Weighted Scoring based PETase expression and activity prediction

Author: Farrel Alfaza, Kayla Queenazima, Melodia Rezadhini*, Nayaka Bagus, Sofyan Maulana Institutions: University Gadjah Mada and Tokyo Tech Uni (*)

Brief Description

The idea of weighting scoring is to approximately measure the combined parameters that will affect PETase protein sequence directly/indirectly on enzymatic process of TPA and its expression performance.

Pipeline

|- main.py |--scoring.py (scoring and final calculation of expression and activity) |--parameter worker |---seq_features.py |---esm clustering |---blossum-alignment |---kcat-km-prediction |---pH-optimum-prediction

---

🔧 Environment Setup (CPU-only)

This project uses Conda to ensure a reproducible Python environment.

Step 1 — Create the Conda environment

conda env create -f environment.yml

What this does:

• Creates a new Conda environment named petase_zero_shot
• Installs Python 3.10 and all core scientific dependencies (NumPy, Pandas, SciPy, Biopython, fair-esm, etc.)

You only need to run this once.

---

Step 2 — Activate the environment

conda activate petase_zero_shot

What this does:

• Switches your shell to use the newly created environment
• Ensures that all Python packages and commands run inside the correct environment

Always activate this environment before running any scripts in this repository.

---

Step 3 — Install PyTorch (CPU version)

pip install -r requirements.txt

What this does:

• Installs PyTorch (CPU-only) and related libraries (torchvision, torchaudio)
• Uses the official PyTorch CPU wheel index for compatibility across systems

We install PyTorch separately because:

• PyTorch CPU wheels are distributed via a special index URL
• Separating it avoids common installation issues on Windows/Linux

---

✅ Verify Installation (Optional)

After setup, you can verify that everything works by running:

python - <<EOF
import torch
import esm
print("Torch version:", torch.__version__)
print("ESM loaded successfully")
EOF

If no errors appear, your environment is ready.

---

ℹ️ Notes

• This setup is CPU-only and does not require a GPU
• For large datasets, ESM scoring may take longer on CPU
• GPU users can create a separate environment if desired

Author Contributions

This project was developed collaboratively by a team from Universitas Gadjah Mada, in collaboration with Tokyo Institute of Technology (Melodia Rezadhini).

Kayla Queenazima (UGM) — Conceptualised the weighted scoring framework integrating multiple biological parameters; designed the overall pipeline architecture combining ESM embeddings, BLOSUM alignment scoring, and kcat/Km prediction; managed submission workflow and final parameter tuning.

Sofyan Maulana (UGM) — Implemented ESM zero-shot scoring module (forge integration) for PETase sequence embedding generation; developed delta-score analysis comparing engineered variants against wild-type.

Farrel Alfaza, Nayaka Bagus (UGM), Melodia Rezadhini (Tokyo Institute of Technology) — Data curation and pre-processing of PETase sequence datasets; biological validation of computational predictions; cross-checking of scoring outputs against known enzymatic behaviour and structural constraints; experimental context interpretation; External validation; biological plausibility assessment of predicted sequences from a protein engineering perspective; advisory input on PETase enzymatic mechanism and structural-functional relationships.

Context

This work was developed as a submission to the alignBio PETase Protein Engineering Tournament (https://alignbio.org/benchmarks/) — a community benchmark for ML-based prediction of PETase enzyme variants. The pipeline combines ESM protein language models with weighted feature scoring to predict expression and activity of engineered PETase sequences.

Note: As of writing, the benchmark organisers have not announced the tournament results.