MOSES: Molecular Sets

TODO: Abstract

Dataset

TODO: Description of dataset

Models

TODO: Check links in models

• Adversarial Autoencoder (AAE)
• Char-RNN (CRNN)
• Junction Tree Variational Autoencoder (JT)
• Objective-Reinforced Generative Adversarial Network (ORGAN)
• Variational Autoencoder (VAE)

Metrics

| Model | Valid | Unique@1k | Unique@10k | FCD | Morgan | Fragments | Scaffolds | LogP | SA | QED | NP | Weight | Internal Diversity | Filters | |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: |:---: | | AAE (scaffolds) | | | | | | | | | | | | | | | | | | | | | | | | | CRNN (scaffolds) | | | | | | | | | | | | | | | | | | | | | | | | | JT (scaffolds) | | | | | | | | | | | | | | | | | | | | | | | | | ORGAN (scaffolds) | | | | | | | | | | | | | | | | | | | | | | | | | VAE (scaffolds) | | | | | | | | | | | | | | | | | | | | | | | |

Description of models

TODO: Check this

• AAE: 1-layer Bi-LSTM (380 hidden size) as encoder and 2-layer LSTM (640 hidden size) as decoder, shared embeddings with size 32. Latent size - 640. Discriminator - MLP (2 layer - 640, 256) with ELU activation. Batch size - 128, number of epochs - 25, lr - 1e-3, optimizer - Adam.

• CRNN: 3-layer LSTM with 600 hidden each, and everyone followed by a dropout layer, with a dropout ratio of 0.2, and a softmax layer on top. Training was with batch size of 64, Adam optimizer with learning rate of 1e-3 for 50 epochs.

• JT: Training was with batch size of 40, KL term weight of 0.005 and Adam optimizer with learning rate of 1e-3 for 5 epochs. KL term was taken into consideration starting from second epoch, i.e we trained model as just autoencoder one epoch. Other parameters were taken from original paper: hidden size is 450, latent dimensionality is 56 and depth of graph message passing is 3.

• ORGAN: TODO: Description of parameters

• VAE: 1-layer bidirectional GRU as encoder with linears at the end, predicting latent space of size 128 distribution parameters. 3-layers GRU decoder with dropout of 0.2 and 512 hidden dimensionality. Training was with batch size of 128, gradients clipping of 50, KL term weight of 1 and Adam optimizer with learning rate of 3 * 1e-4 for 50 epochs.

Calculation of metrics for all models

You can calculate all metrics with:

cd scripts
python run.py 

If necessary, dataset will be downloaded, splited and all models will be trained. As result in current directory will appear metrics.csv with values. For more details use python run.py --help.

Installation

• Install RDKit for metric calculation.
• Install models with python setup.py install

Usage

Downloading of dataset

You can download dataset (and split it) with:

cd scripts
python download_dataset.py --output_dir <directory for dataset>

For more details use python download_dataset.py --help.

Training of model

You can train model with:

cd scripts/<model name>
python train.py --train_load <path to train dataset> --model_save <path to model> --config_save <path to config> --vocab_save <path to vocabulary>

For more details use python train.py --help.

Calculation of metrics for trained model

You can calculate metrics with:

cd scripts/<model name>
python sample.py --model_load <path to model> --config_load <path to config> --vocab_load <path to vocabulary> --n_samples <number of smiles> --gen_save <path to generated smiles>
cd ../metrics
python eval.py --ref_path <path to referenced smiles> --gen_path <path to generated smiles>

All metrics output to screen. For more details use python sample.py --help and python eval.py --help.

You also can use python run.py --model <model name> for calculation metrics.