henhaobio
Sign in

Nature MethodsBiochemistryPeer-reviewed journal article

Atom level enzyme active site scaffolding using RFdiffusion2

Woody AherniD, Jason YimiD, Doug TischeriD, Saman SalikeiD, Seth M. WoodburyiD, Donghyo KimiD, Indrek KalvetiD, Yakov KipnisiD, Brian CoventryiD, Han Altae-TraniD, Magnus S. BaueriD, Regina BarzilayiD, Tommi JaakkolaiD, Rohith KrishnaiD, David A. BakeriD

PDFDOIPreprint
In-depth full-text digest available · Jump to digest ↓

Abstract

De novo enzyme design starts from ideal active site descriptions consisting of constellations of catalytic residue functional groups around reaction transition state(s), and seeks to generate protein structures that can accurately hold the site in place. Highly active enzymes have been designed starting from such descriptions using the generative AI method RFdiffusion [1-3], but there are two current methodological limitations. First, the geometry of the active site can only be specified at the residue level, so for each catalytic residue functional group placed around the reaction transition state, the possible locations of the residue backbone must be enumerated by building side chain rotamers back from the functional group. Second, the location of the catalytic residues along the sequence must be specified in advance, which considerably limits the space of solutions which can be sampled. Here we describe a new deep generative method, Rosetta Fold diffusion 2 (RFdiffusion2), that solves both problems, enabling enzymes to be designed from sequence agnostic descriptions of functional group locations without inverse rotamer generation. We first evaluate RFdiffusion2 on an in silico enzyme design benchmark of 41 diverse active sites and find that it is able to successfully build proteins scaffolding all 41 sites, compared to 16/41 with prior state-of-the-art deep learning methods. Next, we design enzymes around three diverse catalytic sites and characterize the designs experimentally; in each case we identify active catalysts in testing less than 96 sequences. RFdiffusion2 demonstrates the potential of atomic resolution generative models for the design of de novo enzymes directly from their reaction mechanisms.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 2 (10 Apr 2025) · CC BY 4.0

RFdiffusion2 scaffolds unindexed atom-level catalytic active sites via flow matching without inverse rotamer enumeration, solving all 41 in silico benchmark cases and yielding active de novo enzymes across four reactions.

Prior diffusion-based motif-scaffolding models operate on sequence-indexed backbone frames, requiring computationally expensive enumeration of inverse side-chain rotamers and sequence positions to scaffold atom-level enzyme active sites (theozymes). By jointly inferring rotamers, sequence indices, and backbones directly from unindexed functional-group coordinates, RFdiffusion2 scales de novo enzyme design to multi-residue active sites.

Key results

  1. ShownRFdiffusion2 solved all 41 active-site scaffolding problems in the Atomic Motif Enzyme (AME) benchmark, substantially outperforming RFdiffusion.

    Across 41 curated M-CSA/PARITY active sites spanning EC classes 1-5 (100 sampled structures per site, 8 LigandMPNN sequences each, scored by Chai-1 with catalytic heavy-atom RMSD < 1.5 A and no ligand clashes < 1.5 A), RFdiffusion2 succeeded on 41/41 cases versus 16/41 for RFdiffusion, achieving higher success rates on 40 of 41 cases. · Abstract; 3 Atomic Motif Enzyme Benchmark; Figure legends

  2. ShownAllowing RFdiffusion2 to jointly infer both side-chain rotamers and sequence indices outperformed both naive enumeration and fixed native crystal-structure rotamers and indices.

    On representative AME benchmark cases with 3, 4, 5, and 6 discontiguous residue islands (100 designs per setting), inferred rotamers and indices yielded the highest overall success rate and greatest local structural diversity, whereas naive rotamer/index sampling achieved zero successes at >=4 residue islands. · 3 Atomic Motif Enzyme Benchmark; Figure legends

  3. ShownScaffolding atomic theozymes extracted from crystal structures yielded active retroaldolases and multiple-turnover cysteine ester hydrolases while testing <=96 designs.

    Screening 96 IVTT designs scaffolding a Lys-Tyr-Tyr-Asn tetrad from PDB 5AN7 yielded 4 variants with retroaldolase activity on racemic Methodol (best design closest PDB 6SU3, TM-score 0.47); screening 48 designs scaffolding a Cys-His-Asn triad and Gln oxyanion hole from PDB 1PPN on 4MU-Butyrate yielded active enzymes with best acylation kcat/KM = 248 +/- 34 M^-1 s^-1 (closest PDB 5K7V, TM-score 0.53). · Abstract; 4 In vitro results; Figure legends

  4. ShownDesigning zinc metallohydrolases directly from Density Functional Theory (DFT) transition-state geometries yielded catalytically active de novo enzymes with kcat/KM up to 53,000 M^-1 s^-1.

    Testing 96 designs per DFT theozyme yielded 3 active zinc hydrolases for 4MU-Butyrate (best kcat/KM = 77 +/- 10 M^-1 s^-1; PDB 2ZTK, TM-score 0.54), 5 active enzymes for 4MU-Phenylacetate (best kcat/KM = 16,000 +/- 2,000 M^-1 s^-1; PDB 4WIB, TM-score 0.5), and 11 active enzymes with a DFT-placed Glu base (best kcat/KM = 53,000 +/- 5,000 M^-1 s^-1; PDB 2EI4, TM-score 0.60). · 4 In vitro results; Figure legends

Methods, data and reproducibility

Extends RoseTTAFold All-Atom with R^3 Gaussian and SE(3) Riemannian flow matching and stochastic 3D centering to generate protein structures conditioned on unindexed atomic motifs, ligand relative accessible surface area, an origin token, and partial ligands. Assigns sequences with LigandMPNN, filters designs via Chai-1 structure prediction, and validates activity in enzyme assays.

  • Protein Data Bank (PDB) biomolecular structures (proteins, protein-small molecule, protein-metal, and covalently modified complexes) used for flow-matching training · multiple species
  • Atomic Motif Enzyme (AME) benchmark curated from 958 M-CSA active sites cross-referenced with PARITY crystal structures containing all reactants and cofactors across EC classes 1-5 · multiple species · n = 41 active sites (100 structures x 8 sequences per site) · 1qgx
  • Experimental retroaldolase designs scaffolding a minimal catalytic tetrad from PDB 5AN7 assayed on racemic Methodol · none · n = 96 designs screened via IVTT · 5AN7
  • Experimental cysteine ester hydrolase designs scaffolding a catalytic triad and oxyanion hole from PDB 1PPN assayed on 4MU-Butyrate · none · n = 48 designs screened · 1PPN
  • Experimental zinc metallohydrolase designs scaffolding DFT-optimized transition states for 4MU-Butyrate and 4MU-Phenylacetate (with and without a Glu general base) · none · n = 3 campaigns x 96 designs screened
Methods & tools
RFdiffusion2, RFdiffusion, RFdiffusionAA, LigandMPNN, Chai-1, AlphaFold2, AlphaFold3, FoldSeek, RDkit, RosettaMatch
Assay conditions & compute
Final RFdiffusion2 model trained for 17 days using 24 A100 Nvidia GPUs [S3].
Reproducibility gaps
Although the text states that RFdiffusion2 training/inference code and the AME benchmark are freely released, no repository URLs or software version numbers are printed in the text [S4, S6].
Limitations (3)
  • De novo enzymes designed by RFdiffusion2 remain less active than native enzymes, likely because minimal theozyme definitions omit secondary catalytic interactions [S6]. · stated by the authors
  • In the initial 4MU-Phenylacetate zinc hydrolase campaign where a general base was not explicitly included in the theozyme, mutating a chance active-site glutamate to alanine in the top design had little-to-no effect on catalytic efficiency [S5]. · stated by the authors
  • RFdiffusion2 generates backbones before sequence fitting with LigandMPNN rather than co-designing protein sequence and non-motif side chains around the substrate binding pocket [S6]. · stated by the authors

Read this if you develop generative protein structure models or design de novo enzymes and small-molecule binders around atom-level active site geometries.

Citations
119
FWCI
35.22
Field percentile
top 1%

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

Similarity Map
SelectedSame fieldCross-field
THISBraun '25Kim '25Qian '26Li '26Butcher '25Watson '23
81% semantic matchBiochemistry
Markus Braun, Adrian Tripp, Morakot Chakatok+13 more·Nature·Dec 3, 2025· 80 citations