LatentGAN [1] with heteroencoder trained on ChEMBL 25 [2], which encodes SMILES strings into latent vector representations of size 512. A Wasserstein Generative Adversarial network with Gradient Penalty [3] is then trained to generate latent vectors resembling that of the training set, which are then decoded using the heteroencoder. A version of this model is also available in the MOSES platform [5]

The code has been confirmed to work with the environment provided in the environment.yml provided, but not all packages might be necessary.

This model uses the heteroencoder available as a package from [4]. The heteroencoder further requires this package to run properly on the ChEMBL model. Neither of these are available from e.g. anaconda or pip, and have to be installed manually. IMPORTANT: The Deep Drug Coder on the main branch runs with tensorflow 2.0, and is incompatible with the pretrained models contained here. Please use the moses branch of this model.

1. Have Git LFS installed, available from conda.
2. Initialize Git LFS with git lfs install.
3. Clone this repo using git clone https://github.com/Dierme/latent-gan.git
4. Create conda environment from .yml file conda env create --file environment.yml
5. Download dependencies Deep Drug Coder and molvecgen
6. move Deep-Drug-Coder/ddc_pub/ and molvecgen/molvecgen to the root folder (to solve import references from the Deep-Drug-Coder)

mv Deep-Drug-Coder/ddc_pub/ .
mv molvecgen/molvecgen tmp/
mv tmp/ molvecgen/

• A complete file setup after cloning the repo for installing the dependencies, setting up the file structure and installing a conda environment with the name latent_gan_env can be easily done by using the shell script

4. Use shell script bash install-dependencies.sh

One can use the runfile (run.py) for a single script that does the entire process from encoding smiles to create a training set, create and train a model, followed by sampling and decoding latent vectors back into SMILES using default hyperparameters.

Arguments:
-sf <Input SMILES file name>
-st <Output storage directory path [DEFAULT:"storage/example/"]>
-lf <Output latent file name [DEFAULT:"encoded_smiles.latent"], this will be put in the storage directory>
-ds <Output decoded SMILES file name [DEFAULT:"decoded_smiles.csv"], this will be put in the storage directory>
--n-epochs <Number of epochs to train the model for [DEFAULT: 2000]>
--sample-n <Give how many latent vectors for the model to sample after the last epoch has finished training. Default: 30000>
--encoder <The data set the pre-trained heteroencoder has been trained on [chembl|moses] [DEFAULT:chembl]> IMPORTANT: The ChEMBL-trained heteroencoder is NOT intended for use with the MOSES SMILES, just as the moses-trained model is not intended for use with the ChEMBL based SMILES files.

Note that tensorflow outputs a large amount of logging messages when the autoencoder is loaded, which can be overwhelming for a new user. If you want to limit these, we recommend you to use run.py with the following environmental variable

TF_CPP_MIN_LOG_LEVEL=2 python run.py

OR one can conduct the individual steps by using each script in succesion. This is useful to e.g. sample a saved model checkpoint using (sample.py).

1. Encode SMILES (encode.py): Gives a .latent file of latent vectors from a given SMILES file. Currently only accepts SMILES of token size smaller than 128.

Arguments:
-sf <Input SMILES file name>
-o <Output Smiles file name>

2. Create Model (create_model.py): Creates blank model files generator.txt and discriminator.txt based on an input .latent file.

Arguments: 
-i <.latent file> 
-o <path to directory you want to place the models in> 

3. Train Model (train_model.py): Trains generator/discriminator with the specified parameters. Will also create .json logfiles of generator and discriminator losses.

Arguments:
-i <.latent file> 
-o model <directory path>
--n-epochs <Number of epochs to train for>
--starting-epoch <Model checkpoint epoch to start training from, if checkpoints exist> 
--batch-size <Batch size of latent vectors, Default: 64> 
--save-interval <How often to save model checkpoints> 
--sample-after-training <Give how many latent vectors for the model to sample after the last epoch has finished training. Default: 0>
--decode-mols-save-path <Give output path for SMILES file if you want your sampled latent vectors decoded> 
--n-critic-number <Number of of times discriminator will train between each generator number. Default: 5>
--lr <learning rate, Default: 2e-4> 
--b1,--b2 <ADAM optimizer constants. Default 0.5 and 0.9, respectively>
-m <Message to print into the logfile> 

4. Sample Model (sample.py): Samples an already trained model for a given number of latent vectors.

Arguments: 
-l <input generator checkpoint file> 
-olf <path to output .latent file -n number of latent vectors to sample> 
-d <Option to also decode the latent vectors to SMILES> 
-odsf <output path to SMILES file> 
-m <message to print in logfile>

5. Decode Model (decode.py) decodes a .latent file to SMILES.

Arguments 
-l <input .latent file> 
-o <output SMILES file path> 
-m <message to print in logfile>

[1] A De Novo Molecular Generation Method Using Latent Vector Based Generative Adversarial Network

[2] ChEMBL

[3] Improved training of Wasserstein GANs

[4] Deep-Drug-Coder

[5] Molecular Sets (MOSES): A benchmarking platform for molecular generation models