Fig. 2
From: Fate-tox: fragment attention transformer for E(3)-equivariant multi-organ toxicity prediction
Model Architecture of FATE-Tox. The model is roughly composed of two branches: Atom-level and Fragment-level. Each branch passes down the input features and coordinates into a two-component pipeline composed of Molecular transformer module and the E(3)-Equivariant Graph Convolutional Layer. After the atom-level and aggregated fragment-level representations are generated, they are concatentated with ECFP, then passed down to the prediction MLP to yield the prediction for the toxicity label