E(3)-equivariant models cannot learn chirality: Field-based molecular generation

December 2, 2025 · View on GitHub

Official implementation of E(3)-equivariant models cannot learn chirality: Field-based molecular generation

by Alexandru Dumitrescu, Dani Korpela, Markus Heinonen, Yogesh Verma, Valerii Iakovlev, Vikas Garg, and Harri Lähdesmäki

Install environment

Tested on python 3.9.18

run

	pip install -r requirements.txt

Paths

By default, the model and data paths from config_paths_local.yml are used. Change this file accordingly if needed.

Generated samples

The results of this paper are based on the generated molecules found under generated_molecules/geom_drugs and generated_molecules/qm9.

QM9

Data (required for both generation and training)

  • go to data/qm9/data/qm9_edm.
  • run python create_edm_data_splits.py. This will create train.npz, test.npz, valid.npz in data/qm9/data/qm9_edm/qm9.
  • run training after going back to the main project directory (from Section QM9 model training). This automatically creates correct format training file and saves it with the name specified in --data_file argument. Saved in data/qm9/data.
  • if you want to train the model, you can rerun the same training command

Generating

Pre-trained model weights: download from https://drive.google.com/drive/folders/1XpOfCPRvPu22dSgbWgfGRF0Lul7ygdC7?usp=sharing

Before generating, please follow the steps from the previous subsection (Data).

Run:

	python -u test_rdkit_metrics.py --model <model_path> --data <training_data_file> --explicit_hydrogen --timesteps 100 --cls_free_guid --small_mdl --threshold_atm 0.4 --threshold_bond 0.4 --batch_size 50 --optimize_bnd_gmm_weights --beta_schedule cosine --noise_scheduler_conf sched_configs/qm9_explh_sepSched_destrAtmsFirst.yaml --run_name <save_file_name> --discard_fields --num_imgs 300

<model_path>: path to pre-trained model

<training_data_file>: training data file (ensuring resolution, grids, and channel counts are are the same as the training). Using the instructions from QM9 data (required for both generation and training) and having run QM9 model training command, the name should be qm9_data_explicit_hydrogen_033_res.bin.

<save_file_name>: filename for the generated molecules. The path under which it is saved is the same as the "model_path" in config_paths_local.yml.

If many molecules are required, the argument --discard_fields may be handy, since the fields take a large amount of storage space.

Training

For QM9, the best configuration run is:

	python -u train_class_free_guid.py --data QM9EDM_small --batch_size 64 --beta_schedule cosine --data_file qm9_data_explicit_hydrogen_033_res --model unet --run_name test_run_name --resolution 0.33 --lr 3e-5 --explicit_hydrogen --ntimesteps 100 --compact_batch --noise_scheduler_conf sched_configs/qm9_explh_sepSched_destrAtmsFirst.yaml --use_original_atm_pos --test_every 1

GEOM-Drugs

Data (generation only, full training data available soon!)

Generating

Pre-trained model weights: download from https://drive.google.com/file/d/16S2p_ioHxnrt2KSCK40mMoTGphQTv9oX/view?usp=sharing

Before generating, please follow the steps from the previous subsection (Data).

	python generate.py --model <model_path> --batch_size 2 --threshold_atm 0.35 --threshold_bond 0.35 --optimize_bnd_gmm_weights --noise_scheduler_conf  sched_configs/geom_explh_sepSched_destrAtmsFirst.yaml --timesteps 100 --data geom_data --cls_free_guid --small_mdl  --data_type GEOM --explicit_hydrogen --optimize_atom_positions --run_name <training_data_file> --num_imgs 4 --discard_fields

<model_path>: path to pre-trained model

<save_file_name>: filename for the generated molecules. The path under which it is saved is the same as the "model_path" in config_paths_local.yml.

Training a model

Available soon!

Testing generated molecules

Available soon!