Computational enzyme design by catalytic motif scaffolding
1.A new enzyme design method called Riff-Diff combines machine learning with atomistic modeling to scaffold catalytic motifs in de novo proteins. It achieves catalytic activities approaching those of laboratory-evolved enzymes—without the need for high-throughput screening.
2.Riff-Diff robustly designs enzymes from predefined catalytic arrays, enabling the one-shot creation of active biocatalysts. Notably, 91% of the designed retro-aldolases were catalytically active, and several achieved product formation rates multiple orders of magnitude above the screen average.
3.The method was validated on two mechanistically distinct reactions: the retro-aldol cleavage of methodol and the Morita–Baylis–Hillman (MBH) reaction. Both design campaigns yielded enzymes with measurable and in some cases high activity, despite being built from scratch.
4.Key to Riff-Diff is the use of “artificial motifs”—helical fragment assemblies that preorganize catalytic side chains in energetically favorable conformations. These motifs are embedded into scaffolds using a modified RFdiffusion framework.
5.To ensure substrate accessibility, the authors introduced placeholder helices to enforce binding pocket formation during backbone generation. This strategy improves the mimicry of natural enzyme pockets and helps avoid clashes.
6.Iterative refinement cycles are used to improve backbone and sequence quality. Riff-Diff integrates structure prediction (ESMFold), sequence design (ProteinMPNN or LigandMPNN), and Rosetta-based optimization into a cohesive pipeline.
7.Experimental characterization of retro-aldolase variants revealed that several designs (e.g., RAD29 and RAD35) exhibit kcat values of ~0.03 s⁻¹—roughly 5 million-fold faster than the uncatalyzed reaction and surpassing many previous computational designs.
8.These designs also demonstrated high thermostability (folded beyond 90 °C), high chemical stability, and strong stereoselectivity (RAD35 achieved >99% ee). RAD35 also showed 895 turnovers after 48 hours.
9.Crystal structures of several variants showed near-atomic agreement with design models. Yet, catalytic activity did not always correlate with structural precision, highlighting the importance of active site dynamics.
10.Molecular dynamics and AlphaFold3 models revealed that side chain flexibility and substrate engagement, rather than static structural accuracy alone, better explain differences in catalytic performance.
11.For the MBH reaction, Riff-Diff produced enzymes that exceeded the activity of prior computational designs and even rivaled evolved MBHases. MBH48 achieved higher kcat and lower byproduct formation than BH32.8, an evolved variant from 13,590 screened clones.
12.Structural and CD data confirmed proper folding and design accuracy in the MBH enzymes. Crystal structures validated both backbone and side-chain placements, with observed deviations explaining performance differences.
13.The study underscores a crucial insight: active site precision is necessary but not sufficient for high activity. Catalytic geometry, substrate alignment, and conformational dynamics all co-determine enzyme function.
14.Riff-Diff’s design strategy generalizes to arbitrary catalytic arrays and offers a scalable, semi-automated framework for enzyme creation. It paves the way for testing computationally derived catalytic motifs in de novo scaffolds.
15.This work pushes enzyme design closer to predictive, low-throughput workflows and offers rich benchmarking datasets to develop future models of catalytic activity—potentially moving beyond empirical screening altogether.
💻Code:
github.com/mabr3112/riff_dif…
📜Paper:
biorxiv.org/content/10.1101/…
#ProteinDesign #EnzymeEngineering #ComputationalBiology #Catalysis #DeNovoDesign #SyntheticBiology #RiffDiff