Official code release for the paper "Geometric Latent Diffusion Models for 3D Molecule Generation", accepted at International Conference on Machine Learning, 2023.

Install the required packages from requirements.txt. A simplified version of the requirements can be found here.

Note: If you want to set up a rdkit environment, it may be easiest to install conda and run: conda create -c conda-forge -n my-rdkit-env rdkit and then install the other required packages. But the code should still run without rdkit installed though.

python main_qm9.py --n_epochs 3000 --n_stability_samples 1000 --diffusion_noise_schedule polynomial_2 --diffusion_noise_precision 1e-5 --diffusion_steps 1000 --diffusion_loss_type l2 --batch_size 64 --nf 256 --n_layers 9 --lr 1e-4 --normalize_factors [1,4,10] --test_epochs 20 --ema_decay 0.9999 --train_diffusion --trainable_ae --latent_nf 1 --exp_name geoldm_qm9

First follow the intructions at data/geom/README.md to set up the data.

python main_geom_drugs.py --n_epochs 3000 --n_stability_samples 500 --diffusion_noise_schedule polynomial_2 --diffusion_steps 1000 --diffusion_noise_precision 1e-5 --diffusion_loss_type l2 --batch_size 32 --nf 256 --n_layers 4 --lr 1e-4 --normalize_factors [1,4,10] --test_epochs 1 --ema_decay 0.9999 --normalization_factor 1 --model egnn_dynamics --visualize_every_batch 10000 --train_diffusion --trainable_ae --latent_nf 2 --exp_name geoldm_drugs

Note: In the paper, we present an encoder early-stopping strategy for training the Autoencoder. However, in later experiments, we found that we can even just keep the encoder untrained and only train the decoder, which is faster and leads to similar results. Our released version uses this strategy. This phenomenon is quite interesting and we are also still actively investigating it.

We also provide pretrained models for both QM9 and Drugs. You can download them from here. The pretrained models are trained with the same hyperparameters as the above commands except that latent dimensions --latent_nf are set as 2 (the results should be roughly the same if as 1). You can load them for running the following evaluations by putting them in the outputs folder and setting the argument --model_path to the path of the pretrained model outputs/$exp_name.

To analyze the sample quality of molecules:

python eval_analyze.py --model_path outputs/$exp_name --n_samples 10_000

To visualize some molecules:

python eval_sample.py --model_path outputs/$exp_name --n_samples 10_000

Small note: The GPUs used for these experiment were pretty large. If you run out of GPU memory, try running at a smaller size.

python main_qm9.py --exp_name exp_cond_alpha --model egnn_dynamics --lr 1e-4 --nf 192 --n_layers 9 --save_model True --diffusion_steps 1000 --sin_embedding False --n_epochs 3000 --n_stability_samples 500 --diffusion_noise_schedule polynomial_2 --diffusion_noise_precision 1e-5 --dequantization deterministic --include_charges False --diffusion_loss_type l2 --batch_size 64 --normalize_factors [1,8,1] --conditioning alpha --dataset qm9_second_half --train_diffusion --trainable_ae --latent_nf 1

The argument --conditioning alpha can be set to any of the following properties: alpha, gap, homo, lumo, mu Cv. The same applies to the following commands that also depend on alpha.

python eval_conditional_qm9.py --generators_path outputs/exp_cond_alpha --property alpha --n_sweeps 10 --task qualitative

cd qm9/property_prediction
python main_qm9_prop.py --num_workers 2 --lr 5e-4 --property alpha --exp_name exp_class_alpha --model_name egnn

Additionally, you can change the argument --model_name egnn by --model_name numnodes to train a classifier baseline that classifies only based on the number of nodes.

Evaluate the trained property classifier on the samples generated by the trained conditional GeoLDM model

python eval_conditional_qm9.py --generators_path outputs/exp_cond_alpha --classifiers_path qm9/property_prediction/outputs/exp_class_alpha --property alpha --iterations 100 --batch_size 100 --task edm

Please consider citing the our paper if you find it helpful. Thank you!

@inproceedings{xu2023geometric,
  title={Geometric Latent Diffusion Models for 3D Molecule Generation},
  author={Minkai Xu and Alexander Powers and Ron Dror and Stefano Ermon and Jure Leskovec},
  booktitle={International Conference on Machine Learning},
  year={2023},
  organization={PMLR}
}

This repo is built upon the previous work EDM. Thanks to the authors for their great work!