#18 · Structural bio & protein design

Web-deployed models predict kinase selectivity, CYP metabolism, and ADME endpoints

The dominant thread is packaging predictive models as accessible, task-specific web tools for drug discovery triage. DeepKinomeWeb uses a deep learning regression model to predict kinase-inhibitor affinities and visualize panel-level selectivity. DeepCYP applies a multi-task graph neural network to CYP450 pathway-site and product prediction across nine isoforms, reported as outperforming FAME3. PLATE-VS curates assay data, structures, and property-matched decoys into virtual screening datasets. Preprints extend this toward ADME and CNS endpoints: PredHLM uses XGBoost on RDKit descriptors for liver microsome half-life across 11,790 compounds with SHAP interpretation, and BBBP_Atlas applies a structure-aware graph model to blood-brain barrier permeability. The repurposing pipeline is a distinct workflow. What is new is interpretability paired with public deployment, though comparative benefit beyond the FAME3 comparison is not established.