CheMLT-F is a family of pre-trained multitask Transformer models based on DeBERTa for chemical analysis, drug property prediction, and binding affinity prediction. The models are designed to be extendable to new datasets and easy to adapt, retrain, and evaluate through a standardized training pipeline.
Currently, the models support 13 benchmark datasets spanning toxicity and bioactivity, physicochemical property prediction, and binding affinity prediction, with 680+ prediction points across tasks: ToxCast, SIDER, MUV, Tox21, ClinTox, HIV, BACE, BBBP, Lipophilicity (Lipo), Delaney (ESOL), FreeSolv, KIBA, and Davis.
This release provides the full CheMLT-F model stack: pretrained encoder weights plus trained MTL weights for the final models.
