#21 · Structural bio & protein design

Learned models predict ADMET endpoints and prioritize synthesizable small molecules

The dominant thread is learned models over small-molecule representations for property prediction and candidate prioritization. Three preprints address ADMET endpoints. A quantum-inspired Hamiltonian feature set reports solubility MAE matching top TDC AqSolDB entries with 58-fold fewer interaction features. PredHLM applies XGBoost to RDKit 2D descriptors with SHAP analysis for human liver microsome half-life across 11,790 compounds. BBBP-Atlas uses a structure-aware graph model for blood-brain barrier permeability across small molecules and peptides. Published work extends this to target-directed screening, where SVM QSAR classifiers trained on ChEMBL prioritize MAO-B/Sig-1R dual ligands before docking and 150 ns MD. MGVS couples structure-based generation with chemical similarity search for synthesizable analogues, and a benchmark evaluates 2D similarity metrics for retrieval from make-on-demand spaces. What is new is the emphasis on interpretability, compact feature sets and discovery-relevant evaluation.