Identifying the Primary Odor Perception Descriptors by Multi-Output Linear Regression Models
Round 1
Reviewer 1 Report
I suggest to accept the paper with (very) small revision.
Figure 9 and its description could perhaps be improved. It is difficult to understand why POPDs has different values ( 51 and 95), in the figure and the explanation does not completely clarify the understanding.
Also, the reference to figure 9 comes after figure 10.
There are small writing errors (capital letters after the comma, misspelled acronyms, repeated sentence pieces...) that should be corrected.
Author Response
Dear reviewer: Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the re-submitted files. Thank you again! The point-to-point list of responses and actions is as follows:
Reviewer#1, Concern # 1: Figure 9 and its description could perhaps be improved. It is difficult to understand why POPDs has different values (51 and 95), in the figure and the explanation does not completely clarify the understanding. Also, the reference to figure 9 comes after figure 10.
Author response and action: Thank you for pointing this out. We are very sorry for our incorrect writings. We have changed the “POPDs” on the right column into “NPOPDs”, and the reference to figure 9 and figure 10 has been reordered. Please see pages 15 and 16. Also, the descriptions of figure 9 and figure have been revised. Please see lines 364 to 389, pages 14 and 15 for figure 9, and lines 409 to 413, page 16 for figure 10.
Reviewer#1, Concern # 2: There are small writing errors (capital letters after the comma, misspelled acronyms, repeated sentence pieces...) that should be corrected.
Author response and action: Thank you for your carefully reviewing. According to your suggestions, we have checked the whole paper and made some modifications.
In Discussion, “It” after the comma has been changed to “it.” Please see line 498, page 21. “Our” after the comma has been changed to “our”. Please see line 517, page 21.
We have found three misspelled acronyms and corrected them all. Please see lines 480, 481, and 488, page 20. Besides, we have corrected the misspelled “predictable” to “predictable”. Please see the caption of figure 11, page 18.
We have checked the whole paper and revised or removed the repeated sentence pieces.
The brief illustration of our work in Introduction has been revised and streamlined by deleting the duplicated content in Method and Discussion. Please see lines 87 to 106, page 3.
We have modified the content in “2.2 POPDs and the identification frameworks” and deleted the similar sentence pieces with “2.2.1 Disjunctive Models”. Please see lines 149-179, pages 5 and 6.
The first paragraph in “3.4 Performance of Combinational Models and POPDs Identified” has been removed.
The repeated sentence pieces in Conclusion have been deleted. Please see lines 542 to 550, page 22.
The repeated sentence pieces in the last paragraph of 3.4, “It should be noted that after implementing 1000 times and processing with the updating strategy, the results of the disjunctive model are quite different from the results of clustering algorithms. Therefore, these results could not be explained by the theory of clustering.”, has been deleted.
Thank you again for your carefully reviewing and wonderful comments!
we appreciate your warm work earnestly, and hope the correction will meet with approval.
Special thanks to you for your good comments.
Best regards!
Xin Li
Author Response File: 
Author Response.docx
Reviewer 2 Report
The article by Xin Li and co-authors addresses an important issue: the terms used to describe odors. In recent years, there has been a significant increase in the number of descriptors used to describe odors in sensory analysis. The present work aims at doing the opposite by trying to reduce with the help of machine learning the number of descriptors to avoid redundancy. The applications in the field of food, beverages and perfumes are numerous.
Generally speaking, the article is written in a clear and pleasant style, very fluid. The abstract reflects the content of the article and is very clear. The methodology used is also well explained step by step. I am not a specialist in machine learning so I cannot evaluate the scientific quality of the approach used but it is clearly explained for non-specialists. The discussion and the conclusion of the article are also presented in a structured way.
As far as the introduction is concerned, I find that it focuses very quickly on the descriptors of odors and does not mention the difficulty of describing odors. Indeed, the variability between people, according to age is important. Partial anosmia is described and in general the link between the structure of odorant molecules and their odor is not obvious. Close molecules like enantiomers are described in a very different way whereas molecules with very different structures will be described in an identical way. This point should be developed in the introduction because it reinforces even more the relevance of this study by highlighting the difficulty to describe odors. For a recent review on the topics see https://doi.org/10.3390/ijms20123018.
Author Response
Dear reviewer: Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the re-submitted files. Thank you again! The point-to-point list of responses and actions is as follows:
Reviewer#2, Concern # 1: Generally speaking, the article is written in a clear and pleasant style, very fluid. The abstract reflects the content of the article and is very clear. The methodology used is also well explained step by step. I am not a specialist in machine learning so I cannot evaluate the scientific quality of the approach used but it is clearly explained for non-specialists. The discussion and the conclusion of the article are also presented in a structured way.
As far as the introduction is concerned, I find that it focuses very quickly on the descriptors of odors and does not mention the difficulty of describing odors. Indeed, the variability between people, according to age is important. Partial anosmia is described and in general the link between the structure of odorant molecules and their odor is not obvious. Close molecules like enantiomers are described in a very different way whereas molecules with very different structures will be described in an identical way. This point should be developed in the introduction because it reinforces even more the relevance of this study by highlighting the difficulty to describe odors.
Author response and action: Thank you for pointing this out. Your comments are very helpful for us to revise and improve this paper. We have revised the introduction, highlighting the difficulty of describing odors. Please see lines 21 to 28, page 1. We have added several references. Please see lines 579 to 591, pages 22 and 23.
we appreciate your warm work earnestly, and hope the correction will meet with approval.
Special thanks to you for your good comments.
Best regards!
Xin Li
Reviewer 3 Report
The submitted manuscript reports the presentation of an approach able to reduce the number of odor perception descriptors. The matter of primary odors is a challenging and controversial issue. The objective of the work is very interesting.
The purpose of the work is clearly explained, and the current state of the knowledge is very well developed, and precisely supported by the cited references.
Especially this study points out the role of the links between the odors and the difficulty inherent to the sparsity of odors space. The authors present a methodology using clustering calculation by machine learning. The goal is to separate non-predictable odor perception descriptors from predictable odor perception descriptors using a disjunctive model based on the clustering centers. A very important point and advantage is that no odor perception descriptors are ignored, and globally the concept is very interesting.
The non-predictable odor perception descriptors are regarded as primary odor perception descriptors (POPDs), and consequently the predictable odor perception descriptors as non-primary odor perception descriptors (NPOPDs). It seems that such notation could cause confusion between “primary” and “predictable” in the abbreviations.
The limits as well future improvement and developments of the descripted approach are mentioned in the discussion. Indeed, several points appears critical. As emphasized by the author, the Dravnieks dataset is remarkably dense. In the usual odor datasets and databases, the number of molecules largely exceed the number of odor descriptors (several thousand molecules and hundreds odor descriptors, hence some 10 times more molecules than the odor descriptors). With a number of odor descriptors equivalent -and lightly above- to the number of molecules, the Dravnieks dataset seem to be an exception. It will very interesting to test the approach on large datasets (is this planed?).
Besides, there are 95 POPDs: it is a lot comparatively to “only” 51 NPOPDs. In other words, the number on non-primary odors accounts for only one third of all odor descriptors. If that is the same for larger data sets, the number of POPDs would be excessive. Nevertheless, a supposition could be that the same POPDs would emerge, and a larger number of NPOPDs would remain.
The titles and scales of the axes are very small and almost unreadable. It is very difficult to read the legends, which are less than 8 pt. in size (5 pt is really too small). In addition to an increased font size, several supplementary tables would be useful to improve the understanding of graphs.
Line 81-106 describes the way of the identification of POPDs vs NPOPDs. The explanation is clearly stated, but in very dense text, from the introduction (it seems not completely appropriate, to except if required by the recommendations to authors). Besides, the text of the summary of the procedure is well adapted to introduction (lines 113-121). Would the full description not be better placed close to the figures 1 and 2? The approach is clearly described, nevertheless, to facilitate the understanding, I could suggest to separate the text page 3 into several paragraphs (for example at least at line 86), and if possible shorter sentences (especially lines 94-98).
The Figure 5 is placed bizarrely before the part 3 (“Experimental result”) and will be in better placed below the second paragraph of "3.1. Statistical Analysis of Odor Perceptual Data".
Figure 6: there are 14 odor perception descriptors while the legend indicates “ten clustering-center odor perception descriptors “ (“incense” only in figure 6a, “rancid” only in figure 6b). Indeed, four descriptors have no chance to be selected, which results in only ten odor perception descriptors have the probability of being selected. The explanation is clearly provided bellow (line 354-363), nevertheless the text of legend and caption should be modified, especially because the figure is far from text that refers.
Figure 7 and lines 371-278: it is very difficult (unfeasible) to identify the odor perception descriptors. I agree that the names would be unreadable, but a list of the odors in supplementary files would be useful to better appreciate which are ones that have no probability to be a clustering-center.
The figures 7 and 8 are also far from the text which they refer to. Positioning of large figures is not easy, but it is also difficult to understand the meaning of figures without its near explanation. In fact, the figures related to “3.3. Clustering-Center Odor Perception Descriptors Derived from Disjunctive Model” are in “3.4. Performance of Combinational Models and POPDs Identified”.
In the text, the first mention of Figure 9 appears line 445, and the first mention of Figure 10 line 403. That complicates the understanding…
The text of lines 408-409 (“As shown in Figure10(a), (b) and (c), 0.7, 0.8, and 0.9 are the predicting metric thresholds, respectively”) seems not appear to be in line with the legends of figure 10 (“(a) The maximum numbers of the predictable odor perception descriptors under the predicting metric threshold of 0.9.”), etc. …
Equation 4: it is difficult to distinguish between “p” and the notation used for the mean of p. It is possible to use the classical mathematical notation?
Figure 11: predictable (in place of “predictable”).
Figure 9, right: POPDs (51): probably LEMON (in place of “LEMO”).
Line 54: “Madany Mamlouk and Martinetz reported that an upper bound of…. [30], and line 674 the reference 30 appears as “Mamlouk, A.M.; Martinetz, T. …” , which can cause confusion.
Author Response
Dear reviewer: Thanks very much for taking your time to review this manuscript. I really appreciate all your comments and suggestions! Please find my itemized responses in below and my revisions/corrections in the re-submitted files. Thank you again! The point-to-point list of responses and actions is as follows:
Reviewer#3, Concern # 1: The limits as well future improvement and developments of the descripted approach are mentioned in the discussion. Indeed, several points appears critical. As emphasized by the author, the Dravnieks dataset is remarkably dense. In the usual odor datasets and databases, the number of molecules largely exceed the number of odor descriptors (several thousand molecules and hundreds odor descriptors, hence some 10 times more molecules than the odor descriptors). With a number of odor descriptors equivalent -and lightly above- to the number of molecules, the Dravnieks dataset seem to be an exception. It will very interesting to test the approach on large datasets (is this planed?). Besides, there are 95 POPDs: it is a lot comparatively to “only” 51 NPOPDs. In other words, the number on non-primary odors accounts for only one third of all odor descriptors. If that is the same for larger data sets, the number of POPDs would be excessive. Nevertheless, a supposition could be that the same POPDs would emerge, and a larger number of NPOPDs would remain.
Author response and action: Thank you for providing these insights! You have raised an important question, and we agree with your assessment. Compared to 51 NPOPDs, 95 POPDs is much larger indeed. It might be caused by the olfactory system's complexity, the size of the dataset, and the algorithms. In the future, we will try more advanced algorithms to narrow down the number of POPDs. Also, we will test this approach on large datasets. Thank you for your valuable suggestion again!
Reviewer#3, Concern # 2: The titles and scales of the axes are very small and almost unreadable. It is very difficult to read the legends, which are less than 8 pt. in size (5 pt is really too small). In addition to an increased font size, several supplementary tables would be useful to improve the understanding of graphs.
Author response and action: Thank you for your suggestion! We have increased the font size of all the figures and hope the correction will meet with approval. We have carefully reorganized the position of all the figures by placing them close to their descriptions. Besides, we have submitted two supplementary tables about the odorants and odor perception descriptors applied in the Dravnieks dataset. We appreciate your valuable advice earnestly again!
Reviewer#3, Concern # 3: Line 81-106 describes the way of the identification of POPDs vs NPOPDs. The explanation is clearly stated, but in very dense text, from the introduction (it seems not completely appropriate, to except if required by the recommendations to authors). Besides, the text of the summary of the procedure is well adapted to introduction (lines 113-121). Would the full description not be better placed close to the figures 1 and 2? The approach is clearly described, nevertheless, to facilitate the understanding, I could suggest to separate the text page 3 into several paragraphs (for example at least at line 86), and if possible shorter sentences (especially lines 94-98).
Author response and action: Thank you for pointing this out! We have revised and streamlined the explanation of the POPDs identification framework in Introduction, and separated them into three paragraphs. Please see lines 87 to 106, page 3. Also, we have repositioned figure 1 and figure 2 to make the description close to the figures. Please see pages 3 and 4.
Reviewer#3, Concern # 4: The Figure 5 is placed bizarrely before the part 3 (“Experimental result”) and will be in better placed below the second paragraph of "3.1. Statistical Analysis of Odor Perceptual Data".
Author response and action: Thank you for your suggestion! The position of figure 5 has been adjusted according to your suggestion. Please see figure 5 on page 10.
Reviewer#3, Concern # 5: Figure 6: there are 14 odor perception descriptors while the legend indicates “ten clustering-center odor perception descriptors “ (“incense” only in figure 6a, “rancid” only in figure 6b). Indeed, four descriptors have no chance to be selected, which results in only ten odor perception descriptors have the probability of being selected. The explanation is clearly provided bellow (line 354-363), nevertheless the text of legend and caption should be modified, especially because the figure is far from text that refers.
Author response and action: Thank you for your kind suggestion! Figure 6 has been reformatted to make it close to its exploration text, and we have revised the caption of figure 6. Please see page11. The clustering algorithms in the disjunctive model are executed with 100, 200, 300, 400, 500,600, 700, 800, 900, and 1000 times. For each number of clustering algorithm’s execution times, the disjunctive models are implemented 30 times. When the number of execution times of the clustering algorithm is small, the results of the ten CCOPDs are not stable; that is, 14 odor perception descriptors might be selected as the ten CCOPDs during the 30 implementations of the disjunctive model. With the increase of the number of execution times of clustering algorithms, the CCOPDs results will be stabilized on a fixed ten odor perception descriptors.
Reviewer#3, Concern # 6: Figure 7 and lines 371-278: it is very difficult (unfeasible) to identify the odor perception descriptors. I agree that the names would be unreadable, but a list of the odors in supplementary files would be useful to better appreciate which are ones that have no probability to be a clustering-center.
Author response and action: Thank you for your constructive suggestion. We have provided a supplementary file 1 to list all the 146 odor perceptions in order. Please see line 349, page 12.
Reviewer#3, Concern # 7: The figures 7 and 8 are also far from the text which they refer to. Positioning of large figures is not easy, but it is also difficult to understand the meaning of figures without its near explanation. In fact, the figures related to “3.3. Clustering-Center Odor Perception Descriptors Derived from Disjunctive Model” are in “3.4. Performance of Combinational Models and POPDs Identified”.
Author response and action: Thank you for you carefully reviewing! We have repositioned figure 7 and figure 8. Please see pages 14 and 15.
Reviewer#3, Concern # 8: In the text, the first mention of Figure 9 appears line 445, and the first mention of Figure 10 line 403. That complicates the understanding…
Author response and action: Thank you for your kindly reminder! The order of figure 9 and figure 10 has been exchanged, and their explanation text is revised. Please see lines 376 to 389 and lines 409 to 413, pages 14 to16.
Reviewer#3, Concern # 9: The text of lines 408-409 (“As shown in Figure10(a), (b) and (c), 0.7, 0.8, and 0.9 are the predicting metric thresholds, respectively”) seems not appear to be in line with the legends of figure 10 (“(a) The maximum numbers of the predictable odor perception descriptors under the predicting metric threshold of 0.9.”), etc. …
Author response and action: Thank you for your suggestion! We have reordered the subfigures in figure 10, and revised the explanation text. Please see lines 376 to 386, pages 14 and 15.
Reviewer#3, Concern # 10: Equation 4: it is difficult to distinguish between “p” and the notation used for the mean of p. It is possible to use the classical mathematical notation?
Author response and action: Thank you for your advisable comments. Equation 4 has been revised by replacing “p” with “x”, and the explanation of this variables also is revised. Please see lines146 to 147, page 5.
Reviewer#3, Concern # 11: Figure 11: predcitable (in place of “predictable”).
Author response and action: Thank you for pointing this out. We have corrected this misspelling. “predcitable” in Figure 9 has been replace with “predictable”. Please see page 18.
Reviewer#3, Concern # 12: Figure 9, right: POPDs (51): probably LEMON (in place of “LEMO”).
Author response and action: Thank you for your kindly reminder! “LEMO” in figure 10 (figure 9 in the original manuscript) has been replace with “LEMON”. Please see page 16.
Reviewer#3, Concern # 13: Line 54: “Madany Mamlouk and Martinetz reported that an upper bound of…. [30], and line 674 the reference 30 appears as “Mamlouk, A.M.; Martinetz, T. …” , which can cause confusion.
Author response and action: Thank you for pointing this out. We have changed “Madany Mamlouk” to “Mamlouk”. Please see line 62, page 2.
we appreciate your warm work earnestly, and hope the correction will meet with approval.
Special thanks to you for your good comments.
Best regards!
Xin Li
Author Response File: Author Response.docx
