (sequence cleaning, embedding extraction, etc.)
├── train.py            # Training script for model building
├── test.py             # Testing/inference script
├── requirements.txt    # Python dependencies
└── README.md           # Project documentation

---

Installation

esm model and configs are needed while using this model. You can download from https://dl.fbaipublicfiles.com/fair-esm/models/esm1b_t33_650M_UR50S.pt

It is recommended to create a virtual environment (e.g., Conda or venv).

# Example using conda
conda create -n plaszymer python=3.9
conda activate plaszymer

# Install required dependencies
pip install -r requirements.txt

---

Usage

Data Preparation

• Input should be provided as a CSV file containing at least two columns:
  • protein_id: Unique identifier for each sequence
  • sequence: Raw protein sequence (string of amino acids)

Example:

protein_id,sequence,degradable_plastics
seq1,MAADQLTAR...,PET
seq2,MKVLWAALL...,PE

Training

Run the training script with your dataset:

python train.py --train_csv path/to/train.csv --out_dir models/ --batch_size 4 --epochs 50

This will save the trained classifier and label encoder to the models/ folder.

Testing / Prediction

Run the test script with a new dataset:

python test.py --test_csv path/to/test.csv --model_dir models/ --out_csv results/predictions.csv --batch_size 4

Output will include:

• Prediction probabilities for each plastic class
• Top hit per sequence

Example Output

protein_id,PET,PE,PU,PHA,PCL,PBAT,PLA,TopHit
seq1,0.92,0.03,0.01,0.01,0.00,0.02,0.01,PET
seq2,0.05,0.80,0.03,0.01,0.02,0.06,0.03,PE

---

Results

Model performance is evaluated using standard classification metrics (accuracy, precision, recall, F1-score).
Results will be stored in the results/ directory after evaluation.

---

License

This project is licensed under the MIT License.
See the LICENSE file for details.

---

Acknowledgments

• Thanks to FAIR (Meta AI) for releasing the ESM protein language models, which form the basis of sequence embedding in this project.
• We acknowledge the use of scikit-learn as core framework supporting model development, training, and explainability.
• Protein sequence data were curated from publicly available resources such as UniProt and PDB, as well as previously published plastic-degrading enzyme studies.
• This project was inspired by recent advances in protein language models, particularly Protein language models accelerate the discovery of plastic-degrading enzymes (Medina-Ortiz,D. et al., 2024).
• We thank all team members and collaborators for their contributions to data preparation, model design, and experimental validation.
• This work was supported by XJTLU-AI-CHINA.