Deep Learning Architecture Optimization with Metaheuristic Algorithms for Predicting BRCA1/BRCA2 Pathogenicity NGS Analysis
Round 1
Reviewer 1 Report
- Overall a good and well-written manuscript.
- For "2.3. Data selection" any Inclusion and exclusion criteria. Please explain
- In the conclusions section, the author should explain what next and suggestions for future use especially the need for wet laboratory assays.
Author Response
We have been able to incorporate changes to reflect most of the suggestions provided by the reviewers. I have highlighted the changes within the manuscript.
Here is a point-by-point response to the reviewers’ comments and concerns.
1. Overall a good and well-written manuscript.
We are grateful to the reviewer for his insightful comments on our paper.
2. For "2.3. Data selection" any Inclusion and exclusion criteria. Please explain
Thank you for pointing this out. We have actually described inclusion and exclusion criteria in the 2.5 Data encoding (Feature construction) subsection. We decided to pick values that the biologist used to interpret mutation variants which are available in Table 1. We have decided to exclude data that contain multiple values mixing numbers and unusable characters.
3. In the conclusions section, the author should explain what next and suggestions for future use especially the need for wet laboratory assays.
We think this is an excellent suggestion, consequently, we added in the conclusion section:
… They can easily be implemented in other molecular biology laboratories and become a true aid in routine NGS analysis interpretation. After implementation in another laboratory, performance will still need to be evaluated to see if the model generalizes to the new conditions, which may involve, for example, different protocols and samples of different characteristics. The article focused on the optimization of a fully connected deep neural network with metaheuristic algorithms. There are various types of deep learning. For example, the recurrent neural network (RNN), the convolutional neural network (CNN), long short term memory (LSTM), deep belief networks (DBN)... It would be interesting to optimize these other types of neural networks with PSO and GA algorithms for other applications in the medical field.
Author Response File: Author Response.pdf
Reviewer 2 Report
The method and results are sound. Please improve on the quality of figures presented on the clarity of the information e.g. Figure (..the label are too small and not clear
Author Response
We would like to thank reviewer for their comments. I have highlighted the changes within the manuscript. Here are our point-to-point response.
The method and results are sound. Please improve on the quality of figures presented on the clarity of the information e.g. Figure (..the label are too small and not clear
We agree with this and we improved the resolution quality of figures into the manuscript.
Author Response File: Author Response.pdf
Reviewer 3 Report
1. For each first occurrence of abbreviation, please make sure to have the full name. Line 94, 96......
2. There are various types of deep learning, fully connected neural networks, RNN, CNN, etc. From this manuscript, the optimization of neural networks should be limited to fully connected neural networks, the author should explain what type of neural network is optimized in this article.
3. The first step in creating a neural network is the selection and processing of the data to use? I suggest deleting it.
4. The description of Eq (1) is not suitable, R should have a shape to reflect the input data structure.
Author Response
We would like to thank reviewer for their comments. Here are our point-to-point response.
1. For each first occurrence of abbreviation, please make sure to have the full name. Line 94, 96...…
Abbreviations which need to have full names were modified into the manuscript MySQL (My Structured Query Language), BWA-MEM (Burrows-Wheeler Aligner), VEP Ensembl (Variant Effect Predictor), Polyphen (Polymorphism Phenotyping). The other names are not abbreviations, but software names.
2. There are various types of deep learning, fully connected neural networks, RNN, CNN, etc. From this manuscript, the optimization of neural networks should be limited to fully connected neural networks, the author should explain what type of neural network is optimized in this article.
We agree with the reviewer’s assessment. Accordingly, throughout the manuscript, we specified that we focused to tune fully connected neural networks (Multi Layer Perceptron – MLP) with metaheuristic algorithms.
3. The first step in creating a neural network is the selection and processing of the data to use? I suggest deleting it.
This sentence was removed from the manuscript.
4. The description of Eq (1) is not suitable, R should have a shape to reflect the input data structure.
Thank you for pointing this out. The domain should describe the number of predictors and the correct form of the Eq (1) should be:
f: R^15 -> {Pathogenic, Benign) (1)
Author Response File: Author Response.pdf