De novo design of allosterically controlled enzymes 🚀 New preprint from David Baker!🚀 1. The study presents a general de novo strategy to build compact enzymes whose catalytic activity is switched on or off by binding of a separately designed protein effector at a distal site, aiming to mimic natural allostery without domain fusions or large mechanical rearrangements. 2. Core idea: engineer an energy landscape with two nearby conformational states—an active state with catalytically optimal residue geometry and an inactive state with subtly distorted active-site geometry—then design an effector that selectively stabilizes one state to shift the equilibrium (activation or repression). 3. Methodological innovation: use partial RFdiffusion (add ~1–3 Å Gaussian noise then denoise) to generate alternative backbone conformations close to the starting enzyme fold, enriching for subtle but functionally disruptive active-site perturbations rather than large-scale motions. 4. Pipeline overview: start from a de novo serine hydrolase (Win1), generate inactive conformers by partial diffusion (with domain-wise perturbations to keep state energies close), design state-selective binders with RFdiffusion (30–80 aa), and perform multistate-compatible sequence design with tied ProteinMPNN across enzyme/effector and active/inactive contexts. 5. Allosteric activators: from 61 designed enzyme–effector pairs screened in E. coli lysates (4Mu-Ac substrate), two systems (Janus1, Janus2) showed effector-dependent activity increases and were validated with purified proteins; predicted state switching involved remodeling that restores catalytic geometry upon effector binding. 6. Activator kinetics and specificity: Janus1 kcat increased ~5-fold (0.0033→0.017 s⁻1) and Janus2 increased ~2.6-fold (0.00050→0.0013 s⁻1) at 1:2 enzyme:effector; cross-testing showed cognate specificity (non-cognate effectors did not enhance activity). 7. Allosteric inhibitors: designs were built on hydrolases for a bulkier substrate (4Mu-PhAc). A serine-to-cysteine substitution was found to broadly boost baseline catalytic efficiency in these designs and was used to increase dynamic range for observing inhibition. 8. Inhibitor performance: among 24 screened inhibitor pairs, Janus3 showed clear effector-dependent repression; purified-protein kinetics showed ~8-fold reduction in kcat upon effector binding (0.0024→0.0003 s⁻1) with little change in Km, consistent with modulation of catalytic geometry rather than substrate binding. 9. Structural validation and binding biophysics: a 1.7 Å crystal structure of the Janus3 enzyme–effector complex closely matched the design at the interface (Cα RMSD ~0.77 Å), confirming distal binding and conformational coupling. SPR measured strong affinities (Janus1 KD ~2.97 nM; Janus2 ~0.726 nM; Janus3 ~106 nM) and fast on-rates; regulation in spike-in experiments occurred within assay dead time (~21 s). 10. Coupling between catalysis and effector binding: for Janus2 and Janus3, substrate did not significantly affect effector affinity; Janus1 showed weaker binding at high substrate (KD shift ~3→13 nM), and experiments with a non-hydrolyzable analog suggested the effect is linked to catalytic turnover/acyl-enzyme intermediate formation rather than simple substrate occupancy. 📜Paper: biorxiv.org/content/10.64898… #ProteinDesign #Allostery #EnzymeEngineering #GenerativeAI #DiffusionModels #RFdiffusion #ProteinMPNN #SyntheticBiology #StructuralBiology #ComputationalBiology

RFOptimization: Guiding Design Optimization with All-Atom Structure Prediction 1 RFOptimization (RFO) is a training-free framework for rapidly improving “borderline” de novo binders by local, residue-wise mutational search—turning many near-miss designs into AF3-filter-passing candidates in ~1–10 optimization iterations (minutes on a single A100). 2 The core idea is to avoid optimizing against a single predictor’s biases: RFO interleaves two complementary mutation moves—(a) gradient-guided discrete mutations using RoseTTAFold3 (RF3) and (b) structure-prediction + inverse-folding cycling using an orthogonal predictor (Boltz) followed by ProteinMPNN/LigandMPNN redesign. 3 A key technical trick enables gradients with AF3-like diffusion predictors: RFO blocks gradient flow through RF3’s diffusion structure module (stop-gradient) and instead backpropagates through differentiable heads (distogram + confidence). This allows objective-driven sequence updates without differentiating through the expensive denoising trajectory. 4 RFO keeps the optimization state strictly discrete and chemically valid: each accepted mutation updates a one-hot sequence, then re-featurizes from scratch through RF3/AtomWorks so atom types, masks, connectivity, and reference geometry match the new residue identity (reducing “continuous relaxation” exploits). 5 The gradient move is implemented as gradient-guided MCMC: propose residue substitutions from the negative sequence gradient, accept via a Metropolis-style rule with geometric temperature annealing. This mixes fast local improvement with occasional uphill moves to escape local minima. 6 The cycling move provides exploration and debiasing: Boltz predicts an all-atom complex structure for the current sequence, then ProteinMPNN (or LigandMPNN for ligand/enzyme contexts) samples sequences conditioned on that backbone; the next state is chosen by MPNN likelihood while optionally fixing motifs/interface residues. 7 Objectives are modular and user-defined: RFO supports confidence losses (e.g., mean interface PAE, pLDDT), distogram proxy losses (contacts, interface distogram entropy), and explicit atom-level geometric constraints (interval penalties on distances between arbitrary atoms), enabling tasks like active-site scaffolding or metal/covalent geometry control. 8 Benchmark results (AF3 as held-out evaluator; success defined by AF3 iPAE < 2.5 and iPTM > 0.8) show consistent rescue of RFdiffusion/RFpeptides outputs: protein–protein binders improved on average 6.73% → 22.31% across 13 targets; cyclic peptides 1.25% → 12.57% across 4 targets; small-molecule binders 5.62% → 24.91% across 4 ligands. 9 For atomic-motif enzyme scaffolding (24 AME targets), RFO optimizes a composite objective combining atom-level catalytic geometry constraints with confidence (pLDDT), improving the fitness distribution (18/24 targets improved) while increasing predicted substrate confidence—highlighting use of all-atom constraints during sequence refinement. 10 Cross-model generalization and efficiency: under a stringent three-model consensus filter (RF3 + Boltz + AF3), the full two-branch method outperforms gradient-only and cycling-only ablations and an external cycling baseline (Protein Hunter). Amortized compute on PPI targets is ~26 GPU-min per consensus-passing design (vs ~178 min for Protein Hunter and ~34 GPU-hours for BindCraft), emphasizing targeted local optimization over massive resampling. 📜Paper: biorxiv.org/content/10.64898… #ProteinDesign #ComputationalBiology #DeNovoDesign #ProteinBinders #CyclicPeptides #EnzymeDesign #AlphaFold3 #RoseTTAFold #DiffusionModels #ProteinMPNN