Artificial Intelligence (AI) and machine learning-driven automation of complex, neurobiological model reduction: A framework based on deep learning, ensemble learning, and sensitivity analysis methodologies : A thesis submitted in partial fulfilment of the requirements for the Degree of Doctor of Philosophy in Computer Science at Lincoln University

Abstract

Managing complexity is a key challenge in systems biology modelling. While detailed models can provide valuable insights into specific biological processes, they can become computationally intensive and challenging to interpret. Conversely, overly simplistic models may lack accuracy and fail to capture essential aspects of the system's behaviour. Hence, this study focuses on reducing the complexity of the models in systems biology while enhancing accuracy, efficiency, and interpretability. This study uses deep learning (DL), a sub field of machine learning (ML) and artificial intelligence (AI), along with sensitivity analysis using partial rank correlation coefficient (PRCC), and extended Fourier amplitude sensitivity test (eFAST) methods to identify significant system behaviour. DL models can analyse and predict system behaviour using large datasets consisting of biological measurements. They can also handle various omics data and integrate them into comprehensive systems biology models. Key objectives of the study include reducing the complexity of computational models using DL methods, identifying significant subprocesses to reduce pathway diagrams, improving the interpretability of the reduced models using sensitivity analysis, implementing models with minimum knowledge in the parameter space, introducing a framework to automate the complexity reduction process using sensitivity analysis and DL, improving model performance with ensemble learning and automated hyperparameter tuning, and testing the reproducibility of the reduced models using simulations. The significance and contribution of the study lie in its automated approach to developing an interpretable meta-model using AI/ML/DL techniques and reducing complexity while maintaining model accuracy. The study tests neurobiology models and uses sensitivity analysis and DL for complexity reduction. The thesis provides reduced pathway diagrams for each of the models used for testing. The study offers a framework to automate the implementation of meta-models and provides insights into input-output relationships and key processes in biological systems. The methodological approach involves using DL/ML/AI methods, ensemble methods, and dimensionality reduction to reduce complexity in VCell models. The framework reads MATLAB files from VCell, identifies parameters and outputs, performs parameter perturbations and simulations, conducts sensitivity analysis, identifies significant reactions, and trains and validates machine learning models based on the results. The thesis builds on previous efforts to develop more interpretable and reliable computational models that can provide meaningful insights into complex biological systems.