Science's Data Search Engine

Bar chart recreates key kinetic outcomes from RFdiffusion2 metallohydrolase campaigns: top designs reached kcat/KM up to ~5.3×10⁴ M⁻¹s⁻¹, demonstrating near‑natural catalytic efficiency from de novo scaffolds starting only from active‑site/theozyme atoms.

Scatter shows ZETA_2's close structural match (RMSD ≤1.1 Å) to the design and very high catalytic efficiency — direct experimental verification that atom‑conditioned generative scaffolding can produce accurate, highly active active sites in novel folds.

Evidence summary (select excerpts)

• RFdiffusion2 produced de novo Zn‑dependent metallohydrolases from DFT/theozyme active‑site geometries; multiple designs were expressed and purified; top designs showed kcat/KM up to ~53,000 M⁻¹s⁻¹ and X‑ray structures matched models (apo RMSD 1.1 Å; Zn‑bound RMSD 0.8 Å) — experimental validation of atomic‑level scaffolding
  RFdiffusion2 metallohydrolase Nature
• Multimodal diffusion transformers (ProDiT) trained on sequence + structure can condition generation on active‑site motifs or GO terms, recover active residues at atomic fidelity in silico, and scaffold multiple functional states (multistate/allosteric designs) using coupled structure diffusion trajectories — supports the generality of active‑site conditioning beyond RFdiffusion2
  ProDiT multimodal diffusion preprint
• Designing serine hydrolases and other de novo catalysts via RFdiffusion + ensemble preorganization (ChemNet/PLACER) produced functionally active de novo folds for ester hydrolysis with kcat/KM up to ~3.8×10^3 M⁻¹s⁻¹, highlighting that active‑site geometry + ensemble preorganization predicts success and that new folds (not motif grafts) can catalyze chemistry
  De novo serine hydrolases preprint
• Directed evolution and distal mutations often improve turnover by adjusting substrate access, dynamics, and product release; therefore iterative redesign + selection remains essential after active‑site‑conditioned generation to tune kinetics beyond initial catalytic geometry
  Distal mutations preprint

Synthesis — how active‑site‑atom conditioning shifts design paradigms

1. From motif grafting to scaffold-from‑atoms: RFdiffusion2 and ProDiT show that providing atomic coordinates (theozyme atoms) as conditioning allows generation of backbones that place catalytic groups with sub‑Å fidelity, without pre‑specifying residue order — enabling truly de novo folds around a reaction geometry RFdiffusion2 evidence.
2. Atomic fidelity predicts function but is not the whole story: ensemble preorganization metrics (PLACER, ChemNet, Chai‑1) correlate with success; distal positions and dynamics require iterative tuning—experimental selection (screening / evolution) complements generative sampling Ensemble preorganization study.
3. High‑energy chemistries are reachable: designed metallohydrolases catalyzed Zn‑activated hydrolysis (4MU‑PA) at catalytic efficiencies previously unseen in de novo designs, indicating that demanding chemistries (e.g., amide/ester hydrolysis) can be encoded in new folds built around active‑site atoms Metallohydrolase kinetics and validation.

Limitations, blindspots, and falsification tests

• Most demonstrations target small, model reactions (4MU‑PA, ester/retro‑aldol); generality to native amide hydrolysis (peptidase‑like chemistries with larger transition states) remains to be proven experimentally RFdiffusion2 limitations.
• In silico metrics (DeepFRI, PLACER, Chai‑1) can bias selection; independent wet‑lab replication across labs is needed to exclude overfitting to pipeline metrics ProDiT caveats.
• Designs often require iterative rounds (sequence fitting, ensemble selection, experimental screening, directed evolution) to reach peak performance — generative design reduces but does not eliminate experimental optimization Distal mutation effects.

Practical takeaways — how to apply this paradigm

1. Start with a quantum‑chemistry optimized transition‑state/theozyme (atomic coordinates) for the target reaction.
2. Use an atom‑conditioned generative model (RFdiffusion2 / ProDiT / Riff‑Diff) to produce diverse backbones that place active atoms with sub‑Å fidelity.
3. Sequence‑fit (ProteinMPNN / LigandMPNN), predict ensembles (PLACER / Chai‑1 / ChemNet), and rank by active‑site preorganization metrics.
4. Experimentally screen a modest number (~20–100) of top candidates; iterate with mutagenesis and directed evolution to tune distal positions for binding/release and stability.

This pipeline has been shown to reduce the screening burden and produce highly active de novo enzymes, while acknowledging that iterative experimental optimization remains essential for peak kinetics and robustness Ensemble ranking improves hit rate.

Cited primary sources (with detailed extracts)