This is the chemical VAE as presented in this paper. The only difference is that we're not jointly training a mol property predictor.
The files need to contain a single SMILES per row, no trailing comma.
Header needs to be SMILES.
We run with the standard hyperparameters, except for the max KL weight (β), which we set to 1.0 to get
a more disentangled latent space.
Dimension of embedding: 128
# header
echo "SMILES" > ../lincs_trapnell_zinc.csv
# add train SMILES to file
tail -n +2 ../lincs_trapnell.smiles >> ../lincs_trapnell_zinc.csv # skip the first line (`smiles`)
cat ../zinc_smiles_train.csv >> ../lincs_trapnell_zincs.csv
# start the training
./train_chemvae.sh