SummerSim 2016: Simulation of Tumor Necrosis in Primary Melanoma - Response for Reviewers

[Jacob Barhak]

Below s the Response for reviewers submitted as part of the camera ready manuscript

##########

Rebuttal I carefully read the reviews and in the final version of the paper I did my best to comply to all reviewers' objections. Below I would like to explain in details the most important ones. 1) The important remark concerns the differences between unpublished paper I submitted before to ICCS conference, which will be held in San Diego June 5-8, 2016 (entitled: Supermodeling in simulation of melanoma progression) and the SCSC paper entitled: Simulation of Tumor Necrosis in Primary Melanoma. The first paper is available from the Researchgate portal just clicking the following link: https://www.researchgate.net/publication/286930455_Supermodeling_in_simulation_of_melanoma_progression On the one hand the papers are similar in spirit, showing the advantages of melanoma supermodel, but on the other, they are different in both the main focus (context) and �maturity� of the concept presented. The ICCS version appeared a few months earlier thus the version submitted to SCSC is: 1. more advanced in explaining the supermodeling idea, 2. it uses more advanced model of cancer with a crucial vessel remodeling part, 3. the microscopic simulations of melanoma progression by using particle automata model -PAM were performed, commented and compared to macroscopic results. 4. what is the most important, SCSC paper shows a viable biological argument in favor of supermodeling by focusing on tumor ulceration process. The ICCS paper was written with cooperation of professor dr OV Vasylyev from University of Colorado at Boulder, who is the expert in numerical modeling, while the biological context of the SCSC paper was approved by professor dr A. Dudek the oncologist from University of Illinois, College of Medicine. Apart from the model equations and the table with values of parameters, the whole narration is different in these two papers eliminating self-plagiarism. Consequently, I am very particular to discuss the supermodeling approach with various scientific audiences staying behind ICCS and SCSC conferences. (2) The current version of the paper was improved greatly comparing to its former one to show the most evident advantages of supermodeling. Frankly, I have to admit that they were not clearly expressed before, what confuses the first reviewer. First, we postulate that the supermodeling is an alternative to multi-scale modeling, which uses very sophisticated but intractable and overfitted mathematical models �of everything�. Such the multiscale model can have hundreds of free parameters, which have to be adapted from data. Just this fact lets us to forget about serious application of such the models in clinical use (tremendous computational complexity). Unlike multiscale model the supermodel is made of a �base model� of cancer, representing its macroscopic growth by using PDEs. Moreover, it should be the most parsimonious but effective model of cancer, which includes only the most important processes influencing cancer development (in analogy to ensemble classifiers in machine learning, where a single classifier should be very simple � but not simpler). The supermodel consists of a few coupled instances of the base model (I removed the confusing �imprecise model� from the text). The coupling is clearly defined by Eq. 16. The good metaphor explaining the supermodeling is a dynamic system of balls (representing the base model instances) connected with springs (couplings) searching for global minimum in a multimodal function landscape. A ball, which is the lowest, attracts the rest of balls, which start to move towards it. Meanwhile, the other ball may find even lower location, then it attracts the rest ones. And so on. This procedure allows for better exploration of the function domain. Thus our hypothesis is as follows: H: It is possible to obtain reliable prognoses about cancer dynamics by creating the supermodel of cancer, which consists of several coupled instances (the sub-models) of a base cancer model. Its integration with real data can be achieved by employing a prediction/correction learning scheme focused on fitting several values of coupling coefficients between sub-models, instead of matching scores (even hundreds) of tumor model parameters as it is in the classical data adaptation techniques. We postulate that there exist a generic coarse-grained computer model of cancer, which can be used as a computational framework for developing high quality supermodels. The latent fine-grained tumor features e.g. microscopic processes and other unpredictable events accompanying its proliferation not included in the model, are hidden in incoming real data. Summarizing, the main advantage of supermodeling is its greater feasibility of clinical implementation than tremendously complex multiscale models. Integration of a base mathematical model (which consists of the coarse-grained processes of tumor development) with data models (which include fine-grained processes and environmental factors) is possible due to synchronization of sub-models with real data in prediction/correction learning scheme aimed at a reasonable number of coupling coefficients. This is especially novel approach, in the advent of big data era. The formal simple models can be used as additional knowledge in machine learning processes on big data. (3) The other objectives were addressed in the final text of the paper. Thank you very much to all reviewers for their work and contribution to this paper. Witold Dzwinel