Efficient Molecular Conformer Generation with SO(3) Averaged Flow-Matching and Reflow

This is the official code repository for the paper titled Efficient Molecular Conformer Generation with SO(3) Averaged Flow-Matching and Reflow (ICML 2025).

Contribution

• We propose SO(3)-Averaged Flow: A novel flow-matching objective that analytically computes the probability flow from noise to all rotations of the data. When the "correctness" of samples is rotational invariant (such as conformer generation), SO(3)-Averaged Flow improves training efficiency by eliminating the need for rotational data augmentation and further improves model performance.
• We propose to use reflow+distillation to reduce the number of sampling steps of flow-based model for conformer generation and maintain high generation quality.
• We provide a JAX implementation of the diffusion transformer with pairwise biased attention architecture. It is powerful and scalable for generative modeling of molecules.

Installation

Clone this repository:

git clone https://github.com/NVIDIA-Digital-Bio/avgflow.git
cd avgflow

Run the following command to create a conda environment and install the dependencies:

conda env create -f env.yml
conda activate avgflow

Pretrained Checkpoints

We provide 4 model weights through the NVIDIA NGC, including:

1. 52M DiT trained with AvgFlow objective (Link)
2. 52M DiT finetuned with reflow for few-step generation (Link)
3. 52M DiT finetuned with reflow+distillation for 1-step generation (Link)
4. 64M DiT trained with AvgFlow objective. (Link)

If you have NGC CLI tool installed, you can run the following command to download the checkpoints:

bash scripts/download_ckpts.sh

Otherwise, you can create a checkpoints directory by:

mkdir -p checkpoints

and download the checkpoints from the NGC pages above.

Sampling

The model can be used to generate conformers given: 1. a single SMILES string, or 2. a CSV file containing a batch of SMILES strings and the number of conformers to be generated for each molecule.

For conformer generation of a single molecule, run:

python avgflow/generate_from_smiles.py \
--config PATH/TO/CONFIG.yaml \
--smiles SMILES_STRING \
--num_confs N_CONF \
--output_dir PATH/TO/OUTPUT/DIRECTORY \

Example can be found in example/sampling/gen_smiles.sh, which generates 40 conformers for molecule C#CCNC(=O)C1=C[C@@H](c2ccc(Br)cc2)C[C@@H](OCc2ccc(CO)cc2)O1.

For conformer generation of a batch of molecules in a csv file, run:

python avgflow/generate_from_csv.py \
--config PATH/TO/CONFIG.yaml \
--smiles_csv PATH/TO/SMILES.csv \
--output_dir PATH/TO/OUTPUT/DIRECTORY \

Example can be found in example/sampling/gen_csv.sh, which generate various number of conformers for molecules in example/data/toy_gen_csv.csv. Please follow the format of example/data/toy_gen_csv.csv to construct your own csv for sampling.

The config yaml files define the model architecture to be initialized and checkpoint to be loaded. We provide 4 config files for the 4 checkpoints we released:

1. config/generation_config/avgflow_52m_gen.yaml for the 52M DiT trained with AvgFlow objective.
2. config/generation_config/avgflow_64m_gen.yaml for the 64M DiT trained with AvgFlow objective.
3. config/generation_config/avgflow_52m_reflow_gen.yaml for the 52M DiT finetuned with reflow for few-step generation.
4. config/generation_config/avgflow_52m_distill_gen.yaml for the 52M DiT finetuned with reflow+distillation for 1-step generation.

Please choose the config based on your checkpoint choice, and note that the distilled checkpoint only works with 1-step generation.

Training

Preparation of training data

Each molecule with ground truth conformers in the dataset has to be preprocessed before training. We recommend to have a dictionary for each molecule that contains at least 2 keys:

1. features: Features computed from the 2D molecular graph using data_preprocessing.preprocess.mol2features
2. conformers: np.array with dimension [C, N, 3], where C is the number of conformers and N is the number of atoms in the molecule.

For reflow/distill finetuning which requires ($X'_0$, $X'_1$) pairs, we recommend the preprocessed dictionary to contain 2 other keys instead of conformer:

1. x0s: np.array with dimension [C, N, 3], where C is the number of ($X'_0$, $X'_1$) pairs and N is the number of atoms in the molecule. Gaussian noise at $t=0$.
2. x1s: np.array with dimension [C, N, 3], where C is the number of ($X'_0$, $X'_1$) pairs and N is the number of atoms in the molecule. Model generated conformer at $t=1$ from each corresponding x0s.

Please refer to example/data/generate_toy_dataset.ipynb for example of creating the toy training and finetuning datasets.

Launch training

Follow the following steps to launch training:

1. Prepare dataset as illustrated above.
2. Prepare training config. See example in config/train_config/avgflow_52m_train_toy.yaml.
3. (Optional) Change how the preprocessed dataset is loaded in avgflow/train.py line 45-58. You may parallelize the data loading for large training dataset.
4. Launch training (see example in example/train/train_toy.sh) with:

python avgflow/train.py --config PATH/TO/CONFIG.yaml 

Follow the same procedure and use avgflow/reflow_finetune.py for finetuning.

License

Copyright @ 2025, NVIDIA Corporation. All rights reserved.
The source code is made available under Apache-2.0.
The model weights are made available under the NVIDIA Open Model License.

Citation

If you find this repository and our paper useful, please cite our work through:

@article{cao2025efficient,
title = {Efficient Molecular Conformer Generation with SO (3)-Averaged Flow Matching and Reflow},
author = {Cao, Zhonglin and Geiger, Mario and Costa, Allan Dos Santos and Reidenbach, Danny and Kreis, Karsten and Geffner, Tomas and Pellegrini, Franco and Zhou, Guoqing and Kucukbenli, Emine},
journal = {arXiv preprint arXiv:2507.09785},
year = {2025}
}

Disclaimer

This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.