Knowledge Mining of Interactions between Drugs from the Extensive Literature with a Novel Graph-Convolutional-Network-Based Method

Round 1

Reviewer 1 Report

From my point of view, the article is in general well written. The authors propose a novel graph convolutional network-based method for knowledge mining interactions between drugs from extensive literature.

The model called by the authors as DDINN model achieves higher performance results compared to other similar prediction methods.  The authors have demonstrated thru extensive experiments that the model can predict the drug-drug interaction with the best F-score.  

On the other hand, the feature of this novel deep learning model method, that can effectively use the contextual and syntactic information of input literature text at the same time, as well as the proposal of a method to rebalance weight of all edges in the dependency matrix for Graph Convolutional Network, are significant enough contributions in my opinion.   

It must be recognized that the authors carry out a deep and exhaustive bibliographical review. On the other hand, I consider that the References are relevant and recent, as well as of good quality. In addition, there are no inappropriate or unnecessary self-citations or self-references observed. But I do consider that it should be checked that they have complete the information.

I think there are some typographical errors in some parts of the text. For example, on page 5, line 162, the authors may have meant forward. On page 6, at line 211, the authors probably meant pairs. On the other hand, they mention in line 222 of page 6, section 2.2.2. However, I have not been able to locate that section in the article. Also in line 330, of page 10, they probably wanted to say training. A general revision of the manuscript must be carried out

I consider that it is an article with adequate quality to be published after the proposed corrections

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors studied the problem of detecting drug-drug interactions (DDIs). They define a novel graph convolutional network-based method called DDINN to detect potential DDIs, by combining cBiLSTM, graph convolutional networks, and a weight-reblanced dependency matrix.

 

Although the proposal seems to be focused on the specific DDI detection problem (which represents a very interesting problem), the paper does not properly highlight the advancement of the state of the art. 

In particular: 

1) the novelty of the proposal should be better highlighted; 

2) related works should be introduced;

3) the presentation should be improved, by providing real-world examples and ideas underlying design choices;

4) No proper discussion over experimental results has been provided.

 

More specifically,

- Although the authors introduce the research context, they do not provide any example of the kind of predictions that should be considered. Thus, some real-world examples can be provided in section I and/or section 2.1. 

- In section 2, they also introduce a wide reference literature addressing the DDI problem, but no related work section has been introduced.  

- Starting from section 2.2, the authors only describe the architecture and their components. Nevertheless, some characteristics of the considered context remain obscure. For instance, which documents do they refer to? Moreover, design choices should be described more in-depth. 

- Using just one dataset for the experiments does not permit the evaluation of the real effectiveness of the proposed work. Moreover, no detail concerning the dataset has been provided, which kind of drugs it contains?   

- It is strange to have many compared models that have never been described. This further emphasizes the necessity to have a related work section.

- In general, the contents are too superficial. They miss details that are important for a reader to understand why this proposal is better than others. In fact, not only experimental results are important. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have described a graph CNN based model drug-drug interactions (DDIs). The paper is well written and the flow of the manuscript is good. A few of the below questions need clarification:

1. Has there been a class imbalance in the dataset selected? If so, how it is handled has to be explained.

2. In healthcare domain, recall is an important measure of how accurately our model is able to identify the positive data. Typically it is expected to have a high recall value in healthcare domain. But the best recall value is 0.761. Please justify this point. 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

In this revised version, the paper has been improved substantially, and the authors solved most of the remarks. They introduced examples and better described the related literature through a dedicated section. 

Some other suggestions have been provided in the following:

- At the end of section 1, it would be useful to describe the contents in the remainder of the paper.

- In Section 2, Please introduce a final subsection in which your approach is compared to the ones described above.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf