New in REINVENT 4.4
For details see CHANGELOG.md.
- Transformer based Libinvent
- Prior registry to load internal priors more easily
- Strict validation of input configuration to ensure consistency
- Better JSON configuration file writing
- Metadata writing for all created RL and TL models
- Import functionality for scoring runmode
- Stages in staged learning can have their own diversity filters
- More memory efficient transformer models to handle larger numbers of input SMILES
- Additional (fragment) SMILES written to staged learning CSV
- TanimotoDistance renamed to TanimotoSimilarity
- Support for ChemProp multitask models: requires param.target_column
- Allow dot SMILES fragment separator for Lib/Linkinvent input
- Optional [scheduler] section for TL
- Example support script for RAScore
- A more complete RL/TL demo notebook
- Experimental data pipeline to preprocess SMILES for prior creation
- Various code improvements and fixes