Mol2Pro

Code for the paper:

Generalise or Memorise? Benchmarking Ligand-Conditioned Protein Generation from Sequence-Only Data

Generate protein binder sequences from ligand SMILES using an encoder–decoder Protein Language Model.

Requirements

• Python 3.10+
• torch, transformers, numpy

Pretrained Models

The pretrained models presented in the paper are available at Hugging Face

Running inference

Input: a text file with one SMILES per line (or a JSON/JSONL file with translation entries).

Example — model and tokenizers on Hugging Face:

python src/inference.py \
  --model_path AI4PD/Mol2Pro-base \
  --tokenizer_aa AI4PD/Mol2Pro-tokenizer \
  --tokenizer_mol AI4PD/Mol2Pro-tokenizer \
  --input_file data/smiles.txt \
  --output_folder fastas \
  --top_k 15 \
  --seed 0

Outputs: one FASTA per input SMILES (25 sequences per molecule by default), plus inference_metadata.json (per-FASTA SMILES, perplexities, optional ground-truth sequences; run-level greedy perplexity stats) and generation_parameters.json in --output_folder.

Citation

If you find this work useful, please cite:

@article{VicenteSola2026Generalise,
  title   = {Generalise or Memorise? Benchmarking Ligand-Conditioned Protein Generation from Sequence-Only Data},
  author  = {Vicente-Sola, Alex and Dornfeld, Lars and Coines, Joan and Ferruz, Noelia},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.02.06.704305},
}