In this project, we aim to explore the potential of transformer models to optimize molecular properties and/or generate new molecules with desired non-toxic properties.

Unlike currently available machine and deep learning methods, self supervised learning models, e.g., the transformer architecture, provide generalizability by being pre-trained on large unsupervised dataset then fine-tuned on small downstream datasets. The transformer model is highly resourceful with the ability to perform molecular property prediction, optimization, and/or generation.


  • Miriam Mathea · (BASF)
  • Jochen Sieg · (BASF)


  • We thank BASF for financial support