ÖPG-CMD Joint Meeting 2026

A hybrid machine learning and atomistic modeling approach for the design of de novo enzymes

Sep 24, 2026, 5:15 PM

30m

HS 05.01 (University of Graz)

4) Invited talk NESY: Physics of Neutron and Synchrotron Radiation Sources Parallel

Speaker

Gustav Oberdorfer (Institute of Biochemistry, Graz University of Technology, Petersgasse 12/2, A-8010 Graz, Austria)

Description

Reliably introducing function into genetically encodable de novo proteins is still a challenging task. Current design methods mostly produce de novo enzymes with low activities. As a result, they require costly experimental optimization and high-throughput screening to be industrially viable. We developed rotamer inverted fragment finder–diffusion (Riff-Diff), a hybrid machine learning and atomistic modelling strategy for scaffolding catalytic arrays in de novo protein backbones. We show that proficient enzymes can be generated with Riff-Diff while screening as little as 35 designs. Easy access to synchrotron beamlines at the ESRF enabled us to screen several hundred protein crystals and led to the determination of six protein structures of our de novo enzymes. These experimental structures revealed a counterintuitive correlation between design accuracy and catalytic activity, which I will highlight during the talk.