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:

W1 - The application scenario is unclear. I don't think the proposed retrieval method could find relative enzymes in the database just according to the substrate similarity. Our model aims to generate non-natural enzymes, and the retrieval works because the substrate reflects enzyme properties.

Thank you for the comment. Our model aims to generate non-natural enzymes, and the retrieval works because the substrate reflects enzyme properties.

1. Generate non-natural enzymes: The enzymes we aim to generate are indeed non-natural, which is why we designed the zero-shot inference approach.
2. Substrates reflect the desired properties from retrieved enzymes.
   • Substrates reflect enzymes' properties: It is based on the observation that enzymes catalyzing highly similar substrates may also share some similarities [1].
   • Same physics and chemistry rules for natural or non-natural enzymes: While non-natural enzymes may have unique properties, they still adhere to the same physical and chemical principles as natural enzymes. Therefore, their folding and interaction with substrates are governed by the same fundamental rules.
   • Substrate-based retrieval keeps proteins' properties: The retrieval method searches for related enzymes based on the substrates’ similarity to the target substrate rather than similarity to the expected generated enzyme. This approach ensures that the differences between natural and non-natural enzymes do not adversely affect the retrieval results, while the generated enzyme will incorporate beneficial features from the retrieved ones.

Consequently, relevant enzymes can still be identified to assist in generating non-natural enzymes.

[1] Samuel Goldman, Ria Das, Kevin K. Yang, and Connor W. Coley. Machine learning modeling of family wide enzyme-substrate specificity screens. PLOS Computational Biology, 18(2):1–20, 02 2022.ß

W2(1) - Unconditional generation is not fair. Why not use ProGen? Which size of ProGen2 is utilized? Why the pLDDT of ProGen2 is too low?Unconditional generation baselines are for demonstration; ProGen cannot take substrate condition; ProGen2 is 6.4B; Low pLDDT because of lack of fine-tuning.

Thank you for the comment. We will discuss them here.

1. Unconditional generation baselines are for performance demonstration only: Both ProGen2 and ProtGPT2 are indeed unconditional models, and as such, they do not have explicit modules for controllable enzyme design. We include these models as baselines to demonstrate how unconditional generation performs in the context of enzyme design.
2. ProGen cannot take substrate condition: Regarding ProGen, it does allow for controllable generation, but this is achieved by inputting control tags (e.g., Immunoglobulin, Chorismate mutase, Glucosaminidase, Phage lysozyme), which represent prior knowledge about the protein being designed. However, these tags are specific to predefined proteins, meaning ProGen cannot generate enzymes for new substrates without such tags. Consequently, ProGen does not support zero-shot enzyme generation for novel substrates, which is a key goal of our approach.
3. ProGen2 in baseline is the 6.4B parameters version: The version of ProGen2 we used is the ProGen2-xlarge with 6.4B parameters.
4. ProGen2's low pLDDT is because of the lack of fine-tuning: The low pLDDT observed in ProGen2 is likely because it was not fine-tuned on any specific protein families, making its output more random compared to fine-tuned models that specialize in particular types of proteins.

Thanks for the authors' response! After going through the rebuttals, I still have the following concerns:

W1: about goal of designing non-natural enzymes : If the author tends to find novel but natural enzymes, I can think about the proposed method is doable. If it's non-natural enzymes, it's hard for me to trust. Even if the author said the substrates should share some physical constraints, non-natural enzymes have very unique properties that can not be found in the current database. Usually people targeting at designing non-natural enzymes should carefully design the enzyme pocket with strictly designed physical and biochemical rules from domain experts, like luciferases [1]. I can't imagine these rules can be extracted from the current database. Besides, the reported results showed low pLDDT, further validating my concerns.

[1] De novo design of luciferases using deep learning.Nature. 2023.

W2: ProGen cannot take substrate condition: Even though ProGen cannot take substrate condition, you can describe the curresponding enzyme category and property to promote conditional generation.

W2 (2): Enzymes with low pLDDT can be foldable and functional. I agree some cases with low pLDDT can still fold stably, but these cases should be validated in wet lab like the author mentioned in [2]. However, in most cases, the general trend is higher pLDDT leads to candidates that have enzymatic activity with higher probability. If the author check the experimentally confirmed enzymes in PDB, you should find most enzymes with AF2 pLDDT higher than 80.

[2] Sarah Alamdari, Nitya Thakkar, Rianne van den Berg, Alex X. Lu, Nicolo Fusi, Ava P. Amini, and Kevin K. Yang. Protein generation with evolutionary diffusion: sequence is all you need. bioRxiv, 2023.

Overall, I think I still have significant concerns to this paper.

Q1 - How the model generates enzyme sequences in extreme cases such as when there are no retrieval results, and how to ensure the quality of the generation. The molecule encoder can help, and discriminative models filter the generated enzymes.

Thank you for the comment. The molecule encoder can help when the retrieved results are not helpful, and discriminative models filter the generated enzymes for quality control.

1. Relying on molecule encoder.
   • Fixed number of retrieval results: We ensure the model retrieves a fixed number of enzymes, maintaining consistency with the training setup. Different model variations can retrieve varying numbers of enzymes, but even if the retrieved results are unhelpful—or effectively equivalent to “no retrieval results”—our model leverages the molecule encoder.
   • Molecule encoder: This encoder directly processes the target substrate, enabling the generative model to produce enzyme sequences even without meaningful retrieval results.
2. Quality control.
   • kcat: We utilize a discriminative model to predict the catalytic capability of the generated proteins for the target substrate. Filtering can be done based on the value.
   • Foldability: We predicted the enzyme structure with a folding model. Filtering can be done based on the structure and pLDDT.
   • Docking check: We dock the generated enzymes and the target molecule. The docking score and docked structure provide a reference for filtering.

To ensure quality, these evaluation steps assess and refine the quality of the generation.

Q2 - During the model’s sampling phase, how to align the retrieval results to target generation? If all retrieval results are aligned with the first retrieval result, could this potentially introduce errors. The retrieved results are treated equally in the alignment algorithm.

Thank you for the comment. The retrieved results are not aligned to a single sequence. They are treated equally in the alignment algorithm.

1. Align by ClustalW algorithm.
   • ClustalW: It begins by computing pairwise alignments to create a distance matrix, then builds a guide tree to iteratively align sequences or groups. It optimizes scoring with weighted gaps and substitution matrices to achieve biologically meaningful alignments.
   • Not aligned to a single sequence: The retrieved results are treated equally within the alignment algorithm. ClustalW ensures that the order of sequences in the input does not affect the final alignment. This is achieved through the guide tree, which determines the alignment order based only on sequence similarity.
2. Generation process: To initialize the generation process, a fully masked sequence is added at the top of the MSA, serving as the starting state of the sequence to be generated. The MSA, including the retrieval results and the masked sequence, is then fed into the generative model. The generative model iteratively fills in some of the masked positions in the first sequence during each iteration.

Since the retrieval results are not aligned to one specific sequence, this approach avoids introducing errors from unreasonable alignment to a single sequence.

Q3 - The details of how you modified the NOS in the baselines. Only replace the original guidance function with the substrate-enzyme relation scorer in our model.

Thank you for the comment. We only replace the original guidance function with the substrate-enzyme relation scorer in our model.

1. Original NOS.
   • Protein-protein binding affinity scorer: The original NOS is trained as described in its paper, using a discriminator to score the binding affinity between an antibody and an antigen (two protein sequences). This discriminator is appended after the generative model, and the score is used as a loss during training.
   • Inference: During inference, NOS updates the generator's parameters with the current testing set input for ten iterations. Subsequently, inference and sampling are conducted as in a typical discrete diffusion model. This process is repeated iteratively until the full sequence is sampled.
2. Our modification.
   • Substrate-enzyme relation scorer: We replaced the original discriminative model with the enzyme-substrate probability scoring model used in our approach. Additionally, the input for the generator was changed from the target protein sequence to the target substrate molecule.
   • Keep others the same: All other steps, including training and inference processes, remained the same.

W3 - The reasons for 7 tasks selection were not provided. The unconditional generation results outperform the conditional generation results. The 7 target substrates are important in metabolism, and massive pre-trained methods may outperform unsuitable conditioned methods.

Thank you for the comment. The seven tasks were chosen because they involve designing enzymes for seven well-known small molecules that play critical roles in metabolic pathways. We will also introduce why unconditional generation outperforms conditional generation.

1. Task selection.
   • Relevance: These substrates are significant in metabolism and have practical importance in biochemical research and applications.
   • Independence: They are not included in the training or validation sets of any of the models, ensuring the evaluation is unbiased and represents a true zero-shot scenario.
   • Ground Truth: Known enzymes for these substrates exist, providing ground truth for comparison with the enzymes generated by the models.
2. Reasons for superiority in unconditional methods.
   • Unconditional methods with diverse training sets: For unconditional generation baselines such as ProGen2 and ProtGPT2, we used the weights provided by their authors. The pre-training dataset of ProGen2 and ProtGPT2 is extensive and diverse, which may lead these models to generate proteins with functional properties suitable for catalysis, even without explicit guidance. These large-scale protein sequence datasets for pre-training include ground truth enzymes for some test substrates. They are likely to generate proteins well from all perspectives.
   • Some conditional methods only generate protein binder: Some conditional methods, including LigandMPNN, generate protein binder, not the enzyme, for the target. Therefore, they perform badly in kcat.
   • kcat variability: The catalytic activity evaluation (via kcat) is highly sensitive. A single or double point mutation in the catalytic site can lead to significant changes, sometimes resulting in a 5-fold improvement. This inherent variability in kcat measurements contributes to the observed inconsistencies.
   • Multiple functional sites in one protein: The unconditional-generated proteins sometimes have multiple catalytic sites and pockets, and thus have high kcat. When the conditional-generated enzymes cannot leverage the target well, their generated protein will have no function.

Considering the massive pre-training in unconditional methods and the weakness of modeling substrate conditions for conditional baselines, the result is not unpredictable.

W4 - The model's foldability and catalytic performance of sequences should be further analyzed and balanced. We provide versions of the model, and wet lab practitioners can do the selection.

Thank you for the comment. We provide versions of the model in the ablation study, and wet lab practitioners can select the model.

• Metrics should be evaluated in wet lab: In wet lab, practitioners prioritize multiple metrics depending on the specific application. The designed enzymes are tested in initial experiments, and the importance of each metric is then tailored to the task at hand. Therefore, we do not know the importance of the metrics before the wet lab, and it is hard to select the best version of the balanced model.
• Candidate models as the solution: We provide a range of models trained with different numbers of retrieved sequences. These models allow domain experts to choose the most suitable designs for their experimental objectives. For example, in applications where catalytic activity is paramount, sequences with higher Kcat values may be prioritized. Conversely, for tasks requiring robust folding in physiological conditions, foldability may take precedence. Moreover, as shown in Figure 3(f), the structures of our designed enzymes demonstrate satisfactory docking positions, further supporting their functional viability.

W5 - The paper does not include a reproducibility statement. We append it after the conclusion section.

Thank you for the comment. We have added a Reproducibility Statement section between the Conclusion and References in the revised manuscript. The content of the section is as follows:

We have described all necessary details to ensure reproducibility, including dataset information, model architectures, hyperparameters, and evaluation protocols. The code is available at https://anonymous.4open.science/r/SENZ-2BE1/.

We also highlighted it in the manuscript.

W2 - The protein language models for the unconditional generation and sequence generation models are relatively outdated. Moreover, the details of the baseline are not explained in the paper. The baselines are selected based on their importance and relevance, and we appended the details of baselines.

Thank you for the comment. We will introduce the selection of baselines and their details.

1. Select mainstream approaches as baselines.
   • Unconditional generation models: The protein language models, including ProGen2 (Nov. 2023 on Cell systems) and ProtGPT2 (July 2022 on Nature Communications), are unconditional protein generation models with high impact. The protein language models are used as baselines to evaluate unconditional generation performance.
   • Conditional generation models: ZymCTRL is selected because it is designed for enzyme generation, although it needs the Enzyme Commission (EC) number.
2. Details.
   • ProGen2 and ProtGPT2: We utilized the pre-trained weights for both models to generate sequences with a maximum length of 1024. These models serve as benchmarks for the capability of protein language models to generate sequences without specific functional guidance.
   • ZymCTRL: This model employs pre-trained weights and uses the Enzyme Commission (EC) number as a prompt for the autoregressive generation process. It is worth noting that the EC number provides more detailed information about enzymatic function compared to the substrate alone, offering this baseline an advantage in generating enzyme sequences for the given tasks.
   • NOS: We trained NOS following the methodology of its original paper. The original NOS framework uses a discriminator to score the binding affinity between an antibody and an antigen (two protein sequences). We replaced the original discriminator with our enzyme-substrate probability scoring model in our adaptation. Furthermore, we replaced the target protein sequence input with the target substrate molecule. During inference, the NOS generator is updated iteratively for 10 steps using the test set input before sampling, following a discrete diffusion model for sequence generation, as described in the original paper. These adjustments allow NOS to generate enzymes in our setting while preserving its original generative framework.
   • LigandMPNN: This reverse folding model generates a protein sequence based on a protein-ligand complex structure. To adapt it for our task, we randomly generated protein sequences (length: 1024) and predicted their structures using ESMFold [1]. Using RDKit, we generated the structure of the target substrate, and NeuralPLexer [2] was employed to dock the substrate with the predicted protein structure, creating a complex structure. The resulting complex was then input into LigandMPNN for sequence redesign.

[1] Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Salvatore Candido, and Alexander Rives. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123–1130, 2023.

[2] Zhuoran Qiao, Weili Nie, Arash Vahdat, Thomas F. Miller, and Animashree Anandkumar. State-specific protein-ligand complex structure prediction with a multiscale deep generative model. Nature Machine Intelligence, 6(2):195–208, February 2024.

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:

W1 - The model's foundation lacks innovation, as it uses the existing method EvodiffMSA-OADM. The model only considers the sequence level without extending to the structural level. The novelty is to generate enzymes from the substrate, and our model considers structure by MSA.

Thank you for the comment. We will introduce the novelty of our model and its consideration of structure.

1. Innovation.
   • EvoDiff is protein sequence conditioned only: EvodiffMSA is indeed capable of generating new sequences, but it either does so without specific conditions or with only partial sequence information as input. Evodiff per se cannot generate substrate-specific enzymes because it does not take substrate as input.
   • Our model is molecule-conditioned: Our approach enables the direct generation of enzymes from the target substrate in a zero-shot setting, which is a significant difference. We have introduced a retrieval module, molecule encoder, and guidance mechanism to achieve it.
2. Consideration of structure.
   • Structure information is in MSA: As ESMFold [1] showed, MSA can provide enough information for protein structure prediction. In our model, we collect the MSA, from which the generator extracts the structure information. Moreover, since protein structures can be predicted from their sequences using folding models, it has been demonstrated that a protein’s sequence inherently carries information about its structure.
   • Results have valid structure: As evidence of our model’s effectiveness, the generated enzyme shows favorable docking positions, as shown in Figure 3(f), which indicates a valid structure.

[1] Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Salvatore Candido, and Alexander Rives. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123–1130, 2023.

Thank you for the comment. We want to introduce the novelty of our paper further from the perspective of task setting.

1. Zero-shot enzyme generation.
   • Conditioned on substrate: Our model is designed to generate an enzyme given only a small molecule, without any natural language description or specified tags for the desired protein family.
   • Target substrate is new: This target substrate should never been seen by the model before testing.
   • Reason for the zero-shot setting: Some small molecules do not currently have good enzymes. We want to make the model able to generate enzymes for those substrates without enzymes.
2. The reason why our model can achieve zero-shot enzyme generation.
   • Molecule input provides more learnable information: Compared with the label-conditioned generation, molecule condition provides more sharable information. Takes L-glutamate as a target substrate, for example. Its SMILES representation [NH3+][C@@H](CCC([O-])=O)C([O-])=O contains the atom and bond data, enabling the model to learn from other molecules with similar atoms and bonds to leverage the input. Label-conditioned models heavily rely on the information learned from each label because labels have less inner relation with each other.
   • Retrieved proteins provide references for generation: Although the input is only a target substrate, we can further leverage the known enzymes more than in training. The retrieved enzymes help to represent the feature of the target molecule indirectly because they are enzymes of the substrates that share similarities to the target. The retrieved enzymes and the desired output are all proteins, making the generation as easy as extracting from references.

Thanks for the comprehensive response. I have also reviewed the questions and rebuttals from other reviewers. Overall, I now have a better understanding of the model's implementation. But I still have concerns regarding the validity of the evaluation metrics and the quality of the generated outputs. The Kcat predicted by UniKP is a relatively easy-to-trick metric. Did the authors consider using metrics related to enzyme and substrate binding for evaluation? I would like to maintain my score given the current quality of the paper.

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:

W1 - The scores of generated enzymes are much higher than the ground truth. It is common for artificial enzymes to outperform natural counterparts in kcat.

Thank you for the comment. We will explain the high variance in kcat and discuss why generated enzymes outperform ground truth, which is natural enzymes.

1. The high variance in kcat is not unexpected.
   • Sensitive: kcat measurements are known to be sensitive, and it is common to observe a significant (5-fold) improvement with just 1-2 point mutations.
   • Factors influencing kcat: If the catalytic site of an enzyme is replaced with amino acids that do not support catalysis, the enzyme will lose its catalytic ability, resulting in a much lower kcat value.
2. Ground truth enzymes are not necessarily the best.
   • Nature enzymes get the function by evolution: Natural enzymes have evolved their current function and sequence through evolutionary processes that filter out non-functional variants. There is always room for designing more efficient alternatives.
   • Artificial enzymes may outperform their natural counterparts: Enzyme engineering methods have produced better enzymes [1]. Frances H. Arnold was awarded the Nobel Prize for Chemistry 2018 “for the directed evolution of enzymes." Those enzyme optimization experiments are the evidence that enzymes are optimizable.

Therefore, the high variance in Kcat results is not unexpected, and natural proteins are not necessarily the most efficient enzymes for their respective substrates.

[1] Karl-Erich Jaeger, Thorsten Eggert. Enantioselective biocatalysis optimized by directed evolution. Current Opinion in Biotechnology, Volume 15, Issue 4, 2004, Pages 305-313.

W2 - Confusing annotations about probability p and variable x. x is the extreme case of p.

Thank you for the comment. We will explain why we sometimes deduce p to x.

• x: Variable x is a 2D matrix where each row represents a position in the protein sequence, and each column is a one-hot representation of the amino acids, as we highlighted in red on Page 3, lines 115-118.
• p: The probability matrix p has the same shape as x, but instead of one-hot encoding, each column in p represents the probability distribution of selecting each amino acid at a given position.
• Relation between x and p: x is a deterministic matrix with 1 at the specific amino acid for each position, while p represents a probabilistic distribution. When p corresponds to a known distribution from the training set, we deduce x from p as the extreme case, where each position in x contains either 0 or 1.

To summarize, x and p have the same shape, with x representing the deterministic case where each position is restricted to a single amino acid choice (either 0 or 1), while p encodes the probabilistic distribution over amino acids.

Q1 - Why is the score much higher than the ground truth? A way to directly compare the generated enzymes to the ground truth. Compare sequence similarity.

Thank you for the comment. The reason why the generated protein is better than the ground truth can be found in response to W1, and we will introduce how to compare the generated protein to the ground truth directly.

1. The reason generated is better than the ground truth.
   • Response to W1: Please refer to the explanation in response to W1.
2. Directly compare the generated protein to the ground truth.
   • Comparing proteins themselves is unnecessary: Regarding direct comparisons with the ground truth, we aim to generate enzymes with catalytic activity, which does not necessarily imply high sequence similarity to the ground truth enzyme.
   • Sequence identity for direct comparison: To directly compare the generated enzymes to the ground truth, we align the two protein sequences and calculate the identity, which measures the ratio of positions with identical amino acids. For targeting methylphosphonate(1-), the sequence identities between the ground truth enzyme and those generated by models are shown in the following table in percentage.

Typically, a sequence identity below 30% indicates significant differences between two proteins.

Q2 - Re-train the model or just change the generation stage in the ablation study? Re-train the model.

Thank you for the question. Yes, the model has been re-trained with each setting in the ablation study. Additionally, we have trained the model both with and without guidance to assess its impact on performance.

Thank you to the authors for revising the manuscript and providing a link to the code base.

While it is great to provide the code as-is, and certainly better than nothing, I found it very bare-bones and lacking in documentation. There is not even a filled in README. That makes it very difficult to reproduce the results, even if the src/ is available. I recommend the authors to add documentation, including instructions, so that potential users could actually benefit from having the code and perhaps even re-run some of the models trained in the paper.

I also see that there are significant concerns from myself and other reviewers that have only been addressed in the comments here, but not in the main text. I think it is very valuable to consider that if things are not clear to us, perhaps they could be clarified further in the paper, and I would recommend to make changes in the main text to include some of the relevant additional information you have given us below.

I have not updated my score.

Q1 - For the same substrate, how different are the proteins generated via the different benchmarked methods? We compared them in terms of sequence and structural similarity.

Thank you for the comment. In Section 4.6, CASE STUDY TARGETING METHYLPHOSPHONATE(1-), proteins are generated by different models, all targeting the same substrate, methylphosphonate(1-). We will compare their sequence and structure similarity.

1. Sequence similarity.

   • Metric: To directly compare the proteins generated by the baseline methods, we align their sequences and calculate the identity, which is the ratio of positions with identical amino acids.

   • Result: Generally, an identity lower than 30% indicates no significant similarity in sequence. The sequence identities between the proteins targeting methylphosphonate(1-) are shown in the table below.

   	ProtGPT2	ProGen2	ZymCTRL	NOS	LigandMPNN	SENZ
   ProtGPT2
   ProGen2	21.86
   ZymCTRL	21.03	19.13
   NOS	20.18	20.61	16.44
   LigandMPNN	20.23	21.65	20.68	21.81
   SENZ	20.00	20.16	15.63	21.82	18.76

2. Structure similarity.

   • Metric: For structural comparison, we use TM-align to compare the 3D protein structures. TM-align aligns the Cα atoms of protein backbones to find the best structural match, even when sequence identity is low.
   • Result: The TM-scores between each pair of generated enzymes targeting methylphosphonate(1-) are also provided in the table below. For the value in row A, column B, the TM-score is normalized by the length of protein A. TM-scores range from 0 to 1, where a score between 0.0 and 0.30 typically indicates no structural relationship.

   	ProtGPT2	ProGen2	ZymCTRL	NOS	LigandMPNN	SENZ
   ProtGPT2		0.23789	0.12233	0.12550	0.18619	0.12544
   ProGen2	0.22836		0.11656	0.13564	0.18303	0.13386
   ZymCTRL	0.26331	0.25968		0.26866	0.26289	0.33121
   NOS	0.20648	0.22723	0.20648		0.24652	0.23912
   LigandMPNN	0.20929	0.21419	0.14044	0.17088		0.15116
   SENZ	0.21737	0.24737	0.27572	0.25447	0.23319

Q2 - How to handle potential post-translational modifications that could lead to improved/decreased enzyme activity relative to the reference enzyme? It is handled indirectly by the oracle discriminator.

Thank you for the insightful question. Post-translational modifications (PTMs), such as phosphorylation, can introduce structural changes and further affect stability. Currently, there are no specific predictors for the effects of PTMs on enzymatic kinetics. As a result, we rely on an available oracle that has been trained on the substrate-enzyme relationship dataset, which indirectly accounts for the effects of PTMs on enzyme activity.

Q3 - How to incorporate not-used information in future work? The 2D structure of the substrate is handled, and the full reaction can be encoded in the same way.

Thank you for the comment. We will introduce how we leverage 2D structure of substrate and how to take advantage of more information in future generation of the model.

1. Already leverage 2D structure of substrate.
   • Directly: The 2D structure of the substrate is embedded using a graph neural network in the generator.
   • Indirectly: It is implicitly considered in the retrieval process, as the Morgen Fingerprint captures some aspects of local 2D structures.
2. Integrate information of full reaction in the future.
   • Retrieval by reaction: For future work, we plan to design a more sophisticated retrieval strategy to retrieve enzymes based on not only substrate but also all molecules in the reaction, including products and intermediates.
   • Full reaction encoding: The current encoder for the substrate can also encode other molecules in full reaction.

These enhancements can be integrated directly into the current SENZ model to improve data efficiency and overall performance.

Q4 - What are the failure modes of SENZ? Invalid SMILES input.

Thank you for the comment. One of the primary failure modes of SENZ arises from invalid SMILES (Simplified Molecular Input Line Entry Specification) input. The substrate’s SMILES representation must be valid to ensure meaningful enzyme generation. If the user lacks knowledge of SMILES and inputs a string that does not correspond to a real molecule, SENZ may produce enzymes that have no functional or structural relevance.

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:

W1 - Docking score and kcat are insufficient metrics, and rediscovery tasks are recommended. The kcat is important for enzyme generation, and hitting any exact protein has a low likelihood but is unnecessary.

Thank you for the comment. Regarding the proposed rediscovery task, we acknowledge its potential value. We will discuss why the metric of kcat is essential and the role of recovery task in our setting.

1. kcat evaluates catalytic capability and is thus crucial for enzyme generation task.
   • Definition: The catalytic constant, kcat, measures the maximum number of substrate molecules an enzyme can convert into the product per unit time under substrate saturation. It reflects an enzyme’s turnover number and efficiency in catalyzing reactions.
   • Relation to enzyme generation: kcat assesses the efficiency of model-generated enzymes in catalyzing target substrate, and the underlying model’s performance.
2. The Rediscovery task has a low likelihood of hitting and is unnecessary.
   • Low likelihood of hitting: It is because of the vastness of protein space. For instance, a protein sequence of 1000 amino acids has a theoretical space of $20^{1000}$ possible combinations. Thus, the chance of precisely replicating a small set of known enzymes for a given substrate is minimal.
   • Unnecessary: Functional enzymes are often not confined to a single sequence, as variations at non-critical positions typically do not significantly impact performance. Therefore, a valid enzyme is not necessarily restricted to a few known solutions.

W2 - It was not fully clear how SENZ sets itself apart from existing work in enzyme design. SENZ is different in its task and leverage of information.

Thank you for the comment. We will introduce the novelty of SENZ from the perspective of task setting and information utilization.

1. Innovation in task setting.
   • No need for protein-related information: For instance, ZymCTRL [1] is a GPT-2-based model that generates enzymes using the Enzyme Commission (EC) number as a prompt. In contrast, SENZ uses only the target substrate as input, eliminating the reliance on prior knowledge such as EC numbers. This allows SENZ to design enzymes without predefined enzymatic information.
   • Specified for enzymes with catalyzing capability, not binding: Other conditional generation models, such as LigandMPNN [2], focus on designing binders for small molecules or antibody-antigen interactions. These models aim to create proteins that bind tightly to their targets rather than accelerating a specific biochemical reaction. SENZ, however, directly addresses the enzyme design problem, where the objective is to catalyze reactions rather than simply bind substrates.
2. Innovation in leveraging information.
   • Retrieve MSA from substrates: Most protein models collect multiple sequence alignments (MSAs) based on protein sequence similarity, typically by searching for sequences similar to a target protein. SENZ, however, constructs MSAs based on substrate similarity, offering a unique perspective in modeling enzyme activity.
   • Guidance to model catalytic relationship: SENZ integrates discriminative guidance models during the training phase to describe the catalytic relationship to its generative diffusion model, as existing works primarily apply guidance only during the inference stage of diffusion models.

[1] Geraldene Munsamy, Sebastian Lindner, Philipp Lorenz, and Noelia Ferruz. Zymctrl: a conditional language model for the controllable generation of artificial enzymes. In Machine Learning for Structural Biology Workshop. NeurIPS 2022, 2022.

[2] Justas Dauparas, Gyu Rie Lee, Robert Pecoraro, Linna An, Ivan Anishchenko, Cameron Glasscock, and David Baker. Atomic context-conditioned protein sequence design using ligandmpnn. Biorxiv, pp. 2023–12, 2023.

W3 - No code made available. Anonymize repo at https://anonymous.4open.science/r/SENZ-2BE1/

Thank you for bringing this to our attention. We have uploaded the code to an Anonymous GitHub repository, which can be accessed using the following link: https://anonymous.4open.science/r/SENZ-2BE1/.

We also highlighted it in the Reproducibility Statement section between the Conclusion and References in the revised manuscript.

W4 - Minor grammatical mistakes. We refined the manuscript.

Thank you for highlighting this issue. We have thoroughly reviewed the entire paper again and corrected several grammatical errors with the assistance of a grammar checker. We appreciate your suggestion.

Thank you for your patience. We will introduce the anonymized code repo.

1. Dataset construction: We updated a more detailed README with dataset construction. The parts of data preprocessing and data preparation for training are also listed here.

   The dataset split is in enzyme_dataset/full_RHEA_finetune_ProSmith/split/split30_ECnumber_fair_val_test0094400943/.

   Retrieved MSAs are in enzyme_dataset_new/MSA/coldDB/.

   Get raw data

   0. cd enzyme_dataset working directory for raw data.
   1. Download the map of index of all enzymes: wget https://ftp.expasy.org/databases/rhea/tsv/rhea2uniprot%5Fsprot.tsv. Get ./rhea2uniprot_sprot.tsv.
   2. Run get_uniport.py to download fasta files for sequences from UniProt in ./data/. In ./data/ there will be fasta files, each contains only one sequecne. The file name is sequence id.
   3. Run rhea-seq_new.ipynb to merge all sequence with reaction id in ./rhea2uniprot_sprot.tsv. Get ./full_seq_from_rhea2uniprot_sprot_tsv.csv.
   4. Get EC raw file (Will be used for baseline only): wget https://ftp.expasy.org/databases/rhea/tsv/rhea2ec.tsv.
   5. Get reaction SMILES: wget https://ftp.expasy.org/databases/rhea/tsv/rhea-reaction-smiles.tsv.
   6. Get substrate of each reaction with ./seq_smiles.ipynb. Get sub_from_rhea-reaction-smiles_tsv.csv.

   Perform data split

   0. mkdir full_RHEA_finetune_ProSmith, cd full_RHEA_finetune_ProSmith. enzyme_dataset/full_RHEA_finetune_ProSmith is the working directory for data split.
   1. Run enzyme_dataset/full_RHEA_finetune_ProSmith/finetune_mmseq.ipynb. Get fintune_dataset.csv. It consists of substrate-enzyme pairs.
   2. Run finetune_mmseq.ipynb to get finetune_all.fasta, which contains all enyzme sequences.
   3. mmseqs easy-cluster finetune_all.fasta full_finetune_mmseq_30/full_finetune_mmseq_30 tmp --min-seq-id 0.3 -c 0.8 --cov-mode 1. Get enzyme_dataset/finetune_mmseq/finetune_mmseq_30/finetune_all_cluster.tsv and enzyme_dataset/finetune_mmseq/finetune_mmseq_30/full_finetune_mmseq_30_all_seqs.fasta. It is the cluster of sequence with identity of 30 maxium.
   4. Run ./full_RHEA_finetune_ProSmith/gen_full_finetune_fair_val_test.ipynb to generate negative samples and split in ./full_RHEA_finetune_ProSmith/split/split30_ECnumber/. (This folder will be renamed later.)

   Retrieval in advance

   1. Run enzyme_dataset_new/makeMSAfinal_groups.py to perform enzyme retrieval for all substrates. Result in enzyme_dataset_new/MSAgroup/coldDB/. One fasta file for every substrate, containing retrieved sequences. enzyme_dataset_new/retrieve.ipynb is an example of the retrieved similar substrates.
   2. Run enzyme_dataset_new/makeMSA.py to get MSAs from retrieved sequences in enzyme_dataset_new/MSA/coldDB.

2. Training and sampling: We also include this part in README. The main function is in these three files.

   1. Guidance function training: mcgp/ProSmith_finetune_DDP_batch_neg.py.
   2. Generator training: mcgp/TarDiff2loss_DDP_strict_step.py.
   3. Sampling: mcgp/TarDiff_test_all_param.py.