This is the repository for our 2022 ICML paper "LIMO: Latent Inceptionism for Targeted Molecule Generation" by Peter Eckmann, Kunyang Sun, Bo Zhao, Mudong Feng, Michael K. Gilson, and Rose Yu.
You can run LIMO online at https://www.limo-aimd.com/.
Please ensure that RDKit and Open Babel are installed. The following Python packages are also required (these can also be installed with pip install -r requirements.txt):
torch
pytorch-lightning
selfies
scipy
tqdm
Code was tested with Python 3.9, but will likely work on any version of Python 3.
Call python generate_molecules.py to generate molecules with desired properties. Run --help to see the full list of supported properties and other parameters. The default setting is to perform multi-objective binding affinity maximiziation with the filtering step, but other properties can be optimized by specifying --prop (see "Training the property predictor" to optimize your own properties). Model files for penalized logP, binding affinity to ESR1 and ACAA1, QED, and SA are provided in the property_models folder. The default binding affinity model is ESR1 (which is binding_affinity.pt), but to optimize binding to another protein one must make sure the binding_affinity.pt file contains the model for that correct protein. For example, for 2IIK binding optimization one must remove/rename the original binding_affinity.pt file and rename 2iik_binding_affinity.pt to binding_affinity.pt. A trained VAE model is provided that was trained on the ZINC250K dataset, but any SMILES dataset can be used for training (see "Training the VAE").
To optimize molecules for binding affinity, an AutoDock-GPU executable must be compiled, and pointed to with the --autodock_executable flag when generating molecules, training the property predictor, or fine-tuning. A --protein_file must also be specified, but files for ESR1 (1err/1err.maps.fld) and ACAA1 (2iik/2iik.maps.fld) are already supplied. To generate your own protein files, see Steps 1-2 in the AutoDock4 User Guide. The AutoDock-GPU Wiki may also be helpful.
Run python fine-tune "<smiles>" to fine-tune a molecule for increased binding affinity. For each iteration, it will print out the binding affinity and SMILES of the best molecule. Run with 2>/dev/null to silence RDKit warnings.
Train the property predictor with python train_property_predictor.py --prop {logp, penalized_logp, qed, sa, binding_affinity}. Run with --help for a full list of options. When training for binding affinity, you must also provide a --protein_file, with the default being 1err/1err.maps.fld. For properties other than binding affinity, we suggest training with --num_mols 100000 for greater prediction accuracy. The printed r value should be >0.5 for most properties (except binding affinity, which is typically a lower value because it is a more difficult task), so running multiple times until you reach a suitable r value or training with more molecules is recommended.
Before training, you must run python preprocess_data.py --smiles <file>.smi to get data ready. You can train with your own SMILES file, or use the provided zinc250k.smi, which has SMILES for the ZINC250k dataset. Then call python train_vae.py, which will save the trained model as vae.pt.
@article{eckmann2022limo,
title={LIMO: Latent Inceptionism for Targeted Molecule Generation},
author={Eckmann, Peter and Sun, Kunyang and Zhao, Bo and Feng, Mudong and Gilson, Michael K and Yu, Rose},
booktitle={International Conference on Machine Learning},
organization={PMLR},
Year = {2022}
}