MMPT Transformer

Repository containing tools, training/inference code, and analysis for the MMPT project.

Overview

This repo organizes code and data used to train, run inference with, and analyze MMPT models. It includes training and inference scripts, example data, analysis notebooks.

Repository layout

• model/ - Training and inference scripts
  • model/train_MMPT.py
  • model/inference_MMPT.py
• analysis/ - Jupyter notebooks and analysis utilities
  • analysis/analysis_pmv17.ipynb
  • analysis/analysis_pmv17_pmv21.ipynb
  • analysis/calc_properties.py
• data/ - Example input data used by scripts and notebooks
  • chembl_250320_MVR33so.RGP.csv
  • chembl_neutralized_250320_smiles.txt
  • pmv17.mmp.csv
  • pmv_2017_to_pmv_2021_mmps.csv

See the LICENSE file for licensing terms.

Quick start

1. Create and activate a Python virtual environment:

python -m venv .venv
source .venv/bin/activate

2. Install runtime dependencies:

pip install -r requirements.txt

Note: the requirements.txt file has pinned versions to ensure reproducible environments. This ensures the code runs consistently across different systems and time periods.

Common tasks

Key training flags available in model/train_MMPT.py include:

• --model_type (T5 | GPT | T5Chem | RealT5Chem)
• --dataset (dataset name)
• --epochs, --batch_size, --learning_rate, --output_dir
• --use_wandb to enable Weights & Biases logging

Key inference flags available in model/inference_MMPT.py include:

• --model_type, --model_path (required)
• --dataset, --batch_size, --max_length
• --num_beams, --num_return_sequences, --early_stopping

Adjust num_beams and num_return_sequences so that num_beams >= num_return_sequences.

Notebooks and analysis

Open the notebooks in analysis/ for visualizations and exploratory analyses. Example notebooks:

• analysis/analysis_pmv17.ipynb
• analysis/analysis_pmv17_pmv21.ipynb

Note: The analysis notebooks require external sample data files (model outputs, REINVENT results) that are not included in this repository. To use these notebooks:

1. Prepare or obtain the required sample data files
2. Update the DATA_ROOT configuration variable at the top of each notebook to point to your data directory
3. Run the notebook cells with your configured data path

The analysis utilities (analysis/calc_properties.py, analysis/canonicalize.py, analysis/get_umap.py) are provided for post-processing model outputs and can be adapted for your own analysis workflows.

To start a notebook server locally:

pip install jupyterlab
jupyter lab

Data

Some exemplary datasets are stored in data/.

Troubleshooting

• GPU not detected: make sure CUDA and a compatible torch build are installed. Check torch.cuda.is_available().
• Tokenizer / model errors: ensure --model_path points to the trained model directory containing the tokenizer and model weights.

Contact

If you need help or want changes to this README (commands, examples, or extra sections), open an issue.

License

This project is licensed under the terms in the LICENSE file.