Response to Reviewer DL85 (2/2)

Could you provide standard deviations for Mouse and Human for all tools?

Thanks for your helpful reviews! Following your valuable advice, we reproduce the performance of DeepNovo and PointNovo with standard deviations over 5 different random initializations. The results shown in Table Re4 indicate that AdaNovo outperforms previous methods with statistically significant margins.

Table Re4: Empirical comparison of previous models on Mouse and Human datasets in terms of amino acid-level/peptide-level precision (with standard deviations).

The authors indicate limitations in identifying never observed PTMs.

Thank you for your professional reviews! The open search tools such as MSFragger [6] can identify the PTMs the human has discovered, as included in the UniMod database [7]. However, what we mean are the PTMs that have not been found by humans. We will further clarify the distinction in the revised version.

[1] Ad hoc learning of peptide fragmentation from mass spectra enables an interpretable detection of phosphorylated and cross-linked peptides (Altenburg et al., Nature Machine Intelligence)
[2] Large-scale prediction of protein ubiquitination sites using a multimodal deep architecture (He at al., BMC Systems Biology)
[3] De novo peptide sequencing by deep learning (Tran et al., PNAS)
[4] Computationally instrument-resolutionindependent de novo peptide sequencing for high-resolution devices (Qiao et al., Nature Machine Intelligence)
[5] De Novo Mass Spectrometry Peptide Sequencing with a Transformer Model (Yilmaz et al., ICML 2022)
[6] MSFragger: ultrafast and comprehensive peptide identification in mass spectrometry–based proteomics. (Kong et al., Nature Methods)
[7] Unimod: Protein modifications for mass spectrometry. (Creasy et al., Nature Methods)

---

We greatly appreciate your insightful and helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript.

Missing PTMs' the effect on the peptide sequencing overall has to be ignorable. Can you indicate absolute numbers for peptides with/without PTMs in the 9 species benchmark?

Thanks for your insightful and to-the-point reviews! We provide the absolute numbers for peptides with/without PTMs of the 9 species benchmark in Table Re1. As can be observed, peptides with PTMs account for 11.8% to 31.1% across different species. These proportions are relatively large and have significant impacts on the overall performance of peptide sequencing. This is because if there is a single incorrect prediction of a PTM within a peptide, the entire predicted peptide will be erroneous.

Furthermore, we remove peptides with PTMs from the test set and compared the results with those obtained using the model before the removal of PTMs. As shown in Table Re2, the model's peptide-level precision significantly improves after removing the peptides with PTMs in the test set, which verifies that the PTMs have significant impacts on the overall de novo peptide sequencing performance.

Table Re1: The absolute numbers for peptides with/without PTMs in 9 species datasets.

Table Re2: The peptide-level precision of models before and after removing peptides with PTMs in test set.

Can you delineate your approach from existing approaches to identify PTMs and argue why it is beneficial to include this step in de novo sequencing rather than running it separately.

Thanks for your comments! AdaNovo can predict both PTMs' locations and types of the PTMs while previous methods can only predict the location of one specific PTMs type such as phosphorylation [1] or ubiquitination [2]. Therefore, the task studied here is more challenging. If we conduct PTMs identification and de novo sequencing separately, we need to train distinct models to predict various types of PTMs one by one, which is computationally complex and prone to errors. More importantly, as verfied in Table Re2, PTMs have significant impacts on the overall performance of peptide sequencing. Therefore, it is necessary to design novel methods to enhance PTMs identification in de novo peptide sequencing.

The benchmark data set is potentially subject to data leakage. Can you quantify how many of the peptides in the 9 species datasets are actually identical between species or how you protect against data leakage?

Another good point! The nine species datasets are widely used for de novo sequencing methods [3,4,5] and we follow the same experimental settings as them. Specifically, we employ a leave-one-out cross validation framework where the peptides in the training set are almost completely disjoint from the peptides of the held-out species. To illustrate this point, among the ∼26,000 unique peptide labels associated with the human spectra in the test data, only 7% overlap with the ∼250,000 unique peptide labels associated with spectra from the other eight species. The majority (93%) of peptides are not present in the training set, ensuring that the dataset is suitable for evaluating de novo sequencing methods. To further mitigate your concerns, we remove the training samples whose peptide labels overlap with the training set. The results in Table Re3 verify that AdaNovo also exhibits significant advantages on such datasets.

More importantly, in de novo sequencing, we can regard the mass spectra as samples and peptides as labels. The same peptide labels will yield different mass spectra under different experimental conditions. Data leakage occurs when test samples are present in the training set. In our experiments, the training and test spectra come from different experiments, making it impossible for test spectra to be present in the training set. This segregation ensures that there is no data leakage in the benchmark data.

Table Re3: The peptide-level precision on test set without peptides overlap with the training set.

The second part of our response can be found in the next block.

Thanks very much for your professional, insightful, and helpful reviews! We greatly enjoy such in-depth discussions and believe they will significantly enhance the quality of our work.

Q1: I wonder if we have a misunderstanding regarding PTMs here. The motivation of the paper seem to imply that the authors only focus on biological meaningful PTMs ... The frequency numbers the authors describe is significantly above what is commonly described in literature for biological modification...

The motivation of our work lies in that all the PTM types (instead of one specific PTM type) have significant impacts on the overall de novo sequencing performance, which we have explained and experimentally verified in our original manuscript and response above.

Following your previous advice, we provide the ratios/frequency of peptides with PTMs in Table Re1, which can be formulated as, $\frac{\text{The number of the peptides with any PTMs type}}{\text{The number of the peptides}}$. However, the frequency numbers of PTMs you mentioned now can be formulated as $\frac{\text{The number of a specific type of PTMs}}{\text{The number of amino acids}}$. We show the statistics of the frequency of 3 PTM types in Table Re5, from which we can observe that the frequency of PTMs is low (0.051% - 0.914%) in the datasets, just like what is commonly described in literature for biological modification. However, the frequency of peptide with PTMs is relatatively high (see Table Re1), resulting in low peptide-level precision of previous de novo sequencing methods because if a PTM in a peptide sequence is predicted incorrectly, the entire predicted peptide sequence is incorrect in the de novo sequencing task. In other words, the PTMs have significant impacts on the overall peptide-level performance of de novo sequencing even if their frequency is lower than canonical amino acids.

Table Re5: The frequency of 3 PTM types in 9 species datasets.

Q2: Frequency matters a lot for the discussion of training specific models or not and also for the leakage discussion I wonder whether this was considered ...

In some biological applications, we may only care about one specific type of PTMs. For example, in PTMs prediction task, previous methods usually train one model for one specific PTM type. However, to evaluate de novo sequencing models, it is unreasonable to only keep one specific PTM type and remove the other PTM types (such as Ox(M)) in the datasets because all the PTM types have significant impacts on the overall performance of de novo sequencing models, as we have explained and experimentally verified in our manuscript or response above. Additionally, de novo sequencing is primaryly applied to identify the proteomic "darker matters" caused by PTMs and other alterations. It would be better that the de novo sequencing models can identify more PTM types. Therefore, it is reasonable that the widely-used 9 species datasets for evaluating de novo sequencing models contain multiple PTM types. There is NO data leakage in these datasets as we have throughly claried before.

Furthermore, regarding train specific models for each PTM type in PTMs prediction task, some recent works in PTMs prediction also develop for the prediction of multiple types of PTMs. One is CapsNet_PTM which uses a CapsNet for seven types of PTMs prediction [1]. More recently, MusiteDeep has been extended to incorporate the CapsNet with ensemble techniques for the prediction of more types of PTMs [2].

[1] Capsule network for protein post-translational modification site prediction (Wang et al., Bioinformatics)
[2] MusiteDeep: a deep-learning based webserver for protein post-translational modification site prediction and visualization (Wang et al., Nucleic Acids Research)

---

Thank you once again for your thorough, insightful, and constructive reviews! We truly appreciate these detailed discussions and believe they will help resolve any misunderstandings between us. If our responses have addressed your concerns fairly, we respectfully hope you might consider raising your score to support our work. If you have any other questions or concerns, we are very willing to engage in further discussion. Thank you for the time you have dedicated to our work, which will undoubtedly help us improve the quality of our manuscript.

Another good point! Actually, our model AdaNovo is also applicable to the identification of biologically meaningful PTMs, depending on the training dataset used. If the dataset contains various "natural" or biologically meaningful PTMs (such as phosphorylation), AdaNovo can also be used to identify these important biological signals. We conducted some experiments to support our claims. Specifically, we utilized a dataset that encompasses 21 distinct PTMs, referred to as the 21PTMs dataset, as detailed in [1]. For illustrative purposes, we selected data for 5 "natural" or biologically meaningful PTMs—tyrosine phosphorylation, tyrosine nitration, proline hydroxylation, lysine methylation, and arginine methylation—to train and test Casanova and AdaNovo (train:test = 9:1). As shown in Table Re 6, AdaNovo outperforms Casanovo in identifying these ("natural" or biologically meaningful) PTMs, which further lead to AdaNovo’s superiority in overall de novo sequencing performance. We will emphasize this good point in the final version.

Table Re6: The models' performance on the datasets with natural PTMs.

[1] Proteometools: Systematic characterization of 21 post-translational protein modifications by liquid chromatography-tandem mass spectrometry (LC-MS/MS) using synthetic peptides. (Zolg, D. P., et al., Molecular & Cellular Proteomics)

---

We sincerely appreciate the points you have raised. They have helped us identify issues we had not previously considered, particularly regarding AdaNovo’s capability to identify natural or biologically meaningful PTMs. Sincerely hope that our response can address your concerns and further enhance your positive recommendation of our manuscript. If you have any additional questions or concerns, we would be more than happy to engage in a deeper and more thorough discussion. Thank you once again!

Dear Reviewer DL85,

We have provided detailed responses to your reviews. Considering that the deadline for the discussion phase is approaching, we would like to know if our responses have adequately addressed your concerns. If you have any additional concerns or questions, we are more than happy to engage in further discussion. Thank you again for your time and effort in reviewing our manuscript. Your feedback has been instrumental in improving our research:)

Best,
Authors

Dear Reviewer DL85,

We would like to express our sincere gratitude for dedicating your time to reviewing our paper. Your insightful comments on the difference between natural PTMs and accidental PTMs are particularly valuable, inspiring us to apply AdaNovo in the identification of natural or biologically meaningful PTMs.

We have thoroughly considered your feedback and carefully responded to each of your questions with extensive experiments. We would greatly appreciate your feedback on whether our responses have addressed your concerns to your satisfaction.

Once again, we sincerely thank you for your invaluable contribution to our paper. As the deadline is approaching, we eagerly await your post-rebuttal feedback.

Best regards,
Authors.

*Q1: Is "AdaNovo w/o decoder #2 and any reweighting" actually Casanovo? *

Thanks for your helpful reviews! Yes, "AdaNovo w/o decoder #2 and any reweighting" is Casanovo indeed. Therefore, we have performed the ablition study in the original version (Table 2-5).

Q2: Is 72% accuracy on AA level enough in real applications? Maybe some comparisons with database-searching methods help answer this question.

The accuracy of AdaNovo is sufficient for practical applications. In fact, the earliest deep learning model for de novo peptide sequencing, DeepNovo[1], has already been integrated into the commercial software PEAKS. Additionally, what we want to emphasize is that the database search method can only identify proteins already present in the database, whereas de novo peptide sequencing can identify new proteins that are out of the database. Therefore, database searching methods are not suited for our scenarios where the peptide sequences in test set do not exist in the database.

Q3: Section 3.3.1 and 3.3.2 have confusing titles. Consider changing them into amino-acid-level / peptide-level weighting.

Thanks for your valuable advice! We will mitigate this issue in the revised version.

Q4: The normalizations in Eq (5) and (7) seem somewhat arbitrary. Is there any evidence that the setup is optimal?

Thanks for your insightful reviews! The normalization in Eq (5) and (7) is carefully designed. Specifically, we employ the widely-used Z-Score Normalization to standardize variables by transforming them into a standard normal distribution with a mean of 0 and a standard deviation of 1. The cofficients $s_1$ and $s_2$ are designed to control the effects of amino acid-level and peptide-level adaptive training, respectively. Additionally, the inclusion of +1 in the formula typically ensures that the resulting weight is non-negative and that each PSM is fully utlized for training.

Q5: The authors would better articulate the relation between AdaNovo and Casanovo

Casanovo [2] initially employed Transformer encoder-decoder architecture to predict the peptide sequence for the observed spectra. In comparison to Casanovo, AdaNovo's innovation lies in its training strategy specifically tailored for spectrum-peptide matching data to mitigate the data biases (various noise and missing peaks in mass spectra and low-frequency PTMs), rather than the Transformer encoder-decoder architecture. Actually, the de novo peptide sequencing task we have explored, transitioning from mass spectra to amino acid sequences, shares similarities with fields like image captioning [5] (translating images into descriptive texts) and protein inverse folding [6] (deriving amino acid sequences from protein structures), where the encoder-decoder architecture is widely adopted. In these domains, innovation often stems from training strategies rather than the model architectures themselves.

Q6: The studied problem is interesting, while may not be of interest for the major audience in NeurIPS.

Many works [2,3,4] in computational mass spectrum including Casanovo [2] have been published in top-tier machine learning conference including NeurIPS and ICML recently because the tasks in this field can be easily formulated as machine learning problem.

Moreover, the initial phase of drug discovery involves pinpointing disease biomarker proteins or drug target proteins. De novo sequencing serves as a pivotal step in identifying these proteins within the realm of proteomics. Despite remarkable progress in AI for drug design under known targets in top-tier AI conference, the primary bottleneck in the drug discovery pipeline persists in uncovering these crucial protein targets. Our objective is to encourage researchers in the AI community focus more attention on this task, thereby advancing the development of AI-driven drug discovery and development (AIDD).

[1] De novo peptide sequencing by deep learning (Tran et al., PNAS 2017)
[2] De Novo Mass Spectrometry Peptide Sequencing with a Transformer Model (Yilmaz et al., ICML 2022)
[3] Efficiently predicting high resolution mass spectra with graph neural networks (Murphy et al., ICML 2023)
[4] Prefix-tree decoding for predicting mass spectra from molecules (Goldman, Samuel, et al., NeurIPS 2023)
[5] From Show to Tell: A Survey on Deep Learning-Based Image Captioning (Stefanini et al., TPAMI 2023)
[6] ProteinInvBench: Benchmarking Protein Inverse Folding on Diverse Tasks, Models, and Metrics (Gao et al., NeurIPS 2023)

---

We greatly appreciate your insightful and helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research.

Q1: Have the authors tried beam search in AdaNovo?

Thanks for your helpful reviews! Following your valuable advice, we apply the beam search (beam size = 5) in AdaNovo and observe consistent improvements over greedy search in Table Re3 (peptide-level) and Table Re4 (amino acid-level). Thanks again for such helpful advice!

Table Re3: Comparison between greedy search and beam search in terms of peptide-level precision.

Table Re4: Comparison between greedy search and beam search in terms of amino acid-level precision.

Q2: What is the meaning of FDR = 1% in line 176? How to calculate the FDR?

FDR (False Discovery Rate) is a measure used to assess the reliability of peptide or protein identifications in database search. The typical method to calculate FDR involves using a decoy database. Here’s a simplified overview of the process: 1. Database Search: Peptide sequences are identified by matching observed mass spectra to theoretical mass spectra derived from a protein sequence database. 2. Decoy Database: A decoy database is created by reversing or shuffling the protein sequences in the original database. This decoy database contains sequences that mimic the characteristics of the real database but do not correspond to actual proteins. 3. Scoring: Peptide identifications from the real and decoy databases are scored based on parameters such as peptide mass accuracy, retention time, and fragmentation pattern. 4.FDR Calculation: The FDR is then calculated based on the number of accepted identifications from the decoy database compared to the accepted identifications from the real database. This ratio provides an estimate of the proportion of false identifications among all accepted identifications. The FDR calculation can be expressed as:

$$\text{FDR} = \frac{\text{Number of decoy identifications}}{\text{Number of target identifications}}$$

A common approach is to set a threshold for the accepted FDR, such as 1% or 5%, to control the rate of false identifications. This method allows researchers to estimate the proportion of false identifications among their identifications, providing a measure of the reliability of their results.

Q3: Minor issue: The space between the citation and main text is not consistent across the manuscript.

Thanks for your careful reviews! We will mitigate this issue in the revised version.

---

We greatly appreciate your insightful and helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research.

Dear Reviewer #sJoV,

Thank you for your insightful and constructive review! Your feedback has been instrumental in enhancing our research, particularly the beam search method you suggested, which promises to enhance the overall performance of the AdaNovo model.

Best regards,
Authors

Q1: The extra cost of memory compared to Casanovo is significant, which could limit the max length of predicted peptide sequences.

Thanks for your insightful reciews! In mass spectrometry, proteins are enzymatically broken down into peptides for analysis, with peptide lengths typically ranging from 2 to 50 amino acids. This is considerably shorter than full protein sequences, which can contain hundreds or thousands of amino acid residues. Therefore, the memory constraints mentioned might not be a limiting factor for AdaNovo in predicting peptide sequences.

Moreover, we can reduce memory consumption by scaling down Decoder #2. In the rebuttal process, we scale down Decoder #2 to transformer with 3 layers and 6 layers. The experiments shown in Table Re1 indicate a marginal performance drop while significantly reducing memory consumption. Thank you very much for your comments! They have helped us identify issues we hadn't thought of before, significantly improving the efficiency of AdaNovo. We will update the experimental results in the revised version.

Table Re1: The performance of scaling down Decoder #2 from 9 layers to 3 layers on Clam bacteria dataset.

Q2: What are the costs of computing and storage of PointNovo?

Thanks for your valuable advice! We compare the costs of computing and storage of PointNovo, Casanovo and AdaNovo in Table Re2. As can be observed, PointNovo contains less parameters than Casanovo and AdaNovo with similar training and inference speed. However, as shown in Table 1 of the paper, its performance is significantly inferior to Casanovo and AdaNovo.

Table Re2: Costs of Computing and Storage on the same device (Nvidia A100 GPU).

*Q3: What is the detailed number of amino acids in each category in the dataset? In Table 4, why using focal loss would give worse results than cross-entropy loss? *

Thanks for your insightful reviews! We provide the detailed number of amino acids in each category in Table Re2 (Due to the limited sapce, we only show the number of 6 canonical amino acids + 3 PTMs and we will show all the numbers in the revised version). Additionally, focal loss is designed to address moderate class imbalance. However, in our case, canonical amnio acids vastly outnumbers PTMs, focal loss may not be able to effectively handle the imbalance, leading to suboptimal performance. Also, a previous work [3] also report that focal loss is inferior to cross-entropy loss in de novo sequencing. Thanks once again for your valuable advice!

Table Re2: The detailed number of amino acids in each category.

Q4: AdaNovo is built on the Transformer encoder-decoder architecture. Thus I feel the novelty of the proposed model is limited.

AdaNovo's innovation lies in its training strategy (conditional mutual information-based adaptive training) specifically tailored for spectrum-peptide matching data to mitigate the data biases (various noise and missing peaks in mass spectra and low-frequency PTMs), rather than the Transformer encoder-decoder model architecture. The de novo peptide sequencing task we have explored, transitioning from mass spectra to amino acid sequences, shares similarities with fields like image captioning [1] (translating images into descriptive texts) and protein inverse folding [2] (deriving amino acid sequences from protein structures), where the encoder-decoder architecture is widely adopted. In these domains, innovation often stems from training strategies rather than the model architectures themselves [1,2].

[1] From Show to Tell: A Survey on Deep Learning-Based Image Captioning (Stefanini et al., TPAMI 2023)
[2] ProteinInvBench: Benchmarking Protein Inverse Folding on Diverse Tasks, Models, and Metrics (Gao et al., NeurIPS 2023)
[3] De Novo Mass Spectrometry Peptide Sequencing with a Transformer Model (Yilmaz et al., ICML 2022)

---

We greatly appreciate your insightful and helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research.

Dear Reviewer XCqV,

We appreciate your insightful and helpful reviews! Your points on memory consumption are particularly valuable, inspiring us to reduce the size of Peptide Decoder #2 to further enhance AdaNovo's efficiency. If our response has resolved your concerns and clarified any ambiguities, we respectfully hope you might consider raising the score. Should you have further questions or need additional clarification, we would be happy to discuss them. Thank you again for your time and effort in reviewing our manuscript. Your feedback has been instrumental in improving our research.

Best,
Authors

Dear Reviewer XCqV,

We appreciate your insightful and helpful reviews! Your points on memory consumption are particularly valuable, inspiring us to reduce the size of Peptide Decoder #2 to further enhance AdaNovo's efficiency. Considering that the deadline for the discussion phase is approaching, we would like to know if our responses have adequately addressed your concerns. If our response has resolved your concerns and clarified any ambiguities, we respectfully hope you might consider raising the score. Should you have further questions or need additional clarification, we would be happy to discuss them. We truly appreciate your time and effort, which have been crucial in refining our research.

Best regards,
Authors

Q1: For Figure 4, can you supply the dataset size and the performance? How does the amount of PTMs impact the performance of Casanovo and AdaNovo?

Thanks for your helpful comments! We provide the dataset size and performance in Table Re1. We would add these results in Figure 4 of the revised version.

Additionally, following your valuable advice, we remove 50% and 100% peptides with PTMs and report the results in Table Re2, from which we can observe that the performance (peptide-level precision) of Casanovo and AdaNovo significantly improves after removing the peptides with PTMs in the test set. These results indicate that the amounts of PTMs have a significant impact on the performance of de novo sequencing models and AdaNovo is less sensitive to the amounts compared with Casanovo.

Table Re1: The absolute numbers for peptides with/without PTMs in 9 species datasets.

Table Re2: The peptide-level precision of Casanovo and AdaNovo before and after removing peptides with PTMs in test set.

Q2: This is a narrow field of proteomics, there are other problems that could be considered to truly showcase the power of the proposed methodology.

The initial phase of drug discovery involves pinpointing disease biomarker proteins or drug target proteins. De novo sequencing serves as a pivotal step in identifying these proteins within the realm of proteomics. Despite remarkable progress in AI for drug design under known targets, the primary bottleneck in the drug discovery pipeline persists in uncovering these crucial protein targets. Our objective is to encourage researchers in the AI community focus more attention this task, thereby advancing the development of AI-driven drug discovery and development (AIDD).

Moreover, many works [1,2,3] in computational mass spectra including Casanovo [2] have been published in top-tier machine learning conference including NeurIPS and ICML because the tasks in this field can be easily formulated as machine learning problem.

[1] Prefix-tree decoding for predicting mass spectra from molecules (Goldman, Samuel, et al., NeurIPS 2023)
[2] De Novo Mass Spectrometry Peptide Sequencing with a Transformer Model (Yilmaz et al., ICML 2022)
[3] Efficiently predicting high resolution mass spectra with graph neural networks (Murphy et al., ICML 2023)

---

We greatly appreciate your insightful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified any ambiguities, we respectfully hope that you consider raising the score/confidence. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research.

Dear Reviewer xCSd,

Thank you for your concise and insightful feedback! Based on your valuable advice, we have progressively removed peptides with PTMs in the test dataset, ranging from 0% to 100% (kindly note that the earlier range of 5% - 10% might not accurately reflect the trend). The updated results, available at this anonymous link (https://anonymous.4open.science/r/Remove_PTM/Figure_Re1.png), show that the peptide-level precision of both Casanovo and AdaNovo improves significantly as the proportion of removed peptides with PTMs is increased. This demonstrates that the presence of PTMs notably affects the performance of de novo sequencing models, with AdaNovo showing less sensitivity compared to Casanovo.

If our response addresses your concerns and clarifies any uncertainties, we would greatly appreciate your consideration in adjusting the score or confidence. Should you have any further questions or require additional information, please do not hesitate to reach out. Thank you once again for your time and constructive review, which has been invaluable in enhancing our research!

Best regards,
The Authors

Q1: In eq.3 use MI (X , Z; Yj | Y<j ) but in model it is CMI(X , Z; Yj | Y<j ), and should not distinguish between italic rv and bold rv.

Thanks for your careful review! Kindly note that the conditioned mutual information in model is formulated as CMI(X , Z; Yj) or MI (X , Z; Yj | Y<j ) rather than CMI(X , Z; Yj | Y<j ). In eq.3, we can rewrite CMI(X , Z; Yj) as MI (X , Z; Yj | Y<j ) because condition mutual information is defined as the mutual information between two random variables (X , Z) and Yj conditioned on a third Y<j. In other words, CMI is a kind of mutual information. Here, we use the well-known concept of mutual information to help readers understand the calculation process of conditional mutual information.

Additionally, the bold rv denotes specific values while the italic rv denotes the variables. It is necessary to distinguish between values (bold rv) and variables (italic rv) because the conditional mutual information is defined as a measure of the mutual dependence between two variables instead of specific values.

Q2: CMI(X , Z; Yj ) = MI (X , Z; Yj | Y<j )? Please provide a more detailed process.

Yes, CMI(X , Z; Yj ) = MI (X , Z; Yj | Y<j ). Here, CMI(X , Z; Yj) denotes the conditional mutual information between (X, Z) and Yj (conditioned on Y<j). This step can be derived using the definition of Conditional Mutual Information, which is the mutual information of two random variables (X , Z) and Yj given the value of a third Y<j. In other words, CMI is a kind of mutual information. Here, we use the well-known concept of mutual information to help readers understand the calculation process of conditional mutual information.

Q3: The paper does not provide comparisons with other methods that utilize strategies for handling repetitive long-tail data.

We have compared with such methods in section 4.6 of the original paper.

Q4: There is a lack of verification whether the Transformer's predictions align with the model's hypothesized outputs. Additionally, although CMI can technically be calculated, the inputs used in your model do not conform well to standard data formats.

We have elborated on how we feed the mass spectrum to the transformer in Appendix A. More specifically, we regard each mass spectrum peak (m_i, I_i) as a word/token in natural language processing and obtain its embedding by individually encoding its m/z value m_i and intensity value I_i before combining them through summation. The entire mass spectrum can be regarded as a sentence. The precursor can also be encoded using similar way. More details can be found in Appendix A.

Additionally, we feed mass spectrum ($\mathbf{x}$), precuosor ($\mathbf{z}$) and previous sequence ( $\mathbf{y_{<j}}$ ) to the MS Encoder and Peptide Decoder #1 to predict the next amino acid ($y_j$). Therefore, the output of Peptide Decoder #1 is $p(y_{j}|\mathbf{x}, \mathbf{z}, \mathbf{y_{<j}})$. Similarly, Peptide Decoder #2 takes previous sequence ($\mathbf{y_{<j}}$) as input and outputs $p(y_{j}|\mathbf{y_{<j}})$. And then, we can use $p(y_{j}|\mathbf{x}, \mathbf{z}, \mathbf{y_{<j}})$ and $p(y_{j}|\mathbf{y_{< j}})$ to calculate the conditional mutual information using Eq. (4).

---

We greatly appreciate your helpful comments, as they will undoubtedly help us improve the quality of our article. If our response has successfully addressed your concerns and clarified the ambiguities, we respectfully hope that you consider raising the score. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research.

Dear Reviewer XeYm,

Thanks for your prompt and insightful response! We agree that the misunderstanding between us lies in the notation of the conditional mutual information. Following your valuable advice, we would update $CMI(X, Z; Y_{j})$ as $I(X,Z; Y_{j}|Y_{<j})$ in Eq. (3) and Eq. (4) to avoid misunderstandings.

Furthermore, we would like to inquire if our response adequately addresses your concerns. Should you have any further questions or require additional clarification, we would be delighted to engage in further discussion. Once again, we sincerely appreciate your time and effort in reviewing our manuscript. Your feedback has been invaluable in improving our research!

Best,
Authors

It is recommended that you write your formula according to the format provided in https://www.jmlr.org/papers/volume24/21-0482/21-0482.pdf.

Regarding the misunderstanding between us, you directly denote $CMI(X, Z; Y_j)$ as $I(X, Z, Yj | Y_{<j})$. However, CMI involves the mutual information between two variables given a third variable or set of variables.This could potentially lead to a misinterpretation as $I(X, Z | Y_j)$, particularly given that $X, Y_j, Z$ are the only variables under consideration.

In terms of formula notation, there remains significant room for enhancement. Therefore, the original score will be retained, with the expectation that the author will make further refinements.

Dear Reviewer XeYm,

Thanks very much for your feedback on the formula notation! Your suggestions help standardize the mathematical notations in our manuscript and avoid potential misunderstandings. Following your valuable advice, we have updated the notations and uploaded the revised manuscript to the anonymous link: https://anonymous.4open.science/r/NeurIPS24_AdaNovo_Revision-7C65/AdaNovo_NeurIPS2024_Revision.pdf.

Specifically, we replace $CMI(X, Z; Y)$ and $MI(X, Z; Y)$ with $I(X, Z; Y|Y_{<j})$ and $I(X, Z; Y)$, respectively. Also, $\mathcal{X}$, $\mathcal{Y}$, $\mathcal{Z}$, $\mathcal{Y_{<j}}$ are replaced with $X, Y, Z, Y_{<j}$.

These changes do not require substantial modifications to the original manuscript. Sincerely hope that they would resolve your concerns regarding the formula notations. Please let us know if our response has adequately addressed your concerns or if further clarification is needed. We appreciate your time and invaluable feedback in improving our research!

Best,
Authors

Dear Reviewer XeYm,

We have provided detailed responses to your reviews. Considering that the deadline for the discussion phase is approaching, we would like to know if our responses have adequately addressed your concerns. If you have any additional concerns or questions, we are more than happy to engage in further discussion. Thank you again for your time and effort in reviewing our manuscript. Your feedback has been instrumental in improving our research!

Best,
Authors

Dear Reviewer XeYm,

We would like to express our sincere gratitude for dedicating your time to reviewing our paper. Your suggestions help standardize the mathematical notations in our manuscript and avoid potential misunderstandings.

We have thoroughly considered your feedback and carefully responded to each of your questions. We would greatly appreciate your feedback on whether our responses have addressed your concerns to your satisfaction.

Once again, we sincerely thank you for your invaluable contribution to our paper. As the deadline is approaching, we eagerly await your post-rebuttal feedback.

Best regards,
Authors.