Dear Area Chair and Reviewers,

We sincerely thank you for your thoughtful and thorough evaluation of our paper.

Mentioned reference

We will include these six pieces of literature in the related work section in the final version of this paper since modification to the manuscript is not allowed currently. We plan to test our model on the mentioned dataset.

Complexity of catalysis

Our approach to capture the complexity of enzyme catalysis beyond just the structural interactions is by generation based on reference enzymes. Catalytic capability is not fully decided by binding affinity or docking position but also by some unclear features that cannot modeled directly. The reason to use similar molecules' enzymes as references is these enzymes are known to function for similar molecules and thus contain necessary features even if the feature itself is unclear. The discriminator guidance provides clues to necessary features by serving as a loss.

Predictive model and Binding affinity

The reliability of the predictive model can be checked from its original paper. In our paper, the rule-based method generated enzymes' predicted $k_{cat}$ is reasonable in terms of order, since the totally random generated protein << retrieved proteins $\approx$ mutation of ground truth < ground truth. The rule-based baselines are to evaluate if the predictive model can predict values in reasonable order.

To reflect binding affinity, Vina score is provided in the case study. We are still implementing massive automated evaluation by AutoDock-Vina, since there are frequent cases unsuitable for the input check.

Novelty

The approach of expressing substrate conditions to the model is novel, which includes two novel parts.

1. Retrieval based on substrates

   The idea of using a substrate to find related enzymes and using related emzyes to facilitate new enzyme generation is novel. The approach to align related enzymes and serve as the base of generation is novel. This approach makes zero-shot generation practical.

2. Adopt discriminator in training

   The approach of using a discriminative mode to describe the generation target of catalytic capability is novel.

Dataset

Our dataset is created based on Rhea, which is an expert-curated knowledgebase of chemical and transport reactions of biological interest - and the standard for enzyme and transporter annotation in UniProtKB. Our dataset inherits the quality and coverage from Rhea.

To control the quality, we carefully designed the dataset split.

1. Sequence identity threshold

   EnzyBench partitioned training, validation, and testing sets based on a sequence identity threshold of 50%, but ours is 30%.

2. Distribution of enzymes in split

   Unlike EnzyBench split enzymes based on their third-level categories, we split the enzymes by substrates. We guarantee that no two enzymes in different splits share the same substrate.

There are 34982 entries, so it cannot be regarded as limited, given the fact that high-quality annotated enzyme data is precious.

Substrate-enzyme interaction complexity

Our model does not aim to model the substrate-enzyme interactions directly, but to use retrieved reference enzymes to facilitate generation. We use the target substrate to retrieve enzymes, which contain the desired enzyme's properties. The model does not need to model substrate-enzyme interactions directly but needs to learn generation from references.

Vina score

We provide the docking results with the ground truth here. The Vina score of the ground truth enzyme is -3.036, which is very close to our generated enzyme. It means that our generated enzymes bind the substrate slightly better than the ground truth with no discrepancy. We will add a subfigure in Figure 3.

About the value -3.075 and -3.036. The ligand only has five atoms and thus is small. It is hard for a small ligand to get a high score.

$k_{cat}$ for valid sequences

Thanks for your suggestion. When setting plddt threshold as 70%, our model-generated protein above the threshold has an average predicted $k_{cat}$ of 0.362.

The significance of retrieval

Thank you for identifying our novelty. This design aims to provide indirect substrate information in the format of protein sequences. Substrate-based retrieval is adopted for two significant reasons.

1. Without any protein as input, there is no anchor protein sequence to search for similar sequences.
2. With only a target substrate as input, the retrieval method has to use this molecule as the key.

Substrate concatenation

It provides direct information on the target substrate by serving as a single token prompt persistent in the diffusion process. The sequence to generate has known positions from two dimensions then:

1. The rows of aligned sequences.
2. The column of the target molecule's embedding.

7 tasks

Please refer to the response to Reviewer MkMZ.

Dear Area Chair and Reviewers,

We sincerely thank you for your thoughtful evaluation.

Retrieval quality

The quality of retrieved enzymes is reflected in the baseline method named 'Retrieved' in Table 2. That row is as follows:

Random versus unconditionally generated enzymes

1. 'Random' are randomly generated amino acid sequences. They are not randomly selected enzymes.
2. Unconditional generated proteins reflect the distribution of the training set of the baseline method. These models are trained to generate natural-like proteins.

The comparison reflects the features of $k_{cat}$ prediction model in evaluation. A more natural-like protein obtains a higher score than a pure random amino acid sequence.

Retrieved enzymes versus ground truth

The retrieval is from a different set which contains no ground truth enzymes. The retrieval module can never see the ground truth proteins.

Line 297: Natural versus unconditionally generated enzymes

1. Natural enzymes are ground truth, which are the enzymes for target substrates.
2. Unconditionally generated enzymes are generated by baseline models. They do not exist in nature.

Novelty of retrieval and substrate 3D information

Our novelty is searching similar substrates and using their enzymes as reference for the generation model, not the metric about molecule similarity. We use a well-recognized method to calculate the similarity between molecules in the retrieval database. Morgan Fingerprint (MF) is an easy and recognized method to map molecules into the same vector space as representation. Tanimoto distance is designed to measure the similarity between vectors like MF.

MF represents structure without relying on 3D coordinates. It encodes atom environments in a molecule, capturing local atomic neighborhoods. It can indicate whether certain functional groups or motifs are present in a molecule. It allows the comparison of molecules based on shared substructures.

Identity exceeding 30%

It means using mmseqs2 with --min-seq-id 0.3. In Line 143, it means any two protein sequences from different subsets (training set, validation set, or testing set) have a maximum of 30% position overlap. This restriction is to avoid data leakage in protein sequence dataset split. Proteins are not relevant under the threshold.

Eq 5,6 and Line [157-163 Col 2]

Notation for protein:

1. $x$ is a single protein.
2. $\mathbb{P}^{(m)}$ is a set of proteins related to molecule $m$. If the molecule is in the retrieval database, these proteins are the enzymes of the molecule; If the molecule is not in the retrieval database, these proteins are the enzymes of similar molecules.

Equation 6 is a definition relying on equation 5: if the target molecule is not in the retrieval dataset, the proteins for some molecules in the dataset will be collected and serve as the proteins for the target molecule.

Line [157-163 Col 2: ] is the process of retrieving proteins for a target molecule.

1. Calculate the similarity between the target molecule and molecules in the database.
2. Order molecules in the database descending by similarity.
3. Get all the enzymes of each molecule in database in order until a desired amount is obtained.
4. If a molecule has too many enzymes, only part of them is selected. It is to prevent all retrieved proteins from being enzymes of a single molecule.

Diversity and novelty

1. The diversity of generated enzymes for the same substrate.

   It is measured by the metric #clusters. Enzymes for the same target substrate are generated at first. Then they are clustered by a threshold of 30% identity, which means every two proteins from different clusters do not have an overlap of 30%. The number of clusters is recorded. Ground truth enzymes of the same substrate are also clustered, and the number of clusters is obtained. The absolute difference between clusters is the resemblance of diversity to natural distribution. If the generated enzymes have approximately the same number of clusters as natural ones, they have a natural distribution in terms of diversity.

2. The novelty of generated enzymes.

   It is measured by the metric BLASTp. It can be regarded as the distance from a protein to the nearest in all other natural proteins. It is the identity of the generated enzyme with the nearest different known protein, which is obtained by BLASTp2 in the SwissProt database. We also get this value of natural enzymes, and the absolute difference between it and of generated proteins is calculated. If the distance of generated enzymes is similar to that of natural ones, they have a natural distribution in novelty.

Average docking score

We will calculate the docking score in the next version of the manuscript. Runnig docking needs to manually adjust the protein structure file with the substrates'.

Dear Area Chair and Reviewers,

We sincerely thank you for your thoughtful and thorough evaluation of our paper.

UniKP for evaluation

Thanks for your comment. Using the predictive model for scoring is what we do in practice to filter better proteins. Before the wet lab experiment, proteins are scored by the predictive model, and proteins with a higher score are tested with higher priority. The advances in predicted scores suggest higher priority for wet lab experiments. We plan to test the designed enzymes in the wet lab, and the selection is based on the score.

The evaluation of protein properties depends on machine learning model predictions, particularly when no traditional methods are available as alternatives. The direct in silico evaluation of catalytic capability is by adopting the predictive model, and it is also a necessary step for filtering before wet lab experiments. This process is also adopted in protein design tasks for different features with predictive models for different properties.

Guidance for kcat improvement

In this manuscript, we want to show how far the $k_{cat}$ can achieve, because a higher predicted $k_{cat}$ value makes designed proteins a higher priority for wet lab experiments. In real-world enzyme design, experts can decide the preference for predicted $k_{cat}$ or plddt, and then use a version of the model trained with or without guidance. We show the comprehensive effect of the module in facilitating experts' selection of model versions.

AlphaFold3’s support for ligands to filter

Thank you for your suggestion. We will use more structure prediction models in evaluation in future versions of this manuscript.

Sequence similarity of the binding pocket in retrieved and the catalytic site

Thanks for your suggestion. The binding sites are usually not adjacent in sequence, so the sequence's similarity cannot be compared. The binding sites of an enzyme in the Uniprot annotation are usually discrete positions, which are not adjacent to each other in the protein sequence. Therefore, the sequence similarity cannot be compared based on several unconnected positions.

There is no enough annotation related to every enzyme's catalytic site. Our dataset is created based on Rhea, which is an expert-curated knowledgebase of chemical and transport reactions of biological interest. However, there is no pocket or catalytic site annotation on the enzymes.

Influence of database size

Thanks for your suggestion. We will use the full enzyme dataset in real-world new enzyme designing tasks. Of course, the number of substrates in the database decides the quality of retrieved proteins. Therefore, using the largest reliable enzyme knowledge base as the retrieval database is required in real-world tasks.

Max kcat for Table 1

The max $k_{cat}$ value in Table 1 is in the table below.

Distillation of the UniKP into the discriminator

Thanks for your suggestion. In real-world tasks, we will consider utilizing other predictive models, including UniKP, as the discriminator as long as it is end-to-end and preserves the gradient. We will also try to use predictions for further rounds of training. The reason we do not adopt it in the current manuscript is that the evaluation relies on UniKP, which should not be seen by modules in training and generation.

Dear Area Chair and Reviewers,

We sincerely thank you for your thoughtful and thorough evaluation of our paper. Below are our detailed responses to your comments:

Source of the 7 substrates used in the experiments and reason of selection

Thanks for your comment. We selected these substrates because they are common and important in enzymatic reactions. Then, it is worthy to design new enzymes for these substrates.

1. Sepiapterin [1]

   Reason for Designing Enzymes: Enhancing the efficiency and specificity of sepiapterin reductase can improve tetrahydrobiopterin (BH4) production, which is crucial for neurotransmitter synthesis and nitric oxide production. [1]

   [1] Thöny B, Auerbach G, Blau N. Tetrahydrobiopterin biosynthesis, regeneration and functions. Biochem J. 2000 Apr 1;347 Pt 1(Pt 1):1-16.

2. Propylene Oxide [2]

   Reason for Designing Enzymes: Engineering epoxide hydrolases or monooxygenases can provide higher enantioselectivity and stability under industrial conditions for the production of chiral intermediates in pharmaceuticals. [2]

   [2] Erik J de Vries, Dick B Janssen, Biocatalytic conversion of epoxides. Current Opinion in Biotechnology. Volume 14, Issue 4, 2003, Pages 414-420.

3. Levoglucosan [3]

   Reason for Designing Enzymes: Developing specific glucosidases can enable efficient hydrolysis of levoglucosan into fermentable sugars, facilitating biofuel production from biomass pyrolysis products. [3]

   [3] Donovan S. Layton, Avanthi Ajjarapu, Dong Won Choi, Laura R. Jarboe. Engineering ethanologenic Escherichia coli for levoglucosan utilization. Bioresource Technology, Volume 102, Issue 17, 2011, Pages 8318-8322.

4. cGMP [4]

   Reason for Designing Enzymes: Developing specific glucosidases can enable efficient hydrolysis of levoglucosan into fermentable sugars, facilitating biofuel production from biomass pyrolysis products. [4]

   [4] Lucas KA, Pitari GM, Kazerounian S, Ruiz-Stewart I, Park J, Schulz S, Chepenik KP, Waldman SA. Guanylyl cyclases and signaling by cyclic GMP. Pharmacol Rev. 2000 Sep;52(3):375-414.

5. L-Proline (L-Pro) [5]

   Reason for Designing Enzymes: Engineering proline racemase or proline dehydrogenase can enhance the production of D-proline, a valuable chiral building block in pharmaceutical synthesis. [5]

   [5] Tanner JJ. Structural biology of proline catabolism. Amino Acids. 2008 Nov;35(4):719-30.

6. Pyridoxine [6]

   Reason for Designing Enzymes: Designing enzymes for pyridoxine is important for enhancing its role in oxidative stress resistance by modulating its singlet oxygen quenching properties, which can be applied to improving fungal resilience, developing antioxidant therapies, and advancing fluorescence-based imaging techniques. [6]

   [6] Bilski P, Li MY, Ehrenshaft M, Daub ME, Chignell CF. Vitamin B6 (pyridoxine) and its derivatives are efficient singlet oxygen quenchers and potential fungal antioxidants. Photochem Photobiol. 2000 Feb;71(2):129-34.

7. leukotriene A4(1-) [7]

   Reason for Designing Enzymes: Developing selective leukotriene A4 hydrolase inhibitors can lead to new anti-inflammatory drugs with fewer side effects.

   [7] Haeggström JZ. Structure, function, and regulation of leukotriene A4 hydrolase. Am J Respir Crit Care Med. 2000 Feb;161(2 Pt 2):S25-31. doi: 10.1164/ajrccm.161.supplement_1.ltta-6.

To remove structural elements such as signal peptides and transmembrane domains

Thanks for your suggestion. This process can provide the model with clearer sequence patterns by eliminating undetermined sections. In the future version, we will test the performance after removing signal peptides and transmembrane domains in the whole dataset.

Detail of unconditional generation baselines

ProGen2 and ProtGPT2: We utilized the pre-trained weights for both ProGen2 and ProtGPT2 to directly generate sequences with a maximum length of 1024. No information related to target substrates is provided to either model. These models serve as baselines for the capability of protein language models to generate sequences without specific functional guidance.