T-domain-design

Code and workflows accompanying the preprint:

Generative AI designs functional thiolation domains for reprogramming non-ribosomal peptide synthetases
bioRxiv (2026), DOI: https://doi.org/10.64898/2026.03.03.709401

Project Overview

This repository focuses on AI-assisted design and evaluation of thiolation (T) domains in NRPS systems.

The codebase currently provides:

• sequence generation workflows for T-domain redesign with fixed A/C context
• surrogate modeling workflows for protein sequence activity prediction
• utilities for sequence/structure handling and evaluation metrics
• a vendored ProteinMPNN submodule for structure-conditioned design experiments

Repository Layout

• data/:
  • example inputs and generated sequence files used by scripts/notebooks
  • includes dataset_gb1.csv and GB1_WT.fasta for surrogate-model examples
• generation/:
  • generation workflows using ESM3, EvoDiff, and ProteinMPNN
  • see generation/README.md
• surrogates/:
  • notebook tutorial for surrogate modeling (one-hot, PLM embeddings, zero-shot, fine-tuning)
  • see surrogates/surrogates.ipynb and surrogates/README.md
• src/:
  • reusable Python modules:
    • src/esm/ model wrappers for ESM2, ESMC, ESM3
    • src/models/ supervised and fine-tuning model implementations
    • src/eval/ evaluation utilities
    • src/protein/, src/utils/ protein and helper utilities
• ProteinMPNN/:
  • git submodule mirror of ProteinMPNN

Installation

1) Clone and initialize submodules

git clone <your-fork-or-this-repo-url>
cd T-domain-design
git submodule update --init --recursive

2) Create Python environment

conda create -n tdomain python=3.10.13 pip
conda activate tdomain

3) Install dependencies

pip install -r requirements.txt

The main requirements.txt is now profile-oriented with comments so you can switch PLM stacks:

• ESMC/ESM3 path (recommended default): keep esm
• ESM2 path: use fair-esm instead of esm

Do not install both esm and fair-esm in the same environment.

Quick Start

Generation workflows

From generation/:

• ESM3 generation:

python generate_esm3.py

• EvoDiff generation:

python generate_evodiff.py

• ProteinMPNN workflow:
  • open and run generation/generate_pmpnn.ipynb

See generation/README.md for domain-index configuration and expected input files.

Surrogate modeling workflow

Open and run:

• surrogates/surrogates.ipynb

This notebook covers:

1. one-hot encoding based regressors
2. PLM embedding based regressors
3. zero-shot PLM scores
4. PLM fine-tuning (regression and contrastive)

Default PLM in surrogate workflows is ESMC.

Data Notes

Included example files under data/:

• dataset_gb1.csv: cleaned GB1 fitness data used in surrogate tutorial
• GB1_WT.fasta: GB1 wild-type sequence for zero-shot/fine-tuning sections
• gxps_ATC_AF.pdb: example structure used by generation scripts

Generated FASTA files in data/ (for quick inspection/examples):

• esm3_gen.fasta
• evodiff_gen.fasta
• pmpnn_gen.fasta

Citation

If you use this repository or build on the methods, please cite the associated preprint:

• Bülbül EF, Bang S, George K, et al. Generative AI designs functional thiolation domains for reprogramming non-ribosomal peptide synthetases. bioRxiv (2026)

BibTeX:

@article{buelbuel2026generative,
  title={Generative AI designs functional thiolation domains for reprogramming non-ribosomal peptide synthetases},
  author={Buelbuel, Emre F and Bang, Seounggun and Geroge, Kevin and Bianchi, Gabriele and Raj, Prateek and Chung, Seonyong and Pauline, Vincent and Hochstrasser, Ramon and Minas, Hannah A and Elgaher, Walid AM and others},
  journal={bioRxiv},
  pages={2026--03},
  year={2026},
  publisher={Cold Spring Harbor Laboratory}
}

Notes and Limitations

• Several scripts/notebooks are research-grade and may contain hardcoded file paths or environment assumptions.
• GPU is assumed in parts of the generation and PLM workflows.
• For ESM2 vs ESMC/ESM3 usage, prefer separate conda environments.