Implementation of ProtoBind-Diff: A Structure-Free Diffusion Language Model for Protein Sequence-Conditioned Ligand Design by Lukia Mistryukova*, Vladimir Manuilov*, Konstantin Avchaciov*, and Peter O. Fedichev.

ProtoBind-Diff is a masked diffusion language model that generates target-specific, high-quality small-molecule ligands. Trained entirely without structural input, it enables structure-independent ligand design across the full proteome and matches structure-based methods in docking and Boltz-1 benchmarks.

If you have questions, feel free to open an issue or send us an email at lukiia.mistriukova@gero.ai, konstantin.avchaciov@gero.ai, vladimir.manuylov@gero.ai, and peter.fedichev@gero.ai.

You can also try out the model on Hugging Face Spaces.

This repository contains the evaluation toolkit and supporting data for ProtoBind-Diff. It is organized as follows:

We selected 12 protein targets to benchmark molecular generation quality across different models.

• One folder per model (pocket2mol/, targetdiff/, etc.) containing SMILES files of generated ligands.
• bindingdb/ – sets of random molecules (used as reference inactive molecules).
• bindingdb_active/ – sets of true active molecules.
• actives_bindingdb_cl – clustered true actives; one representative per similarity cluster.
• CrossDocked2020/ – cleaned PDB files and corresponding ligands for the targets.
• fasta/ – FASTA sequences of the targets.

Jupyter notebooks to reproduce the paper figures: docking/Boltz-1 score distributions, interpretation of attention maps, chemical property distributions, and UMAP-based target specificity analysis. We also added allign.ipynb to show differences between canonical sequences and PDB receptor sequences.

• Raw docking and Boltz-1 ipTM score tables for each method.

• Model code, train and inference scripts.

• Scripts to download and prepare data for training.

Clone the current repo

git clone https://github.com/gero-science/ProtoBind-Diff.git

You can install the project locally in editable mode:

python -m venv protodiff_env
source protodiff_env/bin/activate    
pip install -e .

Then you'll be able to run:

protobind-infer #inference
protobind-train #train

This script generates potential binding molecules (in SMILES format) for a given protein sequence by computing the protein embeddings on-the-fly.

Note: The script automatically downloads the model checkpoint from Hugging Face Hub and uses the best available hardware.

You need to specify the protein sequence with --fasta_file or --sequence. And you are ready to run inference:

protobind-infer --fasta_file examples/input.fasta

Training the model involves a three-step pipeline: 1) downloading the dataset, 2) pre-processing the data to generate embeddings and similarity matrices, and 3) running the training script with a configuration file.

First, download the necessary dataset files (protein/ligand pairs, tokenizer, etc.) from Hugging Face Hub.

Run the command:

python scripts/download.py --output_dir ./data/experiments/diffusion

This will create a directory containing the raw data.csv and other necessary files.

Next, you need to generate protein embeddings and a molecular similarity matrix from the raw data. This script performs both of these tasks. Note: This is a computationally intensive step. Calculating the ESM embeddings requires a GPU and at least 64 GB of RAM (depending on the size of the categorical_mappings.json file).

Run the command:

python scripts/esm_and_sim_matrix.py --data_dir ./data/experiments/diffusion --out_dir ./data/experiments/diffusion

This uses the files in the --data_dir and saves the new files: all_prots_*.pt (embeddings) and all_smiles_sparse_*.npz (Tanimoto matrix).

You are ready to train your model. Training is configured using .yaml files located in the configs/ directory.

protobind-train -o ./experiment_dir --exp_dir_prefix experiment_name --yaml ./configs/masked_diffusion.yaml

This command uses the following arguments:

• --yaml: Specifies the main model and data configuration file.

• --output_dir: Defines the parent directory where all experiment results will be saved.

• --exp_dir_prefix: Creates a specific folder for this run (e.g., ./experiments/my_first_run).

To tune the model, you can edit the parameters in configs/masked_diffusion.yaml. Details on all tunable hyperparameters can be found in protobind_diff/parsers.py.

If you use this code or the models in your research, please cite the following paper:

@article {Mistryukova2025.06.16.659955,
	author = {Mistryukova, Lukia and Manuilov, Vladimir and Avchaciov, Konstantin and Fedichev, Peter O.},
	title = {ProtoBind-Diff: A Structure-Free Diffusion Language Model for Protein Sequence-Conditioned Ligand Design},
	year = {2025},
	journal = {bioRxiv}
}

The code and model weights are released under MIT license. See the LICENSE file for details.