acceptodds
Under review as a conference paper at ICLR 2027

ANCHORZYME: Anchor-Conditioned Dual-Memory Enzyme Generation for Unseen Substrates

Abstract

Designing enzymes for substrates with no known natural catalyst is a central AI for Science problem because it could expand biocatalysis to new reactions in sustainable synthesis, drug and materials manufacturing, and environmental remediation; we study strict substrate-only zero-shot enzyme generation, where the input is only a small query molecule and the output is a candidate catalytic protein sequence. The main obstacle is an activity–structure trade-off: in this setting, edits that increase predicted enzyme–substrate compatibility often reduce foldability or structural confidence. We show that this trade-off is amplified when retrieved enzymes are used as a single undifferentiated prompt, because global scaffold similarity, which supports foldability, is mixed with local pocket and active-site signals, which support substrate recognition and catalysis. To separate these roles, we propose AnchorZyme, a retrieval-augmented, anchor-conditioned editing framework that selects a scaffold anchor, retrieves donor enzymes, aligns them to define site, shell, and distal edit regions, trains a teacher-guided generator to make better-than-anchor edits, and combines a catalytic-gain discriminator with a frozen sequence-plausibility discriminator for sampling and reranking. On RealUnknown-389, a curated benchmark of 389 currently enzyme-unknown substrates, AnchorZyme achieves the best UniKP and normalized pLDDT, reaching 0.384 and 0.770, and improves over SENZ from 0.310 to 0.384 in catalytic score while raising structural confidence from 0.499 to 0.770, establishing scaffold-preserving local editing as a practical route toward discovery-oriented zero-shot enzyme design for substrates beyond currently known biocatalysis.Our code is available at https://anonymous.4open.science/r/code-EEE1.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Enzyme Design, Zero-Shot Learning, Catalytic Protein Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EDMXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

ANCHORZYME is a retrieval-augmented framework for zero-shot enzyme generation, where the goal is to generate candidate enzyme sequences using only a substrate molecule as input, without access to its known enzyme, EC label, reaction products, or protein structure. The authors identify an activity–structure trade-off in existing methods: improving predicted enzyme–substrate compatibility can reduce the structural confidence of generated proteins. To address this issue, ANCHORZYME separates retrieved enzyme information into two parts: a scaffold anchor that provides a structurally plausible starting sequence, and catalytic donor memories that provide residue-level information for sequence editing. The framework uses a biochemical knowledge graph for retrieval, a frozen UniKP model to construct teacher-supervised edits, and catalytic and sequence-plausibility discriminators to guide a two-stage editing process. The paper also introduce REALUNKNOWN-389, containing 389 molecules with no recorded catalytic enzyme in the database snapshot. On this benchmark, ANCHORZYME achieves a predicted UniKP log10(kcat) of 0.384 and an ESMFold normalized pLDDT of 0.770, outperforming the evaluated substrate-only baselines on both metrics.

Critical strengths

The anchor–donor decomposition is biologically motivated and provides a sensible design strategy. A key strength is the explicit separation of scaffold preservation from local residue editing. A protein must maintain a stable structural framework while allowing changes near functionally important residues to support new substrate interactions. This idea is implemented by selecting a scaffold anchor, retrieving distinct donor enzymes, and aligning their sequences to identify site, shell, and distal regions. The donor-induced edit prior provides alternative residues, while the anchor prior encourages preservation of the original scaffold. The two-stage sampling procedure further separates initial site/shell editing from subsequent distal editing. This is more structured than directly conditioning a generator on an undifferentiated set of retrieved sequences. The anchor-relative teacher objective is also reasonable because it explicitly asks whether a proposed sequence improves over its starting enzyme rather than maximizing an unconstrained score.

Critical weaknesses
  1. The paper does not sufficiently address the distinction between recognizing a substrate molecule and specifying a catalytic transformation. A substrate alone generally does not determine which reaction an enzyme should catalyze. The same molecule may participate in multiple enzymatic reactions, involving different reaction centers, products, cofactors, and catalytic mechanisms. However, the proposed setting explicitly excludes reaction products and EC labels, and REALUNKNOWN-389 does not provide a known target enzyme or experimentally verified catalytic activity. This creates an important ambiguity in the learning objective. It is unclear what constitutes a scientifically correct enzyme proposal for a given molecule, beyond receiving a favorable predicted compatibility score. A generated enzyme might be compatible with a molecule in some sense without catalyzing a useful or intended transformation. I think the benchmark is valuable as a chemically novel screening set, but its current construction does not establish whether the generated enzymes solve a well-defined catalytic task.
  2. The dependence on UniKP throughout both model development and evaluation. I think this creates a risk of surrogate optimization bias: the framework is explicitly designed to produce sequences that score well under UniKP, and its main reported activity improvement is measured using that same predictor. Even though UniKP is frozen, its use as both a training signal and an evaluation metric means that a higher score does not independently demonstrate better catalysis. Plus, for the structural evidence, ESMFold pLDDT measures confidence in a predicted structure, not experimental foldability, stability, or catalytic-site geometry.
  3. The authors argue that the separation of anchors and donor memories enables better enzyme generation. However, an alternative explanation is that starting from natural enzymes already provides strong structural confidence, while local mutations and a larger proposal pool increase the chance of obtaining high surrogate scores.
Questions for authors
  1. How is catalytic success defined when only a substrate molecule is provided?
  2. How can you rule out optimization toward the UniKP surrogate rather than real catalytic improvement?
  3. How much improvement comes from the anchor itself, and how much comes from learned editing?
3 comments

Thank you for recognizing the anchor–donor decomposition and staged editing design. We address your questions with external pairing evaluation, modern generation baselines and structural analysis. Our application goal is to provide ranked sequence candidates when a substrate is available but a target enzyme is unknown, and to test the contributions of retrieval, donor information and learned editing.

1. What task does substrate-only input define?

ANCHORZYME addresses the proposal stage of enzyme discovery: given a molecule, return a ranked set of enzyme sequences for further investigation. Section 3.1 already defines this output. The application starts precisely where a target enzyme, EC label, product or known protein structure is unavailable.

Your observation that one substrate can participate in multiple reactions is consistent with a candidate-set formulation. Requiring a product or EC label in advance would define reaction-specified enzyme design. We focus on the earlier discovery stage, using available biochemical knowledge to supply sequence starting points and new local edit choices for molecules without an assigned enzyme.

“Unknown” in REALUNKNOWN-389 has a specific meaning: the database snapshot contains no recorded catalytic enzyme for the molecule. This is an application cohort for candidate discovery. We assess the proposals through pairing scores and structural analysis, and examine candidate recovery on development targets with reference sequences.

We will foreground this task definition: the paper expands the sequence candidates available for validation in the unknown-enzyme setting, with a ranked set as its output.

2. How do we assess performance beyond the training scorer?

We evaluate frozen candidates with ProSmith-ESP, which was used neither in training nor in generation guidance or selection. Its predicted enzyme–substrate pairing scores directly match the candidate-discovery task.

On 99 full-length targets, all three policies share a requested candidate cap of 20 and the final selector:

Candidate policyProSmith pairing score of final choice
Retrieval R200.06195
R10 + simple editing0.06783
R10 + learned editing0.07389

The learned hybrid has the highest observed mean among the three policies, 0.07389. It uses the same actual candidate count as the simple-editing hybrid: 1,900 candidates per seed, compared with 1,980 for retrieval.

Paired-difference 95% intervals: learned hybrid minus retrieval, [−0.01120, 0.03420]; learned hybrid minus simple-editing hybrid, [−0.00470, 0.01635].

Following your suggestion on structure, we further examined complete sequences and local annotated regions. Across 99 targets, mean Cα pLDDT is 85.96 for learned editing, 86.11 for count-matched simple editing and 88.98 for unchanged anchors. Across all 11 eligible targets with experimental references and annotations, the same mean substitution count gives an annotated-site distance deviation of 1.120 Å for learned editing and 1.436 Å for simple editing; nine targets favor learned editing. This descriptive analysis directly assesses local geometry. Annotations include catalytic, cofactor- and metal-related positions and overlap the existing motif information.

Each addition measures a specific property: ProSmith assesses external pairing scores, full-length pLDDT assesses prediction confidence, and reference-site distances assess local geometric changes after editing.

The modern-baseline comparisons further highlight ANCHORZYME's full-length structural-confidence advantage. Under the same ESMFold settings, our mean pLDDT and that of RFD3 are 86.80 and 74.18 on 74 shared targets; the corresponding values for ANCHORZYME and Proteina/Complexa are 86.45 and 51.58 on 78 shared targets. This direct structural assessment is central to our response to the structural-quality concern. The corresponding returned-set ProSmith means are 0.05113/0.10394 and 0.05045/0.10529, providing a complementary view of predicted pairing propensity. pLDDT evaluates each method's fixed native first complete sequence; ProSmith averages all returned candidates within target. These endpoints characterize structural prediction confidence and pairing propensity, while annotated-site distance deviations further examine local geometry.

3. How do we separate anchor, donor and pool-size effects?

The natural anchor is an intentional starting point in our method; the incremental donor contribution is tested by holding that starting point fixed. Across 114 editable development targets, returned-set full-length recovery is 0.09381 with distinct donors off and 0.11428 with donors retained—a gain of 2.05 percentage points, with a paired 95% confidence interval of [1.49, 2.65]. Mean actual returns increase from 4.54 to 9.06. Both settings share the anchor, editable support and budget, identifying the donor contribution to candidate availability and set-level recovery.

To examine larger pools and additional mutations as alternative explanations, the 99-target full-length experiment gives every policy a requested cap of 20 and the same final selector. Simple editing also matches the learned output's substitution count. Actual candidate totals per seed are 1,980 for retrieval, 1,900 for the simple-editing hybrid and 1,900 for the learned-editing hybrid. The ProSmith results above come from this setting: learned and simple-editing hybrids use the same candidate count and obtain observed means of 0.07389 and 0.06783, respectively.

The experiments have distinct roles: the fixed-anchor intervention tests donor information, and the cap- and edit-count-matched controls test the editing policies. The design uses natural sequences as proposal starting points and distinct donors to expand the residue choices available for exploration.

Thanks for your rebuttal and additional experiments. The rebuttal improves the experimental evidence. But I would keep the original rating. The new independent evidence of catalytic improvement is still limited. the 95% CI both include zero, which don't provide conclusive evidence for the capability of your learned method. Plus, the activity-structure trade-off remains unresolved. The structural analysis is informative but unchanged anchors achieve higher pLDDT than learned edits.

Reviewer_bTxNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ANCHORZYME proposes enzyme candidates for unfamiliar molecules by retrieving existing enzymes and carefully editing them. Its strongest result is improved predicted structural confidence alongside higher predicted activity. However, the study does not establish that the generated enzymes actually catalyze reactions.

Critical strengths

Studying new-to-nature-enzyme design is an important problem and this paper targets at this important task.

Critical weaknesses

Evaluation

The evaluation has big issues:

First, there lacks substantial function evaluation of the designed enzyme sequences. At least the paper needs to evaluate the binding between the enzymes and substrates such as docking, MD simulation.

Second, the pLDDT is too low, which is below 0.8, showing that the designed enzymes are not even foldable.

Additionally, the ProGen2 is a language model without any condition. How did the author implement the baseline?

Finally, the paper needs to compare with some recent SOTA models, like ESM3, RFD2, RFD3 and protina-complexa

Applicability:

Since the paper studies new enzymes, it would be interesting to carefully analyze some designed enzymes with the annotation of active sites or even binding with substrates. This will make us have deeper understanding about the performance of this work.

Questions for authors

See above.

3 comments

Thank you for the suggestions on structural evaluation and recent baselines. We added full-length structure predictions, local-geometry analysis at experimentally annotated sites, and comparisons with RFD3 and Proteina/Complexa. On the shared targets with each of the two new structure-generating baselines, ANCHORZYME achieves mean pLDDT above 86 and higher predicted structural confidence. In the edit-count-matched geometry comparison, learned editing is closer to the experimental reference on nine of 11 targets.

1. Full-length predictions directly address the pLDDT concern

The new full-length comparison places ANCHORZYME above 80 in both shared-target cohorts and above the corresponding modern structure-generating baseline. We use the same ESMFold settings for each method's frozen native rank-1 returned candidate (the first generated candidate for the two baselines). Complete sequences are evaluated without truncation or confidence-based candidate selection.

Shared targetsANCHORZYME full editorComparatorComparator pLDDT
7486.80RFD3 + LigandMPNN74.18
7886.45Proteina/Complexa51.58

ANCHORZYME is higher by 12.62 and 34.87 pLDDT points. The paired 95% intervals are [8.65, 16.75] and [31.22, 38.34], with higher values on 52/74 and 75/78 targets. We completed 578 new baseline structure predictions and reused the completed editor predictions. The editor's two sampling seeds are averaged within each target.

Regarding the inference that a value below 0.8 means the sequences are not foldable, the original 0.770 is a normalized mean corresponding to 77.0 on the 0–100 scale. pLDDT measures local prediction confidence; 80 is not a universal foldability boundary. A mean below 80 therefore does not establish that a sequence is non-foldable. See the EMBL-EBI metric definition.

This new analysis evaluates full-length native first proposals on shared targets and supplements the original 389-target table. In the complete 99-target structure comparison, learned editing, edit-count-matched simple editing and unchanged natural anchors obtain means of 85.96, 86.11 and 88.98, respectively. These results characterize the predicted structural confidence of the output sequences. The next analysis uses experimental reference structures to examine local geometry around annotated sites.

2. Examining edited sequences at annotated sites

With the same mean substitution count, learned editing is closer to the experimental reference geometry at annotated sites on nine of 11 targets. From the 99-target cohort, we selected targets using exact sequence mapping, experimental annotations and reference-structure availability, and included all 11 eligible targets spanning nine anchors. Eligibility does not depend on model outcomes.

MetricLearned editingCount-matched simple editing
Mean substituted residues29.3229.32
RMS deviation of annotated-site Cα pairwise distances (Å; lower is better)1.1201.436
Annotated-residue retention85.03%81.93%

The geometric measure compares distances between corresponding annotated positions in the predicted and experimental reference structures. It directly examines how local edits affect these regions and complements global pLDDT. Annotations cover catalytic, cofactor- and metal-related sites and overlap the existing motif information; this is a descriptive analysis of annotated regions.

To strengthen function-related evaluation, we use enzyme–substrate pairing scores and geometry at experimentally annotated sites. ProSmith-ESP assesses predicted pairing propensity from an enzyme sequence and substrate SMILES, directly matching the candidate-discovery task, and was not used in training or selection.

3. Modern baselines highlight our full-length structural-confidence advantage

ANCHORZYME achieves substantially higher full-length structural prediction confidence in both modern-baseline comparisons. Across the 74 targets shared with RFD3, our mean Cα pLDDT is 86.80, compared with 74.18 for RFD3. Across the 78 targets shared with Proteina/Complexa, the corresponding values are 86.45 and 51.58. This advantage directly addresses the concern about the structural quality of the generated sequences.

To address this concern, we first examine confidence in the full-length structures of native rank-1 proposals, and complement this assessment with an independent score of substrate pairing.

Shared targetsMethodpLDDT ↑ (0–100)ProSmith ↑
74ANCHORZYME full editor86.800.05113
74RFD3 + LigandMPNN74.180.10394
78ANCHORZYME full editor86.450.05045
78Proteina/Complexa, single-pass51.580.10529

pLDDT evaluates each method/seed's native rank-1 returned sequence, fixed before structure prediction (the first generated sequence for the two baselines), using the same ESMFold pipeline on complete sequences. ProSmith averages all returned candidates within target and then over targets. The columns use the same generation batches and shared targets. The editor's two sampling seeds are averaged within target; each baseline uses one fixed generation realization.

A higher pairing score does not replace structural assessment of complete output sequences. The two baselines have higher returned-set ProSmith means, while ANCHORZYME's native rank-1 candidates score 12.62 and 34.87 pLDDT points higher under the same structure-prediction pipeline. Our full-length structural-confidence advantage is therefore the direct evidence addressing the reviewer's structural concern. ProSmith-ESP measures predicted enzyme–substrate pairing propensity, and pLDDT measures local structural prediction confidence.

Both baselines start from the same query-substrate SMILES. We compute one three-dimensional ligand conformer with RDKit ETKDGv3. The design length is taken from the shared frozen retrieved anchor; the baselines receive its length, without its sequence or structure. No native complex, known bound pose, catalytic-site constraint, EC label, reaction or product is supplied.

RFD3 generates ten backbones. LigandMPNN then designs one sequence per generated backbone using the backbone and ligand-atom context. Proteina/Complexa uses ligand-only single-pass joint structure–sequence generation to produce ten sequences, with reward search and subsequent sequence redesign disabled. Candidates are retained in generation order. RFD3 completes 280/389 targets with 2,800 sequences; Proteina/Complexa completes 298/389 targets with 2,980 sequences.

The editor uses 20 proposals per sampling seed and each baseline uses ten; all request ten returned sequences. For the 74-target comparison, the editor per seed and RFD3 return 722 and 740 sequences. For the 78-target comparison, the editor per seed and Proteina return 726 and 780. These are comparisons of pipeline outputs with the generation budgets specified. The separate 99-target retrieval–editing hybrid experiment evaluates the common selector's final choice.

Computational cost. Our generation stage processes 99 targets with 20 proposals per target in about 83 minutes per seed, restoring full-length sequences after local editing. The table also reports the available completed-run timings for the two modern baselines.

Completed runsScope and proposals per targetRecorded generation time
ANCHORZYME99 targets per seed; 20 proposalsAbout 83 minutes per seed; amortized 50.18–50.28 s/target
RFD3 + LigandMPNN280 targets; 10 proposalsMean 267.19 s/target
Proteina/ComplexaTwo completed pilot runs; 10 proposals each468 residues: 169.02 s; 710 residues: 326.14 s

These are observed costs of the evaluated implementations. ANCHORZYME generates within a 512-column editing window and restores the full scaffold sequence, amortizing one model load over 99 targets; retrieval and context preparation are excluded, and the two seed processes run concurrently on one GPU. RFD3 and Proteina/Complexa load models separately for each target. RFD3 timing includes LigandMPNN sequence design; both baselines exclude prior ligand-conformer preparation.

4. How is ProGen2 used in this task?

ProGen2 serves as a substrate-unconditioned reference for generic protein generation. The original REALUNKNOWN-389 experiment uses the official ProGen2-large checkpoint in FP16, generating ten sequences per target, or 3,890 sequences in total.

Requested lengths are sampled from the empirical training-protein length distribution, whose prior spans 50–1,024 residues. The wrapper attaches target identifiers and SMILES after generation for the common downstream evaluation. SMILES is not passed to ProGen2, and lengths do not use true-enzyme labels for the query substrate.

This control compares generic protein generation with substrate-directed candidate proposal. We will explicitly label it as an unconditional baseline and discuss it separately from natively substrate-conditioned methods such as SENZ.

Reviewer_s7xMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies zero-shot enzyme generation for substrates with no known catalytic enzyme, where the goal is to generate plausible protein sequences using only the substrate molecule as input. The authors propose ANCHORZYME, a retrieval-augmented framework that combines scaffold-preserving sequence editing with catalytic information from retrieved enzymes. The method selects a natural enzyme as a scaffold anchor, uses other retrieved enzymes to identify candidate residue modifications, and trains a generator with surrogate-guided supervision and discriminators for catalytic activity and sequence plausibility. The authors also introduce REALUNKNOWN-389, a benchmark of 389 substrates without recorded catalytic enzymes. Experiments show that ANCHORZYME outperforms existing substrate-only generation methods, both predicted catalytic activity and structural confidence.

Critical strengths

Well-motivated integration of retrieval and scaffold-preserving enzyme generation. The paper proposes a sensible decomposition of retrieved enzyme information into scaffold anchors and catalytic donor memories, allowing the generator to preserve a plausible protein structure while making targeted modifications for substrate compatibility. Although scaffold-based design and surrogate-guided editing are established ideas, their integration into a substrate-only zero-shot setting is interesting and provides a useful extension of existing retrieval-augmented approaches such as SENZ.

Encouraging empirical results with supporting ablations. ANCHORZYME achieves the strongest results among the evaluated substrate-only baselines on both predicted catalytic activity and structural confidence. The experiments also go beyond aggregate comparisons, including ablations of catalytic guidance, structural plausibility guidance, and teacher supervision, as well as analyses of donor memories and learned editing.

Critical weaknesses

(Moderate) Limited isolation of the proposed dual-memory architecture's contribution. The paper combines several established ideas, including scaffold-based enzyme design, retrieval-augmented generation, and surrogate-guided sequence optimization. While their integration is interesting, the experiments do not fully establish how much of the improvement is attributable to the proposed separation of scaffold anchors and catalytic donor memories. The authors provide useful ablations showing that donor information and learned editing contribute to performance, but these do not directly compare against simpler scaffold-preserving editing approaches under matched retrieval and optimization budgets. In particular, it remains unclear whether the improvements primarily arise from starting with structurally plausible natural enzymes and optimizing them locally, or whether the dual-memory decomposition provides a substantive additional benefit. A targeted comparison against a simpler anchor-based editing baseline would help clarify the methodological contribution relative to prior work.

(Minor) Limited independent validation of predicted catalytic improvements. The main catalytic evaluation relies on UniKP, which is also used to construct teacher supervision and train the catalytic discriminator. Consequently, improvements in UniKP scores may reflect successful optimization of the surrogate rather than genuine improvements in enzyme–substrate compatibility. This concern is particularly relevant given prior work documenting the limitations of computational enzyme-activity predictors, especially on unfamiliar sequences and substrates (e.g., CatPred by Boorla and Maranas, 2025). Although the authors appropriately acknowledge that their results are computational rather than experimental, evaluating generated enzymes using an independent activity predictor would provide stronger evidence that the improvements generalize beyond the surrogate used for optimization. This is especially important because the proposed benchmark contains substrates without known catalytic enzymes, making direct validation difficult.

Questions for authors

I'd appreciate the authors commenting on the weaknesses above, and incorporate that commentary into the camera-ready.

The authors could also improve clarity by more explicitly distinguishing the central architectural ideas from supporting implementation details, perhaps using a simpler running example to illustrate how scaffold anchors, donor memories, and targeted editing interact.

1 comment

Thank you for recognizing the problem setting and the anchor–donor design. We added simple-editing controls and external evaluation with the official ProSmith-ESP model. ProSmith takes an enzyme sequence and substrate SMILES and returns a score from 0 to 1; higher scores indicate stronger predicted enzyme–substrate pairing propensity. This endpoint directly matches our goal of proposing relevant enzyme candidates for a specified substrate. ProSmith was not used for teacher supervision, generation guidance or candidate selection.

1. What do donors contribute beyond the natural anchor?

Distinct donors provide useful residue alternatives for the same anchor. On 114 editable development targets, we hold the anchor, editable positions, checkpoint and proposal budget fixed and compare retaining versus removing distinct donor information. In the donor-off condition, the anchor fills the donor slots to preserve the input shape.

SettingReturned-set full-length recoveryMean candidates returned per target
Distinct donors off0.093814.54
Distinct donors retained0.114289.06

Recovery increases by 2.05 percentage points, with a paired 95% confidence interval of [1.49, 2.65]. Results from two fixed checkpoints are averaged within each target. This inference-time intervention identifies a measurable donor contribution to candidate availability and set-level recovery, with the anchor unchanged throughout the comparison.

This intervention directly tests the division of roles: the anchor fixes the sequence starting point, while distinct donors expand the available residue alternatives. The full-length hybrid-pool experiment below further compares the final choices from learned editing, simple editing and retrieval alone.

2. How do the candidates perform beyond the training scorer?

We evaluate frozen candidates with ProSmith, which was used neither for training nor for selection. ProSmith did not contribute to teacher supervision, generation guidance or candidate reranking. This directly addresses the request for evaluation with an independent predictor.

On 99 full-length targets, all three policies have a requested candidate cap of 20: retrieval-only R20, R10 plus up to ten simple-editor sequences, and R10 plus up to ten learned-editor sequences. Simple editing matches the learned output's substitution count per target. The same final selector chooses one candidate from each pool, and ProSmith then evaluates that choice.

Candidate-construction policyProSmith score of final choiceActual candidate total per seed
Retrieval R200.061951,980
R10 + simple editing0.067831,900
R10 + learned editing0.073891,900

Two sampling seeds from one frozen checkpoint are averaged within each target. The learned hybrid has the highest observed mean, 0.07389, compared with 0.06195 for retrieval and 0.06783 for the simple-editing hybrid, while returning the same number of candidates as the latter.

Paired-difference 95% intervals: learned hybrid minus retrieval, [−0.01120, 0.03420]; learned hybrid minus simple-editing hybrid, [−0.00470, 0.01635].

This addition separates the training and evaluation signals: the existing pipeline generates and selects candidates, and an independent model assesses their pairing scores.

3. A concrete example of the central design

We will introduce the method with the following example before presenting the training and sampling details.

Given a substrate molecule, retrieval returns enzyme sequences A, B and C. A is the anchor and supplies the complete starting sequence. B and C are donors and supply residue alternatives at aligned positions. Conditioned on the substrate, the generator first edits designated sites and their sequence neighborhoods, then other permitted positions. A fixed selector ranks the resulting candidates.

A answers “which sequence do we start from?”, B and C answer “which residue alternatives are available?”, and the editor answers “which local changes should we propose for this substrate?”. Donors serve as sequence-information sources from related enzymes. This division is the central design; the losses and sampling settings implement it.

Comments on the paper

1 comment

We thank the reviewers for their specific suggestions on the task definition, methodological contribution and experimental validation. We address these points with fixed-anchor donor interventions, retrieval–editing comparisons with a common candidate cap, independent ProSmith evaluation, modern generation baselines and structural analyses. The new results highlight three findings: ANCHORZYME's full-length native first choices achieve higher structural prediction confidence on the 74 and 78 targets shared with RFD3 and Proteina/Complexa, respectively; fixed-anchor interventions show that distinct donors improve candidate availability and returned-set full-length recovery; and, at matched edit counts, learned editing yields smaller annotated-site geometric deviations on nine of the 11 eligible targets with experimental references. In addition, the learned retrieval–editing hybrid achieves the highest observed mean among the three policies in the independent ProSmith evaluation on 99 full-length targets. We also clarify that the query input requires only substrate SMILES, together with each baseline's input conditions, candidate budget and selection rule. Below, we explain how these additions address each reviewer's questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.