Science's Data Search Engine

Evidence-synthesized pipeline (practical recipe)

1. Define atomic motif/theozyme — extract catalytic atoms and ligand/TS coordinates (DFT-derived for TS when relevant). RFdiffusion2 and RFdiffusion3 explicitly condition on these atoms to scaffold active-site geometry without pre-indexing residues RFdiffusion2 atom-level design paper
2. All-atom diffusion generation — run atom-conditioned diffusion to generate backbones that place catalytic atoms precisely; RFdiffusion3/AtomWorks extends to ligands and DNA with improved speed and broader target classes RFdiffusion3 AtomWorks preprint
3. Sequence design (ProteinMPNN / LigandMPNN) — sample multiple sequences per backbone (Rosetta/LigandMPNN used in RFdiffusion pipelines); generate ensembles of 6–8 sequences per backbone to improve odds of experimental success (RFdiffusion2 used 8 sequences/structure) RFdiffusion2 design sequencing
4. Structure prediction & filter — AlphaFold3/Chai-1 used to predict foldability and motif RMSD; retain designs with motif RMSD < ~1.5 Å and high AF3 pLDDT as filtering heuristics (used across RFdiffusion2/3 pipelines) RFdiffusion2 filtering
5. Optional physics refinement — Rosetta relax / ReplicaDock refinements help for flexible complexes (AlphaRED shows value combining AFm with Rosetta docking to improve interfaces) AlphaRED docking + refinement
6. In vitro screening & iterative selection — empirically test many designs; RFdiffusion2 reported several active enzymes from ~96 designs per campaign, and metallohydrolase campaigns reached large kcat/KM improvements; hit rates improve when modeling coordinated waters (ZnO designs outperformed Zn-only in metalloprotease work) RFD2-MI metalloproteases preprint

Key practical metrics (evidence-backed)

• RFdiffusion2: solved all 41 AME benchmark cases and produced experimental enzymes (retroaldol kcat/KM ≈ 6.34 M−1 s−1; metallohydrolases up to ≈53,000 M−1 s−1) — strong evidence atom conditioning helps motif geometry and yields measurable activity RFdiffusion2 outcomes
• Rosetta-era de novo enzyme design (e.g., retro-aldol 2008 Science): many designs with measurable activity (32/72 designs) but required iterative experimental work and exhibited limited efficiencies compared to evolved natural enzymes — atom-level diffusion now reaches similar or higher hit rates with broader chemistries Early Rosetta de novo enzyme design

Limitations, blind spots, and how to guard against them

• In vitro activity is encouraging but frequently remains below native enzymes; catalytic numbers (kcat/KM) vary widely across reactions and designs — do not assume parity with evolved enzymes without iterative optimization and selection RFdiffusion2 limitations
• Benchmark and training-data biases: AME cases are PDB-derived—designs may be biased towards motifs represented in structural databases; verify novelty with FoldSeek/TM and test against orthogonal chemistries RFdiffusion2 novelty checks
• Filtering heuristics (AF3 pLDDT, motif RMSD) help but can miss dynamics and solvent effects; combining MD-derived descriptors or short MD (QDPR) can add confidence before wet-lab tests QDPR MD-informed filtering

Practical reproducible checklist (minimal working pipeline)

1. Obtain atom-level motif (catalytic atoms, ligand/TS geometry via DFT if needed).
2. Run atom-conditioned diffusion (RFdiffusion2/3 or equivalent), generate ~100 backbones per motif case.
3. Design 6–12 sequences/backbone (LigandMPNN/ProteinMPNN); compute AF3 predictions for each.
4. Filter: motif all-atom RMSD < 1.5 Å, AF3 pLDDT high, low clash score; optionally run short MD (10–100 ns) and compute dynamics features (QDPR) to re-rank.
5. Refine selected models with Rosetta relax or AlphaRED-style docking where binding-induced conformational change is expected.
6. Test 50–200 designs in vitro (cell-free expression / in vitro translation enables rapid throughput) and iterate.

Caveat and falsifiability criteria

Claims that atom-level conditioning yields functional enzymes are falsified if: (a) motif-preserving designs (RMSD <1.5 Å) consistently fail to show activity across multiple independent campaigns; (b) improvements in in silico metrics (motif RMSD, AF3 pLDDT) do not correlate with experimental hit rates in blinded tests. RFdiffusion2/3 papers document reproducible experimental hits but note remaining gaps and the need for broader independent replication RFdiffusion2 reproducibility note

Key source extracts (evidence links)