Population-Based Parameter Identification for Dynamical Models of Biological Networks with an Application to Saccharomyces cerevisiae
Round 1
Reviewer 1 Report
The present manuscript "Population-based Parameter Identification for Dynamical Models of Biological Networks with an Application to Saccharomyces cerevisiae" by Ewelina Weglarz-Tomczak, Jakub Tomczak, Agoston Eiben, and Stanley Brul introduces a Python-based software framework for population-based optimization algorithms, POPI4SB, for identification of parameters in systems of ordinary equations. Therein, the python software PySCeS is employed to solve the forward problem, i.e., the system of ordinary equations. Furthermore, two models which combine population-based optimization with a kNN regression surrogate model are introduced. In the manuscript, the considered population-based optimization methods and the model problem, a system of ordinary equations modeling the glycolytic pathway in Saccharomyces cerevisiae, are first described. Then, a numerical study is performed to investigate the performance of the algorithms and the software framework.
As proven by the numerical results, both the considered methods and the software framework work quite well for the considered type of problem. Furthermore, the manuscript is mostly written well and has a clear structure. Therefore, I would suggest the manuscript for publication after a the following comments have been addressed in a revised version:
+ Line 39-40: What does „the framework allows adding new optimizers to a single file“ exactly mean? Moreover, the software framework should be explained in more detail. The illustration in Fig. 1 is quite high-level and probably applicable to most parameter identification software frameworks. The description of POPI4SB could be more specific.
+ Is Fig. 2 just a schematic illustration of population-based optimization in general, or does it correspond to a specific example?
+ I do not understand the point of the sentence „If the time required to obtain a value of the objective function (or the fitness function in the context of EA) is low, then their computational complexity is linear with respect to the size of the population N.“
+ Line 101: Are there any restrictions on the the size of the scaling factor F or could it really be any positive real number?
+ I would be interested in the application of other surrogate models, e.g., neural networks. How would these compare to the kNN regressor? For what reasons has kNN regression been employed instead?
+ In the RevDE+ and EDA+ approaches, how often is forward problem solved and how often is the kNN regressor applied?
+ The investigation of the computing time for the different approaches would be essential here. In particular, the evaluation time of kNN can be significant for large numbers of data. Furthermore, it would be important to know the numerical schemes employed to solve the forward problem in PySCeS. How many time steps are computed, and how much computing time is needed to solve the forward problem?
+ I am wondering if numbering all the equations (11) - (59) is really necessary.
+ Line 199: Why is a Gaussian noise with specifically 3% standard deviation used?
+ Why are the experiments repeated only three times? In particular, when the mean and confidence interval is considered (Fig. 5), the number of experiments should be larger than three.
+ None of the methods seem to be able to get below an objective value of 3.675. Can this be explained?
+ In Fig. 6, only absolute differences are considered. It would be interesting to consider relative errors as well.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The current article is really interesting and its implementation on Python is relevant. The topic is relevant from both ends, practicality and operational modeling when dealing with complex problems, and also computationally important taking into account the wide use of difference libraries in Python.
However, I would like to see a further discussion about the the following:
- What are the differences between the method presented and the Genetic Algorithm optimization? It appears that the presented method resembles many aspects of GA-optimization and Gene Expression Programming.
- If this particular problem would be compared with a derivative-method, what would be the advantages and disadvantages of each of them?
I would like to suggest authors to consider also, if possible to include in the introduction a clear description of the difference between optimization methods. It will direct the reader easier to the reasons for such new adoptions. Besides that, I recommend to consider some extra bibliography about the subject, as for example:
Larson J., Menickelly M., Wild S.M. http://optimization-online.org and the extensive bibliography included there and linked to the Arxiv.org database.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
