Complex enzyme active site scaffolding by iteratively detuned catalytic guidance
Tingli QianiD, Weijin Wang, Shilong GaoiD, Aiko MuraishiiD, Beibei Zhao, Davi Nakajima AniD, Fanglei XueiD, Max Kazman, Johnny MendozaiD, Hannah NguyeniD, Asim Bera, Alex KangiD, Avery Yang, Wenzhao DaiiD, Jia Liang, Yujia WangiD, Xiongyi Huang, Frank DiMaioiD, David A. BakeriD
Abstract
De novo enzyme design requires sequences whose structures reproduce prescribed catalytic geometries. This remains challenging for active sites that couple multiple catalytic residues, cofactors and substrates. Here we introduce the Catalytic Guidance Engine (CaGE), which optimizes designed sequences to satisfy complex sets of functional constraints by alternating constraint guided structure prediction trajectories with sequence design, gradually fading the strength of the constraint information over cycles to nudge the designed sequence towards satisfying all the constraints. We train RF3 to accept atom-pair distances and other design conditions, and use classifier-free guidance to fade their influence over multiple design cycles. We also develop training-free motif guidance protocols which operate at inference time by directly adjusting selected atoms towards the design target geometry during diffusion sampling. CaGE improves catalytic-motif geometry across a diverse enzyme benchmark and yields sequences that recover complex target active sites in independent AF3 predictions. We experimentally demonstrate the power of CaGE by designing active enzymes with complex catalytic sites: a nonheme-iron design positions four active-site residues around an iron centre and a transition-state ligand, and a PLP-dependent design positions four catalytic residues around a PLP cofactor with multiple functional groups. Active site substitutions and a 2.7[A] crystal structure support the designed mechanisms and structures. Thus, CaGE provides a route to functional de novo enzymes with complex active sites.
02 /Specification & Classification
Click any tag to browse matching papersIn-depth summary
Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (16 Oct 2026) · CC BY 4.0
CaGE alternates guided structure prediction with sequence redesign, fading guidance each cycle, so sequences recover complex catalytic sites unaided; designed nonheme-iron and PLP-dependent enzymes were active.
As active-site complexity grows, designs predicted to recapitulate the site become harder to find, and generating large numbers of designs then filtering becomes computationally prohibitive. CaGE trains RF3 on atom-pair distance conditions and withdraws guidance across cycles using classifier-free guidance scales below 1.
Key results
ShownOn enzyme benchmark targets, CaGE gave higher sequence-level pass rates in independent AF3 predictions than an RFD3 plus LigandMPNN baseline without cycling, including targets where the baseline had no passing sequences at the strict cutoff.
At motif r.m.s.d. <1.25 Å with ligand r.m.s.d. <5 Å, CaGE pass rates were 8.11%, 1.57% and 0.32% for PNP, ODC and 4-CBA-CoA DH versus none for baseline. Passing sequences per 1,000 prediction calls: crotonase 0.62 to 8.06, CS 1.25 to 2.46. · CaGE recovers active-site geometry in unguided…
ShownBoth guidance routes pull predictions toward the target motif, and guided and unguided predictions of the same sequences converge as guidance is withdrawn over cycles.
Raising the CFG scale from 0 to 1 lowered median motif r.m.s.d.; atom-pair-trained RF3 beat untrained RF3 at matched scales. Training-free motif guidance with CFG gave comparable accuracy. The guided versus unguided gap in median joint r.m.s.d. narrowed monotonically across all eight targets. · Classifier-free guidance combines and adjusts…; Motif guidance enables precise geometric control…; Guided and unguided predictions converge across…
ShownFor a nonheme-iron transition-state motif, one CaGE cycle sharply raised unguided AF3 recovery, and a scaled-up campaign yielded an active, enantioselective designed enzyme.
Pass rate at 1.25 Å rose from 0.27% (1,491 baseline sequences) to 4.70% (1,701 CaGE sequences), and at 1.5 Å from 0.80% to 10.29%. Median Fe-triad deviation fell from 8.80 to 1.81 Å. Six of 71 tested designs gave product above the no-enzyme control; CaGE-13M gave 83% ee. · CaGE refines a coupled nonheme-iron catalytic…
ShownCaGE produced a de novo monomeric PLP-dependent retro-aldolase from a four-residue catalytic motif, and a second design round on the active hit improved yield.
None of 30,293 pre-CaGE sequences passed the AF3 criteria; CaGE gave 1,029 RF3-passing variants, 96 also passing AF3. CaGE-RA1.0 was the active hit from the 96 screened (22.8 ± 1.6% product); CaGE-RA1.1 reached 41.4 ± 1.7% (n = 3). · De novo design and experimental validation of a…; Hit-guided redesign improves catalytic activity
ShownAlanine substitutions, spectroscopy, mass spectrometry and a crystal structure support the designed catalytic architecture of CaGE-RA1.0.
D27A, H31A, K142A and R159A each reduced product to roughly the no-protein background. Intact mass rose from 21,504 to 21,735 Da after NaBH4 trapping with PLP. The 2.7 Å crystal structure matched the design model with backbone r.m.s.d. 0.702 Å. · Mechanistic and structural validation of the…
Methods, data and reproducibility
RF3 was trained to accept design conditions (including atom-pair distances), and classifier-free guidance or training-free motif guidance steers prediction toward the target geometry. Cycles of guided prediction and LigandMPNN redesign follow, with guidance scaled down to zero; final sequences are checked with AF3 and tested experimentally.
- Eight benchmark enzyme targets with catalytic motifs: PNP, ODC, 4-CBA-CoA DH, crotonase, CS, GalE, GLY and UGDH · n = 8 targets
- Nonheme-iron transition-state motif: RFdiffusion3 backbones, baseline versus CaGE sequences, with 71 designs tested experimentally · n = 1,500 backbones; 1,491 baseline and 1,701 CaGE sequences; 71 tested
- PLP-dependent retro-aldolase campaign: pre-CaGE sequences, CaGE variants passing RF3 and AF3 filters, and 96 designs screened in lysate · n = 30,293 pre-CaGE sequences; 1,029 pass RF3; 96 pass AF3
- Methods & tools
- RosettaFold3 (RF3), AlphaFold3 (AF3), RFdiffusion3, LigandMPNN, FastRelax, classifier-free guidance, motif guidance, intact-protein mass spectrometry, UV-visible spectroscopy, X-ray crystallography
- Reproducibility gaps
- Crystal structure coordinates will be deposited in the Protein Data Bank only after acceptance; the accession is a placeholder. Supporting data are said to be in Supplementary Information and Source Data files, not at a URL.
Limitations (3)
- Reported recovery rates are per evaluated sequence and exclude backbones discarded during cycling; efficiency is normalised by structure-prediction calls. · stated by the authors
- Experimental support rests on few hits: six of 71 iron designs were active above control, and one PLP hit lineage (RA1.0, RA1.1) was characterised in depth. Activity is endpoint product yield and no kinetics are given. · noted in this summary
- The comparison baseline (RFD3 plus LigandMPNN without cycling) and the pass thresholds were chosen by the authors, and the crystal structure is of the apo protein only. · noted in this summary
Who should read:Read this if you design enzymes with multi-residue, cofactor- or metal-containing active sites and want an inference-time way to improve catalytic-motif recovery.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Atom level enzyme active site scaffolding using RFdiffusion2
Woody Ahern, Jason Yim, … David A. Baker·Nature MethodsPeer-reviewed·Dec 3, 2025
- Computational design of serine hydrolases
Anna Lauko, Samuel J. Pellock, … David A. Baker·SciencePeer-reviewed·Feb 13, 2025
- Computational enzyme design by catalytic motif scaffolding
Markus Braun, Adrian Tripp, … Gustav Oberdorfer·NaturePeer-reviewed·Dec 3, 2025
- Computational Design of Metallohydrolases
Donghyo Kim, Seth M. Woodbury, … David A. Baker·NaturePeer-reviewed·Dec 3, 2025
- Function-guided design of active enzymes
Mingyang Hu, Lunjie Wu, … Linchao Zhu·bioRxiv·Jun 29, 2026
- Transplanting enzyme active site geometry into antibody CDRs for catalytic antibody design
Yaojun Zhu·bioRxiv·Jul 28, 2026