Prediction of Linear Cationic Antimicrobial Peptides Active against Gram-Negative and Gram-Positive Bacteria Based on Machine Learning Models

Round 1

Reviewer 1 Report

This is an interesting manuscript where authors present a machine learning methodology for evaluation of antibiotic activity of linear AMPs (antimicrobial peptides). Although, there are number of papers describing computation approach to the AMPs analysis, to our knowledge the classification/clustering machine learning for such group of peptides was not reported before.

Major and Minor comments:

- The list of keywords is rather short, expand it. And actually ‘random forest’ keyword is not the best here.

- Describe abbreviations at their first occurrence for instance random forest (RF), Support Vector Machine (SVM), k-Nearest Neighbor (kNN) algorithms. Check others.

- In introduction, in relation to ref 35-36, authors of this manuscript can explain results along with the method. Consider also these citations: 10.1038/s41598-018-19752-w, 10.1186/s12864-018-5030-1.

- Figure 1 can include more detains of the workflow.

- What was the reason behind the size selection of the peptide (20-50 aa).

- P3, L145, the sentence ‘In other words…’ is not necessary.

- P5, L174, redundancy, change ‘...against Gram-negative bacteria; and 1 implies that the peptide is active against these bacteria.

- The authors can consider performing balance accuracy to measure the performance of the model studied.

- Several research papers have reported higher accuracy (>90%) for different dataset and different models in identifying AMPs. How do you justify your novelty and differentiate from other findings?

- The PCA result from section 3.2 can be combined with section 3.3.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Burcu et al reported a ML application on searching the antimicrobial peptides. The paper is well written. Just have one concern. Could the author pick few predicted peptide for experimental validation? If so, the experimental validation work would make the paper be better.  

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This work, “Prediction of Linear Cationic Antimicrobial Peptides Active Against Gram Negative and Positive Bacteria based on Machine Learning Models” (applsci-1565709) presented by Burcu Bakır-Güngör et al., provides an interesting study in developing two models for the prediction of antimicrobial peptides. With the help of various means including machine learning, the authors provide convincing models for the development of antimicrobial peptides, which present a new clue for the development of AMPs. Before being accepted, one concern should be addressed.

1. Figure 1---Specific numbers should be noted at each step.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The analyzes presented in the manuscript are based on the following data: antimicrobial activity of peptides obtained from the DBAASP database, and some physicochemical data of these peptides estimated using algorithms also offered in the DBAASP database. The results of the analyzes carried out by the authors indicate that the main factors describing the activity of the analyzed peptides are the formal peptide charge and the isoelectric point. The result obtained by the authors is trivial and easily predictable. As the authors themselves admit, "In literature, both net charge and the isoelecetric point of a peptide are known to have a considerable effect in terms of determining the activity of AMPs" (lines 459-461). Antimicrobial peptides have to posse’s net charge and keep this charge over a wide pH range which is described by the isoelectric point, so it was to be expected that these would be the main descriptors.

On the other hand, it should be remembered that the algorithms intended to estimate the macroscopic features of peptides very poorly or not at all take into account the differences in the calculated parameters related to changes in the order of amino acid residues in the peptide sequences. That is, the enumerated characteristics of the peptides are dependent on the amino acid composition, rather than the order of the amino acids in the sequence. The algorithm used by the authors calculates the following peptide properties: Normalized Hydrophobic Moment; Normalized Hydrophobicity; Net Charge: Isoelectric Point: Penetration Depth: Tilt Angle: Disordered Conformation Propensity: Linear Moment: Propensity to in vitro Aggregation: Angle Subtended by the Hydrophobic Residues: Amphiphilicity Index: Propensity to PPII coil.

Parameter calculations performed for two sequences, the numerical values correspond to the parameters listed above.

GWASKIGTQLGKMAKVGLKEFVQ

1.16 -0.28 3.00 10.70 15 76 0.09 0.16 1.96 60.00 1.10 0.86

1.16 -0.28 3.00 10.68 15 104 0.09 0.16 0.00 60.00 1.10 0.86

QVFEKLGVKAMKGLQTGIKSAWG

It is shown above that for two peptides composed of the same amino acids only ascending in reverse order, only two of the twelve parameters calculated differ. The peptide selection procedure used by the authors would recognize these two peptides as completely different sequences, but their biological activity will most likely be completely different, and the macroscopic parameters will be very similar. The authors chose a list of potential descriptors (the method of their determination), which descriptors or the method of their determination is not able to pick the differences in structure / features of individual peptides which correlate with biological activity. Moreover, the authors did not check whether there is a correlation (relationship) between the features / descriptors (this is a standard procedure used in this type of work, see e.g. ref. 58). Therefore, in my opinion, the results presented are practically worthless and the work is not suitable for publication .

Minor points:

It is quite difficult to find a clearly defined research goal in the text. Only in the section Disscusion (lines 367-368) you can find out what was the aim of the work "In this study, we attempted to develop a robust classification model for 367 antimicrobial peptide prediction problem"

Worryingly, the authors probably do not fully understand what the individual parameters / descriptors are. Lines 417-418 "The pI feature refers to the solubility of the peptides under certain pH conditions." The isolectic point has something to do with solubility, but by definition, it describes the pH value at which a molecule has an effective electric charge of 0.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf