Generative models can now build a whole protein around a chosen catalytic site: RFdiffusion2 scaffolds atomic-level catalytic motifs and delivers enzymes that work in the wet lab, and GENzyme and EnzyGen generate pockets conditioned on substrates and reactions. But every one of these pipelines grades its designs by how well they bind something — a substrate, a transition-state mimic — and never asks whether the pocket can let go of the product. This is the exact failure that heterogeneous catalysis knows well as active site poisoning: a site strong enough to activate a reactant is often strong enough to trap the product, so the final release step becomes the bottleneck and the site stalls. Chemists have documented this "sticky product" problem in electrocatalytic CO$_2$ reduction, where sites that bind the $*COOH$ intermediate also bind the $*CO$ product through a linear scaling relationship Atomic Dispersed Hetero-Pairs for Enhanced Electrocatalytic CO2 Reduction, in nitrate-to-ammonia catalysts where bound NH$_3$ throttles turnover Gradient-concentration RuCo electrocatalyst for efficient and stable electroreduction of nitrate into ammonia, and in gold-catalyzed sugar oxidation Selective oxidation of uronic acids into aldaric acids over gold catalyst. Enzymes are not exempt: the lysozyme study below shows product release can be the rate-limiting step in a natural enzyme too.
Proteins, however, have an escape route that solid catalysts lack. A rigid surface must break the scaling relation by atomic-scale electronic engineering, as the Mo–Fe hetero-pair work does with bridging adsorption Atomic Dispersed Hetero-Pairs for Enhanced Electrocatalytic CO2 Reduction; a protein pocket can instead simply breathe open. The clearest evidence is Structure of a low-population intermediate state in the release of an enzyme product, which shows human lysozyme releases its trisaccharide product through a rare (~13%) "unlocked" conformation in which the interdomain angle widens from about 58° to 49° and a transient N44–E35 hydrogen bond appears; removing that bond with an N44A mutation cuts the product release rate to a third. Crucially, this release machinery is invisible in any single static structure — which is why the current generation of design tools, all of which evaluate static bound poses, cannot see it. My proposal is to mine the Protein Data Bank (PDB) — not as a catalogue of single structures but as an ensemble of paired conformational snapshots — to learn a predictor of product-release capacity, and to use it as a poisoning filter and steering signal for generative enzyme design.
Core Claim
The PDB contains, for thousands of enzymes, multiple crystallographic or cryo-EM snapshots of the same sequence in different ligation and conformational states — product-bound, apo, substrate-analog-bound. These pairs are a direct, large-scale record of how evolution solved the product-release problem: which loops move, which channels open, which contacts break. I claim that (i) a curated dataset of such pairs supports a learned, structure-based predictor of whether an enzyme (natural or designed) will retain or expel its product, and (ii) this predictor fixes a systematic blind spot in current generative enzyme design, where maximizing substrate and transition-state affinity implicitly risks tight product binding — the enzymatic analogue of the linear scaling relation. Formally, define the poisoning margin of an enzyme–product pair as
$$\mathcal{P} = \log_{10}\!\left(k_{\mathrm{off}}^{\mathrm{prod}} / k_{\mathrm{cat}}\right),$$
where $k_{\mathrm{off}}^{\mathrm{prod}}$ is the product dissociation rate and $k_{\mathrm{cat}}$ the turnover rate; the enzyme is product-poisoned exactly when $\mathcal{P} < 0$, i.e., the product leaves more slowly than the catalyst turns over. The predictor estimates $\mathcal{P}$ (or an ordinal surrogate) from paired structural information, and designs are filtered or re-weighted to keep $\mathcal{P}$ positive.
This is deliberately different from the closest existing work. GENzyme Reaction-conditioned De Novo Enzyme Design with GENzyme already takes product SMILES as input, but it uses the product only as a conditioning string and scores candidates by static substrate docking (Uni-Mol Docking v2); it never computes product affinity, release barriers, or inhibition. EnzyGen Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates states in its own limitations that it "focuses on binding the substrate" and ignores both transition state and product. RFdiffusion2 Atom-level enzyme active site scaffolding using RFdiffusion2 scaffolds an idealized static theozyme and treats burial as a hand-set dial rather than a release channel. SAKPE SAKPE: A Site Attention Kinetic Parameters Prediction Method for Enzyme Engineering predicts kinetic constants including $K_i$, but purely from sequence plus pre-annotated active sites, with no structural mechanism and no generative use. And the lysozyme study Structure of a low-population intermediate state in the release of an enzyme product is a single-enzyme retrospective analysis that its authors explicitly leave un-generalized. No existing work mines paired PDB conformations to learn product-release machinery, and none injects it into generative design.
Research Program
The program has three parts. First, I will construct ReleasePairs, a curated dataset mined from the PDB: for every enzyme chain with two or more deposited structures at high sequence identity, pair product- or inhibitor-bound states with apo or differently-liganded states of the same sequence, and annotate each pair with pocket-aperture change, loop and interdomain displacement, exit-channel bottleneck radius, and changes in product-pocket contact counts. This follows the spirit of prior PDB ensemble mining — General Protein Data Bank-Based Collective Variables for Protein Folding showed that reference structures across the PDB yield general low-dimensional coordinates, and Prediction of site-specific amino acid distributions and limits of divergent evolutionary changes in protein sequences showed PDB coordinates carry quantitative evolutionary signal — but applied to release geometry rather than folding or sequence statistics. Two cautions shape the construction. The PDB's static-snapshot bias means functional alternate states are underrepresented, as Modeling conformational flexibility of kinases in inactive states demonstrates for inactive kinase conformations, so I will weight pairs by state rarity rather than frequency; and AlphaFold DB AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage can supply only a single predicted state per sequence, so it serves as an apo-side supplement for sequences lacking experimental apo structures, never as evidence about dynamics. Kinetic labels come from the BRENDA/SABIO-RK records already assembled for kinetic prediction (over 31,000 $k_{cat}$ and 53,000 $K_m$ entries per SAKPE), extended with product-inhibition ($K_i$) annotations where available.
Second, I will train a release predictor: an SE(3)-equivariant network over the paired (bound, open) pocket geometries that outputs an estimate of the poisoning margin, with static product-docking features as one input channel and conformational-pair features as the distinctive ingredient. The architectural vocabulary exists — ReactZyme ReactZyme: A Benchmark for Enzyme-Reaction Prediction uses frame-averaging networks for equivariant active-site modeling — but it has only ever been used for retrieving which reaction an enzyme catalyzes, never for release kinetics.
Third, I will integrate the predictor into design: as a reranking filter over the design pools of RFdiffusion2 (on the Atomic Motif Enzyme benchmark) and EnzyGen (on EnzyBench), keeping designs whose predicted margin is positive at fixed substrate affinity, and as a differentiable penalty steering pocket generation. The scientific hypothesis underneath, worth testing in its own right, is that enzyme pockets can break the transition-state/product scaling relation through conformational gating — precisely the move that rigid solid surfaces must achieve electronically Atomic Dispersed Hetero-Pairs for Enhanced Electrocatalytic CO2 Reduction — and that the PDB records the statistics of how.
Proposed Experiments
The retrospective evaluation asks whether the predictor identifies product-inhibited enzymes. Holding out enzyme families from ReleasePairs, I measure rank correlation between predicted margin and measured product-inhibition strength across natural enzymes, reporting Spearman $\rho$ and AUROC for classifying known product-inhibited vs. uninhibited enzymes; for scale of difficulty, ReactZyme's top model reaches only 35.88% Top-1 and MRR 0.495 on time-split reaction retrieval ReactZyme: A Benchmark for Enzyme-Reaction Prediction, so a margin predictor achieving $\rho > 0.4$ across diverse families would be a strong result. A mechanistic case study uses lysozyme as ground truth: the predictor should assign the bound ground state a low release capacity, recognize the unlocked geometry (interdomain angle near 49°, the N44–E35 contact) as release-permissive, and predict the direction of the N44A effect (release rate dropping to roughly a third of wild type, per Structure of a low-population intermediate state in the release of an enzyme product).
The design evaluation reranks the output pools of RFdiffusion2 on the 41-case AME benchmark — where RFdiffusion2 itself achieves 100% scaffolding success against 39% for RFdiffusion Atom-level enzyme active site scaffolding using RFdiffusion2 — and of EnzyGen on EnzyBench (101,000 entries, 3,157 families; EnzyGen reports average pLDDT 87.45 and a 10.79% substrate-affinity gain over baselines Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates). The metric is the fraction of retained designs that (a) keep substrate affinity within a fixed tolerance of the unfiltered pool, and (b) show weaker computed product affinity than transition-state-mimic affinity, with a small subset validated by unbinding molecular dynamics. GENzyme's evaluation infrastructure on EnzymeFill (pocket RMSD 1.88 Å vs. 3.71 Å for RFDiffusionAA Reaction-conditioned De Novo Enzyme Design with GENzyme) provides the protocol template for pocket-level comparison. The decisive end-to-end check is a wet-lab redesign of one AME case — ideally a hydrolase, where RFdiffusion2 already achieved $k_{cat}/K_M$ of 53,000 M$^{-1}$s$^{-1}$ with fewer than 96 screened sequences Atom-level enzyme active site scaffolding using RFdiffusion2 — comparing a margin-filtered design against an unfiltered one under substrate–product competition, measuring whether turnover persists at product concentrations that poison the unfiltered control.
Baselines
The naive reference point is the unmodified RFdiffusion2 pipeline: designs are scaffolded around a static theozyme and selected by folding confidence, with no product consideration whatsoever Atom-level enzyme active site scaffolding using RFdiffusion2. This is the de facto standard in the field and the "do nothing" comparator; beating it means showing that margin-filtered pools measurably reduce predicted and, eventually, measured product trapping without sacrificing the 41/41 scaffolding success or substrate affinity.
The strongest competitor is a static product-docking filter: take the same generated pockets, dock the reaction product with Uni-Mol Docking v2 (the engine GENzyme already uses Reaction-conditioned De Novo Enzyme Design with GENzyme) or Gnina (the scorer in EnzyGen's benchmark Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates), and discard designs whose product affinity exceeds their substrate or transition-state-mimic affinity. This is the obvious cheap fix, and it must be run under identical conditions: same design pools, same seeds, same number of retained designs, and an equal rescoring compute budget, so the comparison isolates the value of the conformational-pair ingredient rather than tuning or compute. Beating it requires the release predictor to rank known product-inhibited enzymes more accurately than static docking does on the retrospective set, and to retain a higher fraction of designs that pass MD-based release checks at equal substrate affinity.
A second predictive baseline is SAKPE SAKPE: A Site Attention Kinetic Parameters Prediction Method for Enzyme Engineering, which reports $R^2 = 0.719$ for $k_{cat}$ (against 0.648 for UniKP and 0.468 for DLKcat), run on the designed sequences to flag likely inhibition. Its dependence on UniProt/EasIFA site annotations, which do not exist for de novo scaffolds, makes it a partial comparator, but it is the best available kinetic regressor and will be given the same inputs wherever annotations can be transferred. Finally, the key ablation is the release predictor trained with static pocket–product geometry features only, with the paired-conformation features removed: if this ablation matches the full model, the PDB-pair mining is not earning its place and the static docking filter suffices.
Why not just…?
Why not just dock the product into generated pockets and filter? This is the real threat, and it may capture most of the cheap wins. The evidence against it being sufficient is the lysozyme result itself: the release rate is governed by a ~13% conformational state that does not appear in the dominant crystallographic pose, so a static dock into the bound-like conformer — exactly what a generated pocket presents — systematically misestimates $k_{\mathrm{off}}$ Structure of a low-population intermediate state in the release of an enzyme product]. The heterogeneous-catalysis literature makes the same point from the other side: desorption is a barrier problem, not an equilibrium-affinity problem, and catalysts engineered specifically to lower the desorption barrier outperform those tuned by binding energies alone Gradient-concentration RuCo electrocatalyst for efficient and stable electroreduction of nitrate into ammonia. Still, the proposal treats this honestly: the static-docking filter is run as the strongest baseline, and the paired-conformation ablation is precisely the experiment that decides whether the extra ingredient matters. If the ablation shows no gain, the correct conclusion is that the filter alone is enough, and the proposal will say so.
Why not just use SAKPE to flag poisoned designs? SAKPE predicts $k_{cat}$, $K_m$, and $K_i$ from sequence and annotated sites SAKPE: A Site Attention Kinetic Parameters Prediction Method for Enzyme Engineering], so it looks like an off-the-shelf solution. Two things block it: it requires active-site annotations from UniProt/EasIFA, which de novo scaffolds lack by construction, and it offers no structural mechanism, so it cannot tell a designer what to change — which loop to move or which contact to weaken. It also reports no accuracy figure for $K_i$ on the merged kinetic dataset (no reported result found on Lacuna for SAKPE $K_i$ $R^2$), whereas its $k_{cat}$ $R^2$ of 0.719 is the figure to beat for any kinetic-prediction use. It enters the proposal as a baseline, not a solution.
Why not just run unbinding molecular dynamics on the top designs? Per design, MD is more faithful than any learned surrogate, and a small MD validation subset is part of the experimental plan. The problem is throughput: generative pipelines screen thousands of candidates per target (RFdiffusion2's own wet-lab workflow screens up to 96 sequences per active design after computational filtering Atom-level enzyme active site scaffolding using RFdiffusion2), and unbinding simulations that must capture rare events like the 13% unlocked state are far too expensive at that scale. The learned surrogate amortizes MD-level insight mined once from the PDB across the whole pool; MD is reserved as the ground-truth check on a fixed, pre-registered subset so the surrogate is validated rather than trusted.
Decisive Checks
Three results would decide the proposal. On the retrospective set, the full predictor must beat the static-docking filter at ranking product-inhibited enzymes, and the paired-conformation ablation must underperform the full model — otherwise the core ingredient adds nothing. On lysozyme, the predictor must recover the known direction and rough magnitude of the N44A mutation effect without being told the answer. And on at least one redesigned enzyme measured in the lab, a margin-filtered design must keep turning over under substrate–product competition where an affinity-matched unfiltered control stalls — the enzymatic counterpart of the desorption-engineered catalysts that stay active for hundreds of hours Gradient-concentration RuCo electrocatalyst for efficient and stable electroreduction of nitrate into ammonia. If the first check fails, the fallback finding is still informative: it would mean static product affinity, not release dynamics, is the binding constraint in current design pools, and the field's effort should go into a two-line docking check rather than a new model.
References
- Atom-level enzyme active site scaffolding using RFdiffusion2
- Reaction-conditioned De Novo Enzyme Design with GENzyme
- Generative Enzyme Design Guided by Functionally Important Sites and Small-Molecule Substrates
- Atomic Dispersed Hetero-Pairs for Enhanced Electrocatalytic CO2 Reduction
- Gradient-concentration RuCo electrocatalyst for efficient and stable electroreduction of nitrate into ammonia
- Selective oxidation of uronic acids into aldaric acids over gold catalyst
- Structure of a low-population intermediate state in the release of an enzyme product
- SAKPE: A Site Attention Kinetic Parameters Prediction Method for Enzyme Engineering
- General Protein Data Bank-Based Collective Variables for Protein Folding
- Prediction of Site-Specific Amino Acid Distributions and Limits of Divergent Evolutionary Changes in Protein Sequences
- Modeling conformational flexibility of kinases in inactive states
- AlphaFold Protein Structure Database 2025: a redesigned interface and updated structural coverage
- ReactZyme: A Benchmark for Enzyme-Reaction Prediction