#2 · ML / AI methods

Physical and quantum-derived references inform learned models of molecular properties

The dominant thread is molecular property prediction in which learned models are anchored to physical or quantum-derived information rather than fitted purely to data. The halogen-π study extends a QM-AI neural network strategy to phenol, imidazole, and indole models of Tyr, His, and Trp, trained on reference data reported at 18M+. TripKa is described as quantum-informed, using triplet interaction networks and physical knowledge to predict aqueous acid-base dissociation, and it outperforms prior methods on SAMPL6-8 and Novartis benchmarks. Other items are less explicitly physical: MURNet fuses descriptors, 2D graphs, and fingerprints for PFAS plasma protein binding, ToxCML combines consensus QSAR with read-across across 18 toxicity endpoints, and the BACE1 preprint integrates meta-ensemble QSAR with docking and molecular dynamics. What is new is the explicit injection of physical or quantum-derived structure into learned models of energies, interactions, and ionization properties.