Modeling DTA by Combining Multiple-Instance Learning with a Private-Public Mechanism

Round 1

Reviewer 1 Report

In this manuscript, Wang et al. introduced a new method to predict the drug-target affinity using advanced machine-learning strategies. This is a quite interesting piece of work that can be potentially useful for the researchers working in the drug design and discovery field, showing theoretical efforts in addressing the challenging problems in this field. First of all, I would like to point out that I am not an expert on algorithm development, so the evaluation of the details of the machine-learning algorithm was not performed. I suggest the editor rely more on the reports from other referees. Overall, the paper reads clearly and the conclusion appears to be based on their results. I don’t have comments.

 

Only one thing needs to be noticed. There is an abbreviation “DTA”, which is not generally used. To me, it also takes me a while to identify what the phrase represents. Please consider to revise.

Author Response

Reviewer 1's Comments:

In this manuscript, Wang et al. introduced a new method to predict the drug-target affinity using advanced machine-learning strategies. This is a quite interesting piece of work that can be potentially useful for the researchers working in the drug design and discovery field, showing theoretical efforts in addressing the challenging problems in this field. First of all, I would like to point out that I am not an expert on algorithm development, so the evaluation of the details of the machine-learning algorithm was not performed. I suggest the editor rely more on the reports from other referees. Overall, the paper reads clearly and the conclusion appears to be based on their results. I don’t have comments.

 

Only one thing needs to be noticed. There is an abbreviation “DTA”, which is not generally used. To me, it also takes me a while to identify what the phrase represents. Please consider to revise.

 

Response: Thank you very much for the careful reading of our manuscript and the valuable suggestions. Because the binding affinity indicates the strength of drug-target interactions, and the main contribution of this work is about the binding affinity prediction. For more clarity, we emphasis this and provide the abbrreviation of DTA in the first sentence of the abstract.

Author Response File: Author Response.pdf

Reviewer 2 Report

This is a really nice piece work proposing a private-public mechanism and MIL to predict DTA. There are sufficient background and detailed description on the method and some good results from the new algorithm.

My only suggestion for the authors is to add more details to the 2.4.2. Public Instance Generator section and also 2.5. MIL section. The authors need to provide more explanation and more detail such that the readers could understand and potentially reproduce what the authors did for this manuscript.

Author Response

Reviewer 2's Comments:

This is a really nice piece work proposing a private-public mechanism and MIL to predict DTA. There are sufficient background and detailed description on the method and some good results from the new algorithm.

My only suggestion for the authors is to add more details to the 2.4.2. Public Instance Generator section and also 2.5. MIL section. The authors need to provide more explanation and more detail such that the readers could understand and potentially reproduce what the authors did for this manuscript.

Response 1: Thank you very much for the careful reading of our manuscript and the valuable suggestions. We add more detailed description in both sections. In 2.4.2 Public Instance Generator section, we explain the details of the Generator, expecially the MHCA module. And in 2.5 Binding Affinity Prediction with MIL section, we provide the combination of the predicted value ^y. The details are as the following.

 

In 2.4.2 Public Instance Generator:

To exploit the drug-target interaction representations from different feature subspaces, MHCA is further employed to perform multiple attention function in parallel h times to generate queries, keys, values matrices Qi, Ki, Vi from i = 1, ..., h. Then the outputs of independent attention are concatenated as the input of a linear transformation to obtain the interaction features, as shown in (14).

 

MultiHead(Q, K, V) = Concat(head1, ..., headh)W^O

where headi = Attention(QW^Q_i , KW^K_i , VW^V_i) (14)

where W^Q_i , w^K_i , W^V_i are the weight matrices in parallel attentions and W^O is the output weight parameter matrix. Take x_1, x_2 as input examples, queris are generated by x_1, keysand values are produced based on x_2. Moreover, the output of the MHCA block Z is computed by:

Z = X1 + MultiHead(Q, K, V) (15)

 

2.5 Binding Affinity Prediction with MIL:

... Therefore, we take the weighted linear combination of the instances as the final binding affinity value as (17), where the weights w1, w2, w3 and w4 are automatically learned during training.

ˆy = w1 ∗ Publicdrug→protein + w2 ∗ Publicprotein→drug + w3 ∗ Publicconcate + w4 ∗ PrivateD

 

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper presents a spontaneous formulation of the drug-target interaction prediction problem as an instance of multi-instance learning. The authors address the issue in three stages, first organizing given drug and target sequences into instances via a private-public mechanism, then identifying the predicted scores of all instances in the same bag, and finally combining all the predicted scores as the output prediction.

The topic is very interesting. The structure of the manuscript is also well. There are several minor questions as follows:

1.       The language of the manuscript does not reach publication in one international Journal. Recommend that the manuscript should be polished by one professional English speaker or company.

2.       There is no discussion section in the paper. Recommend the authors should extend your discussion. The authors should compare the proposed algorithm with the previous literature, such as mentioned algorithms in section 1. Furthermore, the authors may give the advantage and disadvantages of the proposed algorithms.

3.       The conclusions also can be improved. Recommend the authors explore the analysis results briefly and give several valuable points.

 

Hopefully, this will help in the revision of the manuscript.

Author Response

Reviewer 3's Comments:

This paper presents a spontaneous formulation of the drug-target interaction prediction problem as an instance of multi-instance learning. The authors address the issue in three stages, first organizing given drug and target sequences into instances via a private-public mechanism, then identifying the predicted scores of all instances in the same bag, and finally combining all the predicted scores as the output prediction.

The topic is very interesting. The structure of the manuscript is also well. There are several minor questions as follows:

1. The language of the manuscript does not reach publication in one international Journal. Recommend that the manuscript should be polished by one professional English speaker or company.

Response 1: Thank you very much for the careful reading of our manuscript and the valuable suggestions. We have polished this manuscript with AJE (https://aje.com) English language Premium Editing service.  This editing sevice could be verified by their Editing Certificate with the verification code 2812-0B48-C258-2B41-492P via AJE website (https://www.aje.com/certificate). And also, we will still keep following revisions of this manuscript updated by this Premium Editing service.

 

2. There is no discussion section in the paper. Recommend the authors should extend your discussion. The authors should compare the proposed algorithm with the previous literature, such as mentioned algorithms in section 1. Furthermore, the authors may give the advantage and disadvantages of the proposed algorithms.

Response 2:   Thank you for your valuable suggestions. We added the discussions section, and described the comparision with previous algorithms. We also give the advantage and disadvantages in the conclusions section.

 

In this work, we extend computational methods in the field of drug discovery with multiple instance learning which is a popular variation of supervised learning method. Besides, we employ private-public mechanism with different fusion stages to capture interaction's information better. The purpose of this study was to explore a different learning method and deliberate deep model for DTA prediction problem with only raw sequence inputs. The existing works on DTA prediction mostly use different popular techniques to extract useful representations of drugs and proteins, and then the combined representation is fed into complex deep models to find the hidden complex relations between drugs and proteins. The representation learning part has been proven to be efficient in different methods. Some methods, such as DGraphDTA, DeepCDA, DeepConv-DTI and MolTrans, use private representation with late fusion, while some others, such as DeepDTA, GraphDTA, WideDTA use public representation by early fusion. But the contribution of either one to the model and the result is still unknown. So, we want to tackle this challenge more deeply and try to find some insights in the proposed DMIL-PPDTA method.

 

 

3. The conclusions also can be improved. Recommend the authors explore the analysis results briefly and give several valuable points.

Response 3:  We really appreciate your great suggestions and provide aspects for future improvements based on our current work or similar settings in conslusions:

Although DMIL-PPDTA demonstrates good performances, there is still room for further improvements. 1) From the experimental results, all the models show different degrees of performance degradation under blinding setting. This is a kind of classical out-of-distribution (OOD) problems which means that neither the tested drugs nor targets appear in the training set. Thus more effective method should be designed to improve the generalization ability over OOD test set. 2) We formulate the DTA problem as a MIL task, which actually is a Multiple Instance Regression problem. In this work, we utlized the linear regression to fuse the instances information, application-specific fusion method would be beneficial to the performance improvement of predicting drug-target binding affinity.

 

Hopefully, this will help in the revision of the manuscript.

Author Response File: Author Response.pdf