Using Deep 1D Convolutional Grated Recurrent Unit Neural Network to Optimize Quantum Molecular Properties and Predict Intramolecular Coupling Constants of Molecules of Potential Health Medications and Other Generic Molecules
, Emmanuel Gbenga Dada 2, Onyeka Emebo 3 and Olugbenga Oluseun Oluwagbemi 4,5
Round 1
Reviewer 1 Report
Major revision is required
1. Ranale of the study should be explained properly
2. Explain the data set construction process in detail.
3. Table of QM properties should be provided along with the text
4. Quality of figures should be improved
5. Better to explain training loss and validation loss in the manuscript
6. References should be revised according to the journal guidelines.
Author Response
REVIEWER 1 COMMENTS:
- Ranale of the study should be explained properly
RESPONSE: Thank you for your comment. I believe the reviewer means Rationale and not Ranale. The authors have explained the rationale of the study.
- Explain the data set construction process in detail.
RESPONSE: The dataset was obtained from kaggle which was explained in section 3.1
- Table of QM properties should be provided along with the text
RESPONSE: Thank you for that suggestion. The Table of Quantum properties has been provided along with the text in Table 1 line 230
- Quality of figures should be improved
RESPONSE: Thank you. The author has improved the quality of figure 6-13 in the manuscripts.
- Better to explain training loss and validation loss in the manuscript
RESPONSE: Thank you for that suggestion. The training and validation loss curve has been explained in the paper.
- References should be revised according to the journal guidelines.
RESPONSE: Thank you. The author has revised all the references using Chicago reference style which is acceptable to the journal. The changes can be seen on pages 20-22 of the revised manuscript.
Reviewer 2 Report
The authors developed a 1D-CNN-GRU neural network for prediction of scalar coupling constants between atoms belonging to the same molecules.
The data for training/validation were taken from an online open database hosted at kaggle.com.
Application of CNNs to prediction of chemical properties is not novel and it has been extensively investigated in the last 4 years. The concept of assembling a multilayer NN with convolutional, maxpooling, GRU, dropout, and fully connected layer is quite conventional, but it is still refreshing in its application to the problem envisioned by the authors.
However, the authors fail to make a point about the aim of their work.
The declared scope is to predict chemical properties of potential liquid drugs or medication. In the abstract, they even mention (line 31) the potential application of 1D-CNN to target recognition and drug reprofiling, which is not discussed elsewhere in the paper.
How prediction of intramolecular scalar coupling constants can provide insights about the chemical properties of liquid drugs is not clarified, nor discussed. Furthermore, the way normalization was applied to the present work actually defeats the scope of the prediction, reducing it to a mere classification routine of high vs low constant value.
As such, the results do not corroborate the initial statement of the work, and the paper should be rejected in its entirety.
Author Response
REVIEWER 2 COMMENTS:
- The authors developed a 1D-CNN-GRU neural network for prediction of scalar coupling constants between atoms belonging to the same molecules.
The data for training/validation were taken from an online open database hosted at kaggle.com.
Application of CNNs to prediction of chemical properties is not novel and it has been extensively investigated in the last 4 years. The concept of assembling a multilayer NN with convolutional, maxpooling, GRU, dropout, and fully connected layer is quite conventional, but it is still refreshing in its application to the problem envisioned by the authors.
RESPONSE: Thank you for your comment. Application of 1D-CNN has been investigated but to the best of our knowledge. There is no publication (journal paper, conference proceeding or book chapter) that has studied molecular property with 1D-CNN-GRU. This shows the novelty of the work. Moreover, the parameters and approach considered in this paper is totally different from other papers.
- However, the authors fail to make a point about the aim of their work.
RESPONSE: Thank you for your comment. The aim of this research has been stated on page 3 of the paper.
- The declared scope is to predict chemical properties of potential liquid drugs or medication. In the abstract, they even mention (line 31) the potential application of 1D-CNN to target recognition and drug reprofiling, which is not discussed elsewhere in the paper.
RESPONSE: Thank you for that observation. It has been corrected in the Abstract.
4, How prediction of intramolecular scalar coupling constants can provide insights about the chemical properties of liquid drugs is not clarified, nor discussed.
RESPONSE: Thank you for this suggestion. The authors have explained it in the conclusion sections.
- Furthermore, the way normalization was applied to the present work actually defeats the scope of the prediction, reducing it to a mere classification routine of high vs low constant value.
RESPONSE: Normalize data is very important in prediction. The following are the importance of data normalization which was also explained in the manuscripts.
- Data normalization aids in the improvement of data speed, accuracy, and efficiency, as well as the removal of undesired outliers.
- It aids in the reduction of data alteration issues.
- Data normalization reduces duplication by ensuring that only relevant data is stored in each table.
- As such, the results do not corroborate the initial statement of the work, and the paper should be rejected in its entirety.
RESPONSE: The main purpose of this research is to predict molecular property using 1D-CNN-GRU which was explained in the manuscript. The results have substantiated the initial statement of the work mentioned in the abstract and introductory sections of the paper. It is therefore our believe that the paper should be accepted.
Reviewer 3 Report
Reviewer comments:
It is challenging to predict the molecular properties of liquid pharmaceuticals or therapies to address health concerns in drug development. Machine learning methods for precise prediction can help speed up and lower the cost of identifying new medication. In this study, a 1D-CNN-GRU method is proposed to offer the powerful capability of signal feature extraction as well as the capacity to forecast successive scalar coupling constant signals. The experimental results indicate that the proposed method achieves better prediction accuracy than other existing methods.
This manuscript is well-written and well-organized. The objective is well-articulated and reached. The figures and tables are presented in a clear and appropriate manner and are consistent with the description in the text. The results and analysis presented in the manuscript are interesting for this field, and Applied Sciences is the appropriate journal to submit it. But there are still some points that the authors should consider, as described in the following. Also, some suggestions are provided, in case the authors consider them interesting to carry out.
In line 150, “mechanics-based machine learning (QML) combines …”. Should “mechanics” be capitalized at the beginning of this sentence? However, the abbreviation QML doesn’t make sense. Maybe it should be quantum mechanics-based machine learning (QML).
In line 156, “Montavon et al. [23] use a learning-from-scratch technique for …”. “use” should be “used” here.
In line 195, “Finally, they use SchNet to predict potential-energy areas …”. “use” should be “used” here.
In Fig. 1, please label the X and Y axis.
In Table 2, please adjust the font size and show each number in the same row.
In line 318, there is an extra space before (ht-1). In line 329, there is an extra space after (ht-1).
In line 326, U should be Uz according to the Eq. (5). And both Wz and U are explained as the weight of the update gate. The author should clearly distinguish the definitions between them.
In line 336, U should be Ur according to the Eq. (6). And both Wr and U are explained as the weight of the reset gate. The author should clearly distinguish the definitions between them.
In Eq. (8), is the ht in the right side of Eq. (8) equivalent to the ht in the left side of Eq. (7)? Please provide more details.
In line 353, there are two commas (,) between t and ht-1.
In Eq. (10), please provide explanations of all parameters in this equation.
In line 515-518, “The performance of the Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), One Dimensional Convolutional Neural Network (1DCNN) and the proposed 1D Convolutional Gated Recurrent Unit Neural Network (1D-CNN-GRU) based molecular property systems.” There is no verb in this sentence.
In line 519, “The mean, Standard deviation, minimum, 25%, 50%, 75% and maximum.” There is no verb in this sentence. And “Standard deviation” shouldn’t be capitalized.
In line 521, “Atom_index_1 also has small volatility of 3.27 compared with other molecular property …”. Should Atom_index_1 be Atom_index_0 according to Table 2?
In Fig. 5, please label the X and Y axis.
In line 548-549, “… in this study Such as RNN, LSTM, GRU, 1D-CNN and the proposed method …”. “Such as” shouldn’t be capitalized.
In line 551, “Figure 5-10” should be “Fig. 6-10”.
The author can consider combining Fig. 6-10 into one figure with multiple subplots to make the comparison in an easy way. And the background of Fig. 6-8 is different from that of Fig. 9-10.
The author can consider combining Fig. 11-15 into one figure with multiple subplots to make the comparison in an easy way. And the background of Fig. 11-13 is different from that of Fig. 14-15.
The order of methods in Fig. 6-10, Fig. 11-15 and Table 4 should be consistent, e.g., LSTM, RNN, GRU, 1D-CNN, 1D-CNN-GRU.
In line 562, Fig 11-15 should be Fig. 11-15.
In Fig. 13-15, please label the X and Y axis.
In the comparison in Fig. 6-10, Fig. 11-15, and Table 4, we can see the method RNN. However, RNN is not mentioned in line 115-118, 515-518. Please keep the description consistent in the text.
Author Response
REVIEWER 3 COMMENTS:
- It is challenging to predict the molecular properties of liquid pharmaceuticals or therapies to address health concerns in drug development. Machine learning methods for precise prediction can help speed up and lower the cost of identifying new medication. In this study, a 1D-CNN-GRU method is proposed to offer the powerful capability of signal feature extraction as well as the capacity to forecast successive scalar coupling constant signals. The experimental results indicate that the proposed method achieves better prediction accuracy than other existing methods.
RESPONSE: Thank you for your comment.
- This manuscript is well-written and well-organized. The objective is well-articulated and reached. The figures and tables are presented in a clear and appropriate manner and are consistent with the description in the text. The results and analysis presented in the manuscript are interesting for this field, and Applied Sciences is the appropriate journal to submit it. But there are still some points that the authors should consider, as described in the following. Also, some suggestions are provided, in case the authors consider them interesting to carry out.
RESPONSE: All the authors appreciate your effort in improving this research work. We hope all your suggestions will be carried out.
- In line 150, “mechanics-based machine learning (QML) combines …”. Should “mechanics” be capitalized at the beginning of this sentence? However, the abbreviation QML doesn’t make sense. Maybe it should be quantum mechanics-based machine learning (QML).
RESPONSE: Thank you. In line 150 mechanics-based machine learning was replaced as Quantum mechanics-based machine learning (QML)
- In line 156, “Montavon et al. [23] use a learning-from-scratch technique for …”. “use” should be “used” here.
RESPONSE: Thank you. We have replaced use with used.
- In line 195, “Finally, they use SchNet to predict potential-energy areas …”. “use” should be “used” here.
RESPONSE: Thank you. We have replaced use with used.
- In Fig. 1, please label the X and Y axis.
RESPONSE: Figure 1 contains X and Y values.
- In Table 2, please adjust the font size and show each number in the same row.
RESPONSE: The font size used in this manuscripts was suggested from the applied sciences journal.
- In line 318, there is an extra space before (ht-1). In line 329, there is an extra space after (ht-1).
RESPONSE: Thank you for the suggestion. It has been corrected in line 318 and 329.
- In line 326, U should be Uz according to the Eq. (5). And both Wz and U are explained as the weight of the update gate. The author should clearly distinguish the definitions between them.
RESPONSE: It has been replaced with Uz.
- In line 336, U should be Ur according to the Eq. (6). And both Wr and U are explained as the weight of the reset gate. The author should clearly distinguish the definitions between them.
RESPONSE: It has been replaced with Ur
- In Eq. (8), is the ht in the right side of Eq. (8) equivalent to the ht in the left side of Eq. (7)? Please provide more details.
RESPONSE: Thank you for that observation. The result obtained from equation (7) which is ht was transferred to equation (8)
- In line 353, there are two commas (,) between t and ht-1.
RESPONSE: Thank you, it has been removed.
- In Eq. (10), please provide explanations of all parameters in this equation.
RESPONSE: The authors have provided all the parameters in eqn. 10.
- In line 515-518, “The performance of the Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), One Dimensional Convolutional Neural Network (1DCNN) and the proposed 1D Convolutional Gated Recurrent Unit Neural Network (1D-CNN-GRU) based molecular property systems.” There is no verb in this sentence.
RESPONSE: Thank you for that observation. We have included the verb.
- In line 519, “The mean, Standard deviation, minimum, 25%, 50%, 75% and maximum.” There is no verb in this sentence. And “Standard deviation” shouldn’t be capitalized.
RESPONSE: We have replaced it with small letter and included the verb.
- In line 521, “Atom_index_1 also has small volatility of 3.27 compared with other molecular property …”. Should Atom_index_1 be Atom_index_0 according to Table 2?
RESPONSE: Thank you for that suggestion.
- In Fig. 5, please label the X and Y axis.
RESPONSE: The atom_index represents values from 0 to 25. Most recent paper didn’t label X and Y axis.
- In line 548-549, “… in this study Such as RNN, LSTM, GRU, 1D-CNN and the proposed method …”. “Such as” shouldn’t be capitalized.
RESPONSE: Thank you. It has been replaced with small letter.
- In line 551, “Figure 5-10” should be “Fig. 6-10”.
RESPONSE: Thank you for your suggestion. It has been corrected.
- The author can consider combining Fig. 6-10 into one figure with multiple subplots to make the comparison in an easy way. And the background of Fig. 6-8 is different from that of Fig. 9-10.
RESPONSE: Thank you for your suggestion but most journal prefer single plot. The background of Fig. 6-8 has been removed.
- The author can consider combining Fig. 11-15 into one figure with multiple subplots to make the comparison in an easy way. And the background of Fig. 11-13 is different from that of Fig. 14-15.
RESPONSE: Thank you for your suggestion. The background has been removed.
- The order of methods in Fig. 6-10, Fig. 11-15 and Table 4 should be consistent, e.g., LSTM, RNN, GRU, 1D-CNN, 1D-CNN-GRU.
RESPONSE: Thank you for that observation. The figure and Table 4 is consistent.
- In line 562, Fig 11-15 should be Fig. 11-15.
RESPONSE: Thank you for the suggestion.
- In Fig. 13-15, please label the X and Y axis.
RESPONSE: X and Y axis are values.
- In the comparison in Fig. 6-10, Fig. 11-15, and Table 4, we can see the method RNN. However, RNN is not mentioned in line 115-118, 515-518. Please keep the description consistent in the text.
RESPONSE: Thank you for that observation. We have keep the description consistent throughout the paper.
Round 2
Reviewer 1 Report
Thanks for the correction. The quality of the manuscript may be accepted for publication.
Author Response
Thank you for accepting our manuscripts for publication in your journal, after rigorous corrections.
Reviewer 2 Report
In their revised articles, the authors failed to address the points previously highlighted.
The title and declared aim of the work "Optimizing Quantum Molecular
Properties of potential Health 2 Liquid Medication fluids" is not
met in the results of their work. What they do is predicting intramolecular coupling constants of molecules taken from an online database.
An additional reason of concern is that nothing is said about the nature of the molecules and the database cannot be accessed without joining the competition it was created for. From a first look to the
database, it seems to contain molecules of all kinds, not just compounds that exist in a liquid form. I would like to stress here that I am not claiming that the authors' work is without merit. What
I would like to point out is that there is a clear discrepancy between what they claim and what they do, making the paper highly misleading.
Lastly, I am concerned about the approach followed in the paper. The
training/validation curves reported in figures 6 to 10 go to plateau
in just a couple of epochs, which could be due to overfitting or the
incapability of the NN to properly learn the provided information. Looking at the MAE, RMSE and MSE, the latter can be excluded. However, I would like to see a AUC-ROC curve of the values predicted
by the NN, because in fig. 15 it looks like the network is unable to
actually learn that there are coupling constants with values higher
than 0.45. Indeed, the RMSE is quite high (0.15) after normalization
(values ranging from 0 to 1). In addition, from table 4 is can be seen that the statistics related to the performances of the NN do not
significantly improve in the tested models of NN. As such, I would
like to ask the authors to provide additional analyses, like the
AUC-ROC curve, to better show the capability of their NN in comparison with the other ones.
Author Response
REBUTTAL LETTER FOR MANUSCRIPT NUMBER: applsci-1759896
We appreciate the time and efforts you and other anonymous reviewers have spent providing input and information on ways to strengthen our paper. It is with great pleasure that we are invited to submit a revised draft of our manuscript, titled, “Optimizing Quantum Molecular Properties of potential Health Liquid Medication fluids using Deep 1D Convolutional Grated Recurrent Unit Neural Network” to Applied Sciences.
We have made improvements that reflect the detailed recommendations made by the reviewers. To the best of our knowledge, we have addressed all of the issues and concerns raised in the last email. Our responses to the reviewers’ comments sent to our email on 29th June 2022 are outlined below.
REVIEWER 2 COMMENTS:
- In their revised articles, the authors failed to address the points previously highlighted.
The title and declared aim of the work "Optimizing Quantum Molecular Properties of potential Health 2 Liquid Medication fluids" is not met in the results of their work. What they do is predicting intramolecular coupling constants of molecules taken from an online database.
RESPONSE: Thank you for calling our attention to this concern. We have addressed this concern by updating the title of the revised manuscript to “Using Deep 1D Convolutional Grated Recurrent Unit Neural Network to Optimize Quantum Molecular Properties and Predict Intramolecular Coupling Constants of Molecules of potential Health Medications and Other Generic Molecules”
- An additional reason of concern is that nothing is said about the nature of the molecules and the database cannot be accessed without joining the competition it was created for.
RESPONSE: The database can be accessed freely available from the Kaggle website.
- From a first look to the database, it seems to contain molecules of all kinds, not just compounds that exist in a liquid form. I would like to stress here that I am not claiming that the authors' work is without merit. What I would like to point out is that there is a clear discrepancy between what they claim and what they do, making the paper highly misleading.
RESPONSE: We have addressed this by updating the title of the revised article to address the concern that the reviewer has just raised in order to resolve any misleading claim.
4, Lastly, I am concerned about the approach followed in the paper. The training/validation curves reported in figures 6 to 10 go to plateau in just a couple of epochs, which could be due to overfitting or the incapability of the NN to properly learn the provided information. Looking at the MAE, RMSE and MSE, the latter can be excluded. However, I would like to see a AUC-ROC curve of the values predicted by the NN, because in fig. 15 it looks like the network is unable to actually learn that there are coupling constants with values higher than 0.45. Indeed, the RMSE is quite high (0.15) after normalization (values ranging from 0 to 1). In addition, from table 4 is can be seen that the statistics related to the performances of the NN do not significantly improve in the tested models of NN. As such, I would like to ask the authors to provide additional analyses, like the AUC-ROC curve, to better show the capability of their NN in comparison with the other ones.
RESPONSE: Since the focus of the paper is the use of deep 1D Convolutional Grated Recurrent Unit Neural Network to optimize quantum molecular properties and predict intramolecular coupling constants of molecules of potential health medications and other generic molecules – this is a predictive approach.
Many problems that require the generation of AUC-ROC curve are usually problems that require a classification approach. The problem we have proposed solution to does not require an AUC-ROC curve because we adopted a predictive approach. Our manuscript is based on a prediction problem and not a classification problem, for this reason, it does not require an AUC ROC curve. We have tried different high epochs but we discovered that we are getting the same result. Most of the time higher epoch does not result in obtaining result.
