Molecule Generation using Transformers and REINFORCE
Researchers propose Taiga, a transformer-based architecture for the generation of molecules with desired properties
An article in Nature caught my attention today, "Molecule generation using transformers and policy gradient reinforcement learning".
Here Eyal Mazuz et al. from the University of the Negev, Israel, propose Taiga, a transformer-based architecture for the generation of molecules with desired properties
They show that transformers combined with policy gradient RL, REINFORCE specifically, can be used to accelerate the search for novel valid molecules. Traditionally, chemists and pharmacologists use their intuition and expertise to identify new molecules.
The language to describe the structure of chemical species is SMILES (simplified molecular-input line-entry system) and the idea to rely on generative models to create SMILE strings that describe valid molecules has a history. However previous systems produced a large percentage of invalid SMILES strings.
The paper is the first attempt to combine RL with transformer models to generate new molecules. They approach the problem in two stages
- The model learns to embed SMILE string representations in a vector space.
- The model optimizes the vector space in order to generate molecules with the desired properties.
As many said, the combination of deep learning and synthetic biology could usher in a new golden age of bio tech.