Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
BP: Fachverband Biologische Physik
BP 14: Poster Session II
BP 14.63: Poster
Dienstag, 10. März 2026, 18:00–21:00, P2
Assessing the performance of quantum-mechanical descriptorsin physicochemical and biological property prediction — •Alejandra Hinostroza Caldas1, Artem Kokorin2, Alexandre Tkatchenko2, and Leonardo Medrano Sandonas1 — 1TUD Dresden University of Technology, 01069 Dresden, Germany — 2University of Luxembourg, L-1511 Luxembourg City, Luxembourg
Understanding how molecular structure relates to physicochemical and biological properties is essential for computer-aided drug design. A major challenge in applying machine learning (ML) to this problem is defining numerical representations that capture both geometric and electronic information. We introduce the QUantum Electronic Descriptor (QUED) framework [ChemRxiv, doi:10.26434/chemrxiv-2025-hj4dc], which integrates geometric descriptors (BOB, SLATM) with a quantum-mechanical descriptor (DQM) that encodes global and local electronic properties computed efficiently with the DFTB method. We validate QUED on the QM7-X dataset of small drug-like molecules and show that incorporating electronic structure information substantially improves ML predictions of physicochemical properties. For biological endpoints, QUED also demonstrates predictive value for acute toxicity (LD50) in TDCommons-LD50 and for lipophilicity in the MoleculeNet benchmark. Our findings underscore the benefits of integrating electronic structure information with geometric descriptors and highlight the role of conformational diversity in improving the robustness of molecular property prediction.
Keywords: property prediction; ADMET endpoints; physicochemical properties; machine learning; quantum mechanics