Automatic discovery of dynamic regulatory mechanisms from spatial and temporal information

Abstract

This dissertation addresses the critical challenge in biology and medicine of extracting mechanistic insights from temporal and spatial data. Overcoming this challenge is essential for advancing many scientific and medical fields, including developmental biology, synthetic biology, stem cell research, and clinical practice. Crucially, the inference of the mechanisms underlying spatial and temporal phenomena, such as pattern formation, cellular differentiation, and disease dynamics, is hindered by the complexity of nonlinear interactions and feedback loops within biological processes. To address this fundamental problem, this thesis presents innovative methodologies for inferring interpretable mechanistic models that can provide a deeper understanding of biological processes and disease progression. A major contribution of this work is a novel high-performance machine-learning methodology for inferring regulatory and signaling mechanisms capable of generating specific spatial patterns in two-dimensional domains. Leveraging heuristic machine-learning algorithms, the method identifies network structures and interaction mechanisms as well as the optimal number of necessary genes. High-performance computing with graphical processing units accelerates simulations, addressing the computational demands of evaluating numerous candidate models. Another significant advancement developed in this thesis is a computational framework for automatically designing synthetic gene regulatory mechanisms (GRMs) that can generate any target spatial pattern using orthogonal morphogen gradients. Leveraging evolutionary computation and high-performance computing, the framework accelerates GRM design. It features a versatile mathematical modeling approach for nonlinear regulatory interactions, an innovative error calculation using box blur kernel convolution, and adjustable algorithmic tolerances to create patterns ranging from simple shapes to complex biological structures. This dissertation also contributes to the advancement of clinical research by introducing an explainable machine-learning method to infer predictive models from temporal datasets of patients with acute myeloid leukemia (AML). Using these heterogenous patient data, such as clinical information, genetic and cytogenetic abnormalities, treatments, and disease progression markers, the method can produce mathematical models capable of capturing disease dynamics, identifying key signaling mechanisms, and offering insights into AML progression and treatment outcomes. Finally, this work advances bioinformatics approaches to analyze dynamic single-cell datasets from flow cytometry and transcriptomics experiments to study human hematopoiesis. Using in vitro cultures of human mobilized CD34+ hematopoietic stem and progenitor cells (HSPCs) under intrinsic and extrinsic conditions, the pipeline integrates data filtration, normalization, dimension reduction, clustering, cell annotation, trajectory analysis, and cell velocity analysis. This comprehensive approach enhances understanding of HSPC biology and the factors that influence their differentiation.