Continous-discrete approach of tumor growth: multiscale modeling and Bayesian inference using Gaussian Process Surrogates

• Autor(a) principal: Rocha, Heber Lima da
• Data de Publicação: 2019
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do LNCC
• Texto Completo: https://tede.lncc.br/handle/tede/319
• topic: Câncer; Processos gaussiano; CNPQ::CIENCIAS DA SAUDE::MEDICINA::CLINICA MEDICA::CANCEROLOGIA

The reliability of computational predictions of complex physical and biological events is one of the most relevant aspects of predictive computational science. Without a rigorous approach to assessing the reliability of models subject to various types of uncertainties, computational models are of little applicability in medical science or science in general. Thus, the success and clinical use of mathematical and computational models ultimately depend on how key issues in predictive science are addressed. In this doctoral dissertation research, we extend the two-scale hybrid model of tumor avascular growth developed previously to include mechanisms of intracellular regulation and angiogenesis. The vascular model has been implemented in a very flexible way, so that it is possible to include different mechanisms in the various scales as well as their simplification, aiming to focus on some specific dynamics. A three-dimensional version of the model has also been developed. Model parameters associated with the three scales (tissue, cellular and subcellular) were estimated at first to represent the typical dynamics of non-specific carcinomas. Usually, such parameters cannot be directly determined and should be inferred from experimental evidence/data. As a step in this direction, we have developed a platform for integrating data at the cellular scale of the multiscale model in order to calibrate it using a Bayesian approach. As the model is stochastic and of high computational cost, we study the application of several approximate Bayesian approach techniques. In particular, the Approximate Approximate Bayesian Computation (AABC) approach methodology, based on the use of a metamodel built from a limited number of model simulations, has proved to be computationally efficient. To further improve the construction of the metamodel, we developed a reduced model based on Gaussian processes (GPR). From parametric space samples via the sparse Latin hypercube method, we developed an adaptive strategy for the construction of the metamodel. ABC-GPR combination methods were used to calibrate the model at the cellular scale using in vitro data from the confluence of BT474 cells (human breast cancer cells). They present significantly higher computational efficiency, being potential strategies to be used for the calibration of our complete multiscale model. Because the developed parameter calibration approach is conceptually model-independent, it has the potential to be used in the inference of parameters of computationally expensive models used in various fields of science and engineering.