Update note of v_0.20
1 - Train data expended to 77M molecules
2 - Conditions includs drug-likeness and synthetic accessibility score
3 - Incresed model depth and width
3 - Incresed maxium molecule generation length to 220
Introduction:
This app is based on training of a toy dataset for proof-of-concept with more than 1.8 million
molecules taken from PubChem database and filtered by the molecule string lenghth
less than 80 chracters.
https://pubchem.ncbi.nlm.nih.gov/
The max-length for generation is limited to 100, so big molecules will not be
generated.
The models is based on autiregerasive transformer with attention mechanism applied (GPT like).
Conditions for generation is applied for facilitate generate desired molecules.
After generation the molecules is shown and calculted using
RDkit
The project is to demostract a proof of concep of conditional GPT generator for molecules
design.