Protein-Synformer

A protein‑conditioned fork of SynFormer Generative modelling and synthesis‑planning for drug discovery.

---

1. Overview

Prot2Drug‑SynFormer extends the original SynFormer framework to generate synthetic routes for small‑molecule ligands conditioned on a target protein. Key additions come from our MSc research (see paper draft in /docs/) and include:

New feature	Where to look
Protein‑aware decoder using per‑residue embeddings from ESM‑C 600 M	src/synformer/models/protein_decoder.py
Updated scripts/train.py – PyTorch‑Lightning datamodule, mixed‑precision, gradient accumulation	scripts/train.py
scripts/hyperopt.py – Optuna‑powered hyper‑parameter search	scripts/hyperopt.py (requires Optuna)
scripts/evaluate.py – unified evaluation (route validity, length, novelty, docking proxy)	scripts/evaluate.py
Pre‑computed protein embeddings (16‑bit, ESM‑C 600 M)	data/protein_embeddings/

The fork retains SynFormer’s fast transformer‑based inference and support for Enamine REAL building blocks while adding protein context and modern training utilities.

---

2. Installation

2.1 Environment

# create conda env\conda env create -f env.yml -n prot2drug
conda activate prot2drug

# install package in editable mode
pip install --no-deps -e .

The provided env.yml includes PyTorch 2.3, PyTorch‑Lightning 2.2, Optuna 3.x, and RDKit nightly.

2.2 Building Block Database

We provide pre‑processed building‑block data (fpindex.pkl and matrix.pkl). You can download these files from our Hugging Face page and place them in the data directory specified in your YAML config.

Important: the files are derived from Enamine’s REAL building‑block catalogue and are available only upon request. Please request the US Stock catalogue from Enamine and store the raw SMILES under data/building_blocks. The processed data are for academic research only; any commercial use requires permission from Enamine.

2.3 Trained Models

Pre‑trained checkpoints can be downloaded from the same Hugging Face page and stored in data/trained_weights/.

2.4 Train with your own templates and building blocks

If you prefer a custom set of reaction templates or building blocks, edit the paths on lines 6–7 of the relevant config file and run:

python scripts/preprocess.py --model-config your_config_file.yml

to regenerate fpindex.pkl and matrix.pkl, then train:

python scripts/train.py configs/dev_smiles_diffusion.yml

Tweak the batch size and number of epochs according to your hardware.

2.2 Data assets

Asset	Location	Notes
Building blocks	data/building_blocks/	Raw Enamine REAL US-in-stock catalogue. Request access via Enamine.
Processed block data (fpindex.pkl, matrix.pkl)	data/processed/comp_2048/	Fingerprint index and reaction template matrix, derived from building blocks.
Protein embeddings (*.pth)	data/protein_embeddings/	ESM-C 600 M per-residue embeddings stored in 16-bit PyTorch format.
Synthetic routes (*.pth)	data/synthetic_pathways/	Pickled synthetic pathways from SynFormer runs; used for supervised pretraining or evaluation.
Protein–molecule pairs (*.csv)	data/protein_molecule_pairs/	Tabular file containing known binding pairs for training and evaluation.
Trained checkpoints	data/trained_weights/	Pretrained model weights: both small (~17 M) and big (~178 M) model variants available.

Note: For access to the synthetic pathways, protein embeddings, trained weights and protein molecule pairs, please contact the maintainers of this repository directly.

Download links are available on our Hugging Face hub.

---

3. Quick start

3.1 Training

python scripts/train.py configs/prot2drug.yml \
  --batch-size 128 \
  --devices 1 \
  --num-workers 8 \
  --log-dir ./logs

Key flags:

• --debug – runs a single forward & backward pass to sanity‑check the pipeline.
• --resume <ckpt> – warms‑start the decoder and head weights from a checkpoint (see in‑script comments for fine‑tuning logic with LoRA or last‑N‑layers freeze).
• --seed, --num-nodes, --num-sanity-val-steps – exposed for reproducibility & multi‑node training.

Tip 💡 batch_size must be divisible by devices.

---

3.2 Hyper‑parameter optimisation

python scripts/hyperopt.py configs/prot2drug_big.yml \
  --n-trials 100 \
  --batch-size 196 \
  --devices 1 \
  --n-jobs 4        # run four Optuna trials in parallel

---

3.3 Inference

python sample.py \
  --model-path checkpoints/prot2drug_big.ckpt \
  --protein-fasta P12345.fasta \
  --output results/P12345_routes.csv

By default the script searches data/protein_embeddings/ for a pre‑computed embedding; otherwise it will generate one on‑the‑fly (GPU recommended). Sampling parameters (beam width, temperatures, max steps) can be modified via CLI – run sample.py --help.

---

3.4 Evaluation

python scripts/evaluate.py data/trained_weights/epoch=23-step=28076.ckpt \
  --repeat 100 \
  --n-examples 50

---

4. Citation

If you use Prot2Drug‑SynFormer in your research, please cite:

@mastersthesis{euskirchen2025prot2drug,
  title = {x},
  author = {Janik Euskirchen, Gerben Prikanowski, Casper Bröcheler, Humam Ahmed},
  school = {Maastricht University},
  year = {2025}
}

@article{gao2024synformer,
  title={Generative Artificial Intelligence for Navigating Synthesizable Chemical Space},
  author={Gao, Wenhao and Luo, Shitong and Coley, Connor W.},
  year={2024},
  note={arXiv:2410.03494}
}

---

5. License

This fork inherits the MIT licence from SynFormer. Use of the Enamine REAL building blocks remains subject to Enamine’s terms.

---

Happy synthesizing!