Quantum Mechanics and AI Can Help Treat Epilepsy (thepotentialsurface.substack.com)

Researchers are demonstrating that general-purpose Neural Network Potentials (NNPs) — machine-learned surrogates trained on massive quantum chemistry datasets — can simulate enzyme systems at quantum accuracy and scale, providing a new route for drug discovery for conditions like drug-resistant epilepsy. Using the OMol25 dataset (over 100 million high-level quantum calculations) and a universal model called Orbmol, teams simulated a full carbonic anhydrase (CA) enzyme system of 20,000+ atoms and captured the zinc‑centered active site chemistry that underlies CA’s regulation of neuronal pH. Because CA inhibitors bind the catalytic zinc and shift local pH to reduce neuronal excitability, correctly modelling zinc charge, spin state and proton-transfer/pKa dynamics is critical for predicting inhibitor binding and mechanism; the NNP achieved this by conditioning on electronic properties (total charge and spin multiplicity). Technically, this approach trains NNPs on expensive first‑principles data so they can predict energies and forces orders of magnitude faster — turning weeks of supercomputer time into minutes on a single GPU — while preserving quantum-level detail in complex, heterogeneous molecular environments. The result is a practical, “out‑of‑the‑box” general model that generalises to unseen biomolecular targets, enabling high‑throughput screening, longer-timescale dynamics, and richer mechanistic insight. The shift removes compute as the primary bottleneck and redirects focus to model architectures and novel applications across drug discovery and materials design.