A deep learning method for predicting the minimum inhibitory concentration of antimicrobial peptides against Escherichia coli using Multi-Branch-CNN and Attention

ABSTRACT Antimicrobial peptides (AMPs) are a promising alternative to antibiotics to combat drug resistance in pathogenic bacteria. However, the development of AMPs with high potency and specificity remains a challenge, and new tools to evaluate antimicrobial activity are needed to accelerate the discovery process. Therefore, we proposed MBC-Attention, a combination of a multi-branch convolution neural network architecture and attention mechanisms to predict the experimental minimum inhibitory concentration of peptides against Escherichia coli. The optimal MBC-Attention model achieved an average Pearson correlation coefficient (PCC) of 0.775 and a root mean squared error (RMSE) of 0.533 (log μM) in three independent tests of randomly drawn sequences from the data set. This results in a 5–12% improvement in PCC and a 6–13% improvement in RMSE compared to 17 traditional machine learning models and 2 optimally tuned models using random forest and support vector machine. Ablation studies confirmed that the two proposed attention mechanisms, global attention and local attention, contributed largely to performance improvement. IMPORTANCE Antimicrobial peptides (AMPs) are potential candidates for replacing conventional antibiotics to combat drug resistance in pathogenic bacteria. Therefore, it is necessary to evaluate the antimicrobial activity of AMPs quantitatively. However, wet-lab experiments are labor-intensive and time-consuming. To accelerate the evaluation process, we develop a deep learning method called MBC-Attention to regress the experimental minimum inhibitory concentration of AMPs against Escherichia coli. The proposed model outperforms traditional machine learning methods. Data, scripts to reproduce experiments, and the final production models are available on GitHub.

Thank you for submitting your manuscript to mSystems.We have completed our review and I am pleased to inform you that, in principle, we expect to accept it for publication in mSystems.However, acceptance will not be final until you have adequately addressed the reviewer comments.
Thank you for the privilege of reviewing your work.Below you will find instructions from the mSystems editorial office and comments generated during the review.

Preparing Revision Guidelines
To submit your modified manuscript, log onto the eJP submission site at https://msystems.msubmit.net/cgi-bin/main.plex.Go to Author Tasks and click the appropriate manuscript title to begin the revision process.The information that you entered when you first submitted the paper will be displayed.Please update the information as necessary.Here are a few examples of required updates that authors must address: • Point-by-point responses to the issues raised by the reviewers in a file named "Response to Reviewers," NOT IN YOUR COVER LETTER.
• Upload a compare copy of the manuscript (without figures) as a "Marked-Up Manuscript" file.
• Each figure must be uploaded as a separate file, and any multipanel figures must be assembled into one file.• Manuscript: A .DOC version of the revised manuscript • Figures: Editable, high-resolution, individual figure files are required at revision, TIFF or EPS files are preferred ASM policy requires that data be available to the public upon online posting of the article, so please verify all links to sequence records, if present, and make sure that each number retrieves the full record of the data.If a new accession number is not linked or a link is broken, provide production staff with the correct URL for the record.If the accession numbers for new data are not publicly accessible before the expected online posting of the article, publication of your article may be delayed; please contact the ASM production staff immediately with the expected release date.
For complete guidelines on revision requirements, please see the journal Submission and Review Process requirements at https://journals.asm.org/journal/mSystems/submission-review-process.Submission of a paper that does not conform to mSystems guidelines will delay acceptance of your manuscript.
Please return the manuscript within 60 days; if you cannot complete the modification within this time period, please contact me.If you do not wish to modify the manuscript and prefer to submit it to another journal, please notify me of your decision immediately so that the manuscript may be formally withdrawn from consideration by mSystems.
Corresponding authors may join or renew ASM membership to obtain discounts on publication fees.Need to upgrade your membership level?Please contact Customer Service at Service@asmusa.org.
Thank you for submitting your paper to mSystems.
The ASM Journals program strives for constant improvement in our submission and publication process.Please tell us how we can improve your experience by taking this quick Author Survey.

Reviewer comments:
Reviewer #1 (Comments for the Author): Jielu Yan et al. conducted a study on the use of deep learning techniques to predict the minimum inhibitory concentration (MIC) of antimicrobial peptides against Escherichia coli.Overall, the paper presents interesting findings and has significant scientific merit.However, there are several areas that require modification before publication.

Major suggestions:
(1) The authors should provide a brief explanation of the hyperparameter selection process for the machine learning method used, including reference to relevant literature.For example, the authors selected an epoch of 200 and a learning rate of 0.0005 (compared to a common learning rate of 0.001).It would be helpful to know the reasons behind these selections and the primary reference(s) used.
(2) Considering that the study was conducted with a relatively small dataset of 7,861 data entries, it is necessary to verify whether this sample size is adequate for training the model.The authors should provide additional references to support the reliability of their analysis.
(3) Given the many factors that influence MIC values for E. coli in biological experiments, such as growth media, temperature, and humidity, the authors should exercise caution in their introduction and discussion of the study's findings.

Minor suggestions:
(1) The authors should provide a brief introduction to the average Pearson correlation coefficient (PCC) in line 107 to aid readers who may not be familiar with this statistical measure.PCC reflects the degree of correlation between two datasets, with a value of -1 indicating a totally negative correlation, 0 indicating no correlation, and 1 indicating a completely positive correlation.
(2) There may be a typo in Table 5, where "PAAC pseudo-Amino Acid Composition" and "SOCNumber Sequence-Order-Coupling Number" are listed.
(3) The font size in Figures 1 and 3 (lines 138 and 191) may be too small, and the resolution could be improved to facilitate reading.(4) The capitalization of words in Figure 6 (line 419) should be consistent across all images.
Overall, this is a well-written paper, and with the recommended modifications, it would be suitable for publication.
Reviewer #2 (Comments for the Author): The paper "A Deep Learning Method for Predicting the Minimum Inhibitory Concentration of Antimicrobial Peptides against Escherichia coli using Multi-Branch-CNN and Attention" is a natural extension of the authors' previous work.In this paper, they develop a machine learning model that uses deep learning to predict the concentration of antimicrobial peptides meaning that they developed a regression model instead of a binary classifier.
The analysis presented in the paper is robust.The authors show the feature selection process, and they compare the performance metrics of several regression models, including the one they propose.The final comparison tests several modalities of the proposed architecture, showing its versatility and how the user can fine-tune the model to a specialized scope.
The authors show that their proposed model is superior to traditional machine learning algorithms by a small margin.However, it performs better overall.This is a solid research paper that includes the corresponding checks for a machine learning project.

Type of manuscript: Research Article
Title: A Deep Learning Method for Predicting the Minimum Inhibitory Concentration of Antimicrobial Peptides against Escherichia coli using Multi-Branch-CNN and Attention We sincerely thank the Associate Editor and the anonymous reviewers for their detailed, insightful and valuable comments.We have carefully considered the reviewers' comments and revised the paper accordingly.In the following, please find our responses to each reviewer's comments and a summary of the revisions.For ease of reading, we included the original review comments in a box, followed by our responses.

Please note that:
"main-marked.pdf": the marked-up version of the revised manuscript, in which the deleted text has been shown with a strikethrough and the added text with red color.

Jielu Yan et al. conducted a study on the use of deep learning techniques to predict the minimum inhibitory concentration (MIC) of antimicrobial peptides against Escherichia coli.
Overall, the paper presents interesting findings and has significant scientific merit.However, there are several areas that require modification before publication.

Major suggestions:
AQ:1 = The authors should provide a brief explanation of the hyperparameter selection process for the machine learning method used, including reference to relevant literature.For example, the authors selected an epoch of 200 and a learning rate of 0.0005 (compared to a common learning rate of 0.001).It would be helpful to know the reasons behind these selections and the primary reference(s) used.
Thanks for the suggestion.We have presented the hyperparameter selection procedure in the section "Hyperparameter tuning of MBC-Attention" as shown below.
Different data sets use different parameters, and the final choice of hyperparameters is often based on the results of parameter optimization.There are many works dealing with different strategies for selecting the optimal parameters [1,2].Overall, the final learning rate is determined based on the validation performance in the training datasets, model architecture, and combination with other hyperparameters.To our understanding and observation, it is not always 0.001.
Moreover, we first stopped at 200 epochs to empirically select the CNN layer, the number of filters, the dropout rate, and the loss function.After we decided on the CNN layer, number of filters, dropout rate, and loss function, we used an early stop function to terminate the training phase and maintain a stable performance.We manually set the parameters of the early stop function and finally determined that the early stop would be monitored with "loss", with a learning rate of 0.0005, a patience of 15, a decay rate of 0.92, and decay steps of 25 as our final selection.
We are aware that there are many different algorithms for hyperparameter optimization [3].However, since the MBC-Attention model is relatively small, in our case, the grid search is sufficient for this purpose.

AQ:2 = Considering that the study was conducted with a relatively small dataset of 7,861 data entries, it is necessary to verify whether this sample size is adequate for training the model. The authors should provide additional references to support the reliability of their analysis.
Thanks for the critical suggestion.Just a note on the numbers, the total number of sequences downloaded is 7,861, where the number of sequences finally used after preprocessing is 3,929, with 3,536 sequences forming the training set and 393 sequences forming the test set.Although the trend is clear, the optimal data size for this AMP prediction task cannot be determined due to unavailability of an even larger data set.
The figure and the analysis are added to the manuscript, page 7-8, as shown below: This analysis result was restated in the discussion part: AQ:3 = Given the many factors that influence MIC values for E. coli in biological experiments, such as growth media, temperature, and humidity, the authors should exercise caution in their introduction and discussion of the study's findings.
Thank you for this suggestion.We are aware of the different MIC values that can be obtained for the same AMP under different experimental conditions.However, since the dataset is not comprehensive enough to train specific (or condition-dependent) DL models, our approach in this work is to predict the overall antimicrobial potential of an AMP by correlating the sequence with the average MIC from multiple experimental entries.
We improved the manuscript with this idea in the Introduction (page 3) and Method (page 9).We believe that learning the correlation between different experimental conditions and the experimental MIC measure will provide deeper insight on the AMP activity, that should be further investigated in future works (see page 9).Thank you for the quick and comprehensive revision of your manuscript.
Your manuscript has been accepted, and I am forwarding it to the ASM Journals Department for publication.For your reference, ASM Journals' address is given below.Before it can be scheduled for publication, your manuscript will be checked by the mSystems production staff to make sure that all elements meet the technical requirements for publication.They will contact you if anything needs to be revised before copyediting and production can begin.Otherwise, you will be notified when your proofs are ready to be viewed.
ASM policy requires that data be available to the public upon online posting of the article, so please verify all links to sequence records, if present, and make sure that each number retrieves the full record of the data.If a new accession number is not linked or a link is broken, provide production staff with the correct URL for the record.If the accession numbers for new data are not publicly accessible before the expected online posting of the article, publication of your article may be delayed; please contact the ASM production staff immediately with the expected release date.
As an open-access publication, mSystems receives no financial support from paid subscriptions and depends on authors' prompt payment of publication fees as soon as their articles are accepted.

Publication Fees:
We have partnered with Copyright Clearance Center to collect author charges.You will soon receive a message from no-reply@copyright.com with further instructions.For questions related to paying charges through RightsLink, please contact Copyright Clearance Center by email at ASM_Support@copyright.com or toll free at +1.877.622.5543.Hours of operation: 24 hours per day, 7 days per week.Copyright Clearance Center makes every attempt to respond to all emails within 24 hours.For a complete list of Publication Fees, including supplemental material costs, please visit our website.
Corresponding authors may join or renew ASM membership to obtain discounts on publication fees.Need to upgrade your membership level?Please contact Customer Service at Service@asmusa.org.
If you would like to submit a potential Featured Image, please email a file and a short legend to msystems@asmusa.org.Please note that we can only consider images that (i) the authors created or own and (ii) have not been previously published.By submitting, you agree that the image can be used under the same terms as the published article.File requirements: square dimensions (4" x 4"), 300 dpi resolution, RGB colorspace, TIF file format.
For mSystems research articles, you are welcome to submit a short author video for your recently accepted paper.Videos are normally 1 minute long and are a great opportunity for junior authors to get greater exposure.Importantly, this video will not hold up the publication of your paper, and you can submit it at any time.
Details of the video are: • Minimum resolution of 1280 x 720 • .movor .mp4.video format • Provide video in the highest quality possible, but do not exceed 1080p • Provide a still/profile picture that is 640 (w) x 720 (h) max • Provide the script that was used We recognize that the video files can become quite large, and so to avoid quality loss ASM suggests sending the video file via https://www.wetransfer.com/.When you have a final version of the video and the still ready to share, please send it to mSystems staff at msystems@asmusa.org.
W. I. Siu University of Saint Joseph Institute of Science and Environment Macau China Re: mSystems00345-23 (A Deep Learning Method for Predicting the Minimum Inhibitory Concentration of Antimicrobial Peptides against Escherichia coli using Multi-Branch-CNN and Attention) Dear Prof. Shirley W. I. Siu: To examine if our sample size is adequate to train the proposed MBC-Attention model, we conducted four experiments trained with randomly selected 1000, 2000, and 3000 sequences from the training set and the full training set (3536 sequences).Each draw-and-run was repeated 5 times, all trained models were tested with the same test set and we measured their performance by PCC.As shown in Figure4below, the resulting boxplot of the PCC values exhibit a clear trend of improvement, indicating that the model is able to learn from more training data and make predictions more accurately.The result also shows a clear trend toward convergence, suggesting that the current training data size is a reasonable amount to train the MBC-Attention model.