Machine learning methods for predicting cancer drug effects from signaling and cell fate data

• Main Author: Mishra Shital Kumar
• Other Authors: Zheng Jie
• Format: Theses and Dissertations
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Computer science and engineering::Information systems::Models and principles
• Online Access: http://hdl.handle.net/10356/70489
• Institution: Nanyang Technological University

An important area of research within the discipline of the Computational Systems Biology is to investigate the mechanisms of determination of cell fate. As the cellular development progresses, a cell goes through several phenotypes (e.g. Apoptosis, Proliferation, G1, S, G2 and M). Apoptosis, which is a process of programmed cell death in case of a healthy cell, is a special case of cell fate. For example, development of fingers in human embryo occurs due the apoptotic cell death of the cells between digits. Regulatory networks such as transcriptional regulatory network for regulating cell differentiation in the embryonic development, signaling networks for regulating activity level of proteins, etc., determine the course of cell fate and lineage commitment. Signaling pathways play important regulatory roles in apoptosis. However, in cancer cells signaling pathways are in an atypical forms that lead to continuous survival, growth and proliferation of tumor cells. A major objective in cancer research is to investigate the dynamics of signaling pathways that influence the apoptosis of tumor cells. The changes in signaling pathways supporting uncontrolled cell growth, termed as rewiring, can lead to dysregulation of cell fates. Hence comparative analysis of normal and oncogenic signal transduction pathways may provide insights into dynamics of cancer drug induced signaling changes and enable the discovery of novel and effective anti-cancer therapies. A better understanding of rewiring is essential for developing effective anti-cancer therapeutic strategies. In this thesis, we first demonstrate how rewiring of signaling pathways regulate diverse signaling molecules in an apoptotic network. We use a simple yet intuitive differential equation based computational model to understand the roles played by certain essential regulatory proteins in the regulation of programmed cell death. Inspired by the success of dynamical modeling and data analysis in cancer biology, we propose a hybrid modeling approach combining computational models with experimental phosphoproteomics data. We use a hybrid modeling approach based on ordinary differential equation (ODE) models and machine learning techniques to map network rewiring in the apoptotic pathways that may be responsible for the increase of drug sensitivity of tumour cells in triple-negative breast cancer. Our method employs Genetic Algorithm to search for the most likely network topologies by iteratively generating simulated protein phosphorylation data using ODEs and the rewired network and then fitting the simulated data with real data of cancer signaling and cell fate. Most of our predictions are consistent with experimental evidence from literature. Combining the strengths of knowledge-driven and data-driven approaches, our hybrid model can help uncover molecular mechanisms of cancer cell fate at the systems level. Secondly, we investigate the context-specific drug effects on diverse cancer cell lines. We construct a knowledge-based model of ordinary differential equations (ODEs) for the apoptotic signaling network and subsequently infer model parameters (e.g. reaction rates) from real phosphoproteomics data for three breast tumor cell lines, i.e. BT-20, MCF7 and MDA-MB-453 using a Bayesian framework of inference. The model is used to predict apoptosis in response to various perturbations such as caspase knockdown for each of the three cell lines which can be validated using the experimental literature. The inferred changes of the parameters reveal drug effects on diverse cell lines under the treatment with the drug of Erlotinib. Despite having a limited amount of real data, we believe that our in silico apoptosis model is able to capture several dynamic characteristics of drug sensitivity of tumor cells in a coherent way. We hope that as new and larger datasets become available in future, our proposed model can be used to gain more in-depth and quantitative understanding about the regulation of cancer cell death. Our studies offer a novel method of hybrid modeling for linking the computational model with biological data. It is demonstrated to be promising for explaining and predicting the impact of anti-cancer therapies on cancer cells at the systems level.