Deep drug design for molecular scaffold hopping
Year: 2026
Type: preprint
Abstract: Structure-based virtual screening serves as a primary approach for ligand prioritization from large molecular libraries, but accurately predicting the activity of chemical entities on novel drug targets remains a formidable challenge. Here, we present an interactome-based graph transformer network for rapid bioactivity prediction and de novo ligand design. Integrating three-dimensional pocket geometry with two-dimensional ligand graphs yields a joint representation for direct affinity prediction, which generalizes well across sparse datasets. To validate practical utility, we performed a zero-shot generative ligand design to identify scaffold hops from the complex natural product ibogaine, a noncompetitive inhibitor of the serotonin transporter. The top-ranked design and its rational analogs were synthesized and exhibited equipotent isofunctional in vitro activity alongside favorable pharmacokinetic properties directly upon initial generation. This workflow establishes a blueprint for molecular engineering that seamlessly bridges the automated design of lead-like, synthetically accessible small molecules with efficient bioactivity prediction.
Source: ChemRxiv
Institutions: ETH Zurich, F. Hoffmann-La Roche AG
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7208705350