FLOWR: Flow Matching for Structure-Aware De Novo, Interaction- and Fragment-Based Ligand Generation

December 5, 2025 · View on GitHub

FLOWR Overview

FLOWR is a research repository that investigates continuous and discrete flow matching methods applied to structure-based drug discovery. It provides a complete workflow for training models, generating novel ligand molecules, and evaluating the generated structures.

⚠️ Note: This repository is no longer actively maintained and is provided primarily for reproducing the results from our paper. For continued development and new features, please use our follow-up work FLOWR.root.

Installation
Getting Started
Contributing
License
Citation

Installation

GPU: CUDA-compatible GPU with at least 40GB VRAM recommended for training and generation

Package Manager: mamba
Install via:

curl -L -O https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-$(uname)-$(uname -m).sh
bash Miniforge3-$(uname)-$(uname -m).sh

Create the Environment
Install the required environment using mamba:
```
mamba env create -f environment.yml
```
Activate the Environment
```
conda activate flowr
```
Set PYTHONPATH
Ensure the repository directory is in your Python path:
```
export PYTHONPATH="$PWD"
```
Install ADFR Install the ADFR suite. Adjust the prepare_receptor_bin_path in genbench3d/configs/default.yaml to match the ADFR installation.

Getting Started

We provide the full SPINDR data in both .smol and .cif format, as well as a fully trained FLOWR model checkpoint and generated samples. For training, generation and evaluation, we provide basic bash and SLURM scripts in the scripts/ directory. These scripts are intended to be modified and adjusted according to your computational resources and experimental needs.

Checkpoints

There are two checkpoints, one model trained without and one with explicit hydrogens. Download the FLOWR checkpoints here: Zenodo.

Data

Download the SPINDR dataset (and generated samples if needed) here: Zenodo.

The 'raw' SPINDR data for both pocket and ligand data can be downloaded here: Zenodo.

Download the genbench3d_data.tar and place it untarred in the repo (we will get rid of this dependency in the future).

To train a model, unzip the smol_data.zip to get the smol-files. Specify the directory they are placed in the respective scripts (see below). We also provide the cif-files for all protein pockets splitted into train, validation and test.

Generation of novel molecules either de novo or fragment-based can be done with the provided checkpoint (flowr.ckpt).

Training the Model

Start by training the model using the provided training script. This script sets hyperparameters such as batch size, learning rate, and network architecture.

Modify scripts/train_spindr.sh as needed, then run:

bash scripts/train_spindr.sh

Generating Molecules

After training (or with the provided flowr.ckpt), generate novel molecules using the provided generation script.

Modify scripts/gen_spindr.sl according to your requirements, then submit the job via SLURM:

sbatch scripts/gen_spindr.sl

Generating Molecules from PDB/CIF

After training (or with the provided flowr.ckpt), you can generate novel molecules given a PDB or CIF protein/pocket file.

If you provide a protein PDB/CIF file, you need to provide a ligand file as well to cut out the pocket (default: 6A cutoff - modify if needed). We recommend using (Schrödinger-)prepared complexes for best results with the protein and ligand being protonated.

Note, if you want to run conditional generation, you need to provide a ligand file as reference.

Modify scripts/gen_pdb.sl according to your requirements, then submit the job via SLURM:

sbatch scripts/gen_pdb.sl

Generated ligands are saved as an SDF file at the specified location (save_dir) alongside the extracted pockets.

Evaluating Molecules

Evaluate the generated molecules using the evaluation script. This step calculates metrics including molecular validity, uniqueness, and interaction recovery.

Modify scripts/eval_spindr.sh as needed, then run:

bash scripts/eval_spindr.sh

@misc{cremer2025flowrflowmatchingstructureaware,
      title={FLOWR: Flow Matching for Structure-Aware De Novo, Interaction- and Fragment-Based Ligand Generation}, 
      author={Julian Cremer and Ross Irwin and Alessandro Tibot and Jon Paul Janet and Simon Olsson and Djork-Arné Clevert},
      year={2025},
      eprint={2504.10564},
      archivePrefix={arXiv},
      primaryClass={q-bio.QM},
      url={https://arxiv.org/abs/2504.10564}, 
}