Trained JTVAE Autoencoder Model on ZINC Dataset

This asset contains a Junction Tree Variational Autoencoder (JTVAE) model trained on the ZINC dataset, along with the train/test split used for training and validation, the datasets vocabulary file, and a documentation file further describing the model. The JTVAE method was originally released in a 2018 paper called ���Junction Tree Variational Autoencoder for Molecular Graph Generation���. The JTVAE method is an autoencoder that is trained to encode molecules from their SMILES string representation into a latent vector then decode the vector, and then decode the vector back to the original SMILES string. This allows for the application of math-based operators on the vector representation, which can be decoded into a new SMILES string. The main purpose of this tool is to facilitate the use of generative AI methods for creating new optimized molecules. This model was trained using our implementation of the JTVAE method which can be found in the associated GitHub link. We modified the original code to implement GPU���s for training. To speed up the training process, we upgraded the code from Python2 to Python3, and increased the reconstruction accuracy. Once you have a trained JTVAE model using our source code, the model can be plugged into our Generalized Generative Molecular Design Loop. This means that anyone can use this model as an autoencoder for molecular optimization. Please refer to the documentation to learn more about getting started with our model and JTVAE code.

jtvae_zinc

Model

AutEncoder

PyTorch

None

None

Black, Sean

sean.black2@nih.gov

No

33.8 MB

Unverified