70040 research outputs found
Sort by
Modelling Environmental Influences on Cell Phenotype with Mechanistic and Machine Learning Approaches
No full textMicroenvironmental conditions are critical in shaping cell phenotype. Recent advances in two modalities, single-cell transcriptomics and live imaging, have revealed that environmental conditions influence both gene expression and cellular morphology, respectively. In this PhD thesis, I develop and describe two computational models targeting each modality, both of which evaluate the extent to which environmental factors govern cell morphology and transcriptomic states. The two models, described in Chapters 3 and 4 respectively, differ in the level of embedded biological assumptions, which allows for both hypothesis-driven and data-driven exploration of environmental influences on cell phenotype.
The first model consists of an agent-based mechanistic model that examines how local environmental conditions influences the morpho-spatial phenotype of tingible body macrophages (TBMs) during the germinal centre reaction in mice. Subsequent in silico perturbation of various parameters reveals that the TBM choice of morpho-spatial phenotype maximizes the clearance rate for the local density and motility of apoptotic fragments.
The second model consists of an agnostic machine learning model, called ENTRAIN, which prioritizes ligand-receptor pairs in cellular differentiation. By integrating trajectory analysis with ligand-receptor-transcription factor networks, the first known method to do this, ENTRAIN identifies ligand-receptor pairs most likely to be influencing cell differentiation trajectories. ENTRAIN’s results are strongly concordant with published literature yet orthogonal to existing methods, demonstrating that the recovered signals are overlooked by traditional cell-cell communication methods. In contrast to the hypothesis-driven model in chapter 3, ENTRAIN operates without relying on strong prior assumptions.
Collectively, the two models illustrate the versatility of computational methods across contexts with varying levels of biological assumptions. While Chapter 3’s method embeds numerous mechanistic assumptions, Chapter 4 is a data-first method, operating in the absence of predefined rules. I then conclude this thesis in Chapter 5 with a discussion on a logical extension of this work: developing models that predict transcriptomic states in response to environmental perturbations. Such models have potential applications in combinatorial drug-response prediction, cell-state engineering, and whole-cell modelling
Engaging in a community of practice in visual arts: A systematic literature review
No full textThis systematic literature review explored the potential of collaborative art investigations of
digital artworks in schools to engage disengaged students and foster a sense of connectedness and
belonging within a community of practice (CoP). The review also focused on the barriers and
enablers of effective CoPs in digital creative practice and the effects on students’ outcomes. The
review comprised a systematic search of peer-reviewed journals in ProQuest Education, ProQuest
Design and the Arts, Taylor and Francis, ERIC, and Arts and Architecture Complete databases
between 2012 and 2022. PRISMA guidelines were used to locate and select 13 articles that met
the inclusion criteria. A social realist perspective was used as a lens to guide the analysis of the
review. Results indicated that collaborative art investigations in digital artmaking including animation
offer valuable opportunities for enhancing student engagement, promoting a sense of community, and nurturing a positive learning environment
Distributed Optimization in Networked Systems
No full textOptimization is a research area with a wide range of applications in mathematics, engineering, geophysics, economics, and finance. Among various optimization problems, network optimization has emerged as a significant area of research, concentrating on optimizing networked systems such as social, economic, computer, and power networks. This problem has been a flourishing area of research in the past two decades.
Traditional centralized optimization approaches, where all problem data are processed by a central entity, become impractical for large-scale systems due to computational and communication constraints. Distributed optimization addresses these challenges by enabling multiple agents to collaboratively solve an optimization problem while sharing limited information locally. The inherent advantages of distributed optimization include improved scalability, resilience to single-point failures, and reduced communication and computational overhead. As a result, distributed optimization has gained significant interest, particularly in applications such as sensor fusion, robotic coordination, smart grids, and deep learning. However, substantial challenges remain, particularly in handling nonconvex optimization problems, time-varying constraints, and disturbances in dynamic environments.
This thesis develops robust and scalable distributed optimization frameworks tailored for three key application domains: pose graph optimization (PGO) problem, deep neural network training, and constrained time-varying optimization for first order and second-order systems.
First, a distributed optimization framework is introduced for solving PGO, a fundamental problem in robotics and sensor fusion. Unlike conventional convex relaxation techniques, the proposed approach directly addresses the nonconvex nature of the problem while preserving the underlying geometric constraints. A novel Split Orthogonality Constraint (SOC) method is developed by integrating the Alternating Direction Method of Multipliers (ADMM) with split Bregman iterations, ensuring efficient and accurate pose estimation. Furthermore, an Extended Proximal Alternating Linearized Minimization with Augmented Lagrangian (EPALMAL) approach is introduced to improve convergence speed and scalability in large-scale PGO problems. Numerical simulations validate the effectiveness of these methods, achieving near-global optimality in distributed settings.
Following this, an accelerated ADMM-based framework is developed for deep neural networks, offering an alternative to traditional Stochastic Gradient Descent (SGD) methods. While SGD and its variants are widely used for deep learning, they suffer from slow convergence, sensitivity to poor conditioning, and high variance in gradient updates. To address these inefficiencies, Anderson Acceleration is integrated into ADMM to enhance convergence speed while maintaining robustness. Additionally, matrix inversion is replaced with backtracking and quadratic estimation, significantly reducing computational complexity. Theoretical analysis provides convergence guarantees, and extensive numerical experiments on benchmark datasets demonstrate superior performance compared to existing ADMM and SGD-based techniques.
Finally, a continuous-time distributed optimization framework is proposed for multi-agent systems with time-varying cost functions and constraints. The approach extends previous work by incorporating both nonlinear inequality and linear equality constraints without imposing restrictive assumptions on Hessian matrices. To enhance robustness, an integral sliding mode control strategy is integrated, ensuring stability in the presence of both bounded disturbances and disturbances with bounded derivatives. Theoretical analysis using Lyapunov functions and nonsmooth analysis establishes convergence guarantees. Numerical simulations confirm the framework's effectiveness in handling dynamic constraints and disturbances in second-order systems
GDP-B: Accounting for the Value of New and Free Goods
No full textThe welfare contributions of new goods and free goods are not well-measured in standard statistical agency metrics like GDP or productivity. We derive explicit terms for the contributions of these goods and introduce a new framework and metric, GDP-B, which quantifies their benefits. We apply this framework to several empirical examples, including Facebook and smartphone cameras, and estimate their valuations through incentive-compatible choice experiments. Our new approach can help measure welfare changes over time and reveal which goods and innovations contribute the most to economic growth and well-being. (JEL D13, E01, E23, O30
An Immersed Boundary-Lattice Boltzmann Method for Fluid-Structure Interactions Involving Compressible Flows
No full textThis thesis develops an efficient numerical method using the standard lattice Boltzmann method (LBM) for studying fluid-structure interaction (FSI) problems in high-speed flows. The LBM is chosen due to its easy-to-use algorithm and inherent parallel nature. Since the standard lattice models lack the moments required to fully recover the compressible Navier-Stokes equation (NSE), a correction body force is applied numerically to account for highly compressible flows. In this work, a two-dimensional recursive regularised (RR) collision model is used to solve the mass and momentum equations for compressible flows. The finite difference method (FDM) for solving the energy equation in its entropy form is employed due to its simplicity and high computational efficiency. The LBM and FDM are coupled using the equation of state. The Newmark-β time integration is used to update the position of the structure. Additionally, the iterative feedback immersed-boundary method (IBM) is utilised, providing an efficient approach for handling the boundary conditions for moving geometries.
The following validations are conducted at various Mach numbers (Ma) and Reynolds numbers (Re) to assess the accuracy and robustness of the solver: weakly compressible flow (Ma=0.1) over a cylinder at Re=40; subsonic flow over a cylinder at Re = 100 and 0.2 ≤ Ma ≤ 0.5; transonic flow over a NACA0012 aerofoil at Re = 500 and Ma=0.8 and 0.9; and supersonic flow over a cylinder at Re = 300 and 1.2 ≤ Ma ≤ 1.5. Finally, to validate the solver for moving bodies, the flow-induced vibration (FIV) of an elastically mounted cylinder is conducted at 70 ≤ Re ≤ 150 for Ma=0.1 and 0.5. The results produced by the current solver agree well with the data reported in previous references across all tested Mach number ranges. Therefore, the hybrid IB-LB-FD solver can be used to study FSI problems in both subsonic and supersonic flows with satisfactory accuracy.
To demonstrate the versatility of the developed solver, the FIV of an elastically mounted cylinder in compressible flows is investigated. The simulations are conducted for a range of parameters including mass ratios (m*) between 10 and 100, Reynolds number in the range 70 ≤ Re ≤ 300, reduced natural frequency (FN) between 0.11 and 0.25, and the Mach number in the range 0.1 ≤ Ma ≤ 1.5. First, the vibration amplitude decreases as Ma increases and eventually becomes zero in supersonic flows at low mass ratios. For the first time, a heavy cylinder with a 2S mode of vortex shedding is observed at Ma=1.4 and Re=300 with an initial excitation. The results suggest that the mass ratio plays an important role, and at m* ≥ 81.625, the self-sustained vibration of a cylinder is observed. Spectral characteristics and flow fields confirm that the interaction between the cylinder and its wake is driven by a shock-induced feedback mechanism. Second, in the range 70 ≤ Re ≤ 150, the FIV properties are similar to incompressible flows at Ma<0.3. A linear reduction in both the vibration amplitude and the synchronisation range is observed as the Ma increases up to 0.7. Beyond that, a more prominent λ shock appears intermittently, leading to significant reduction in the vibration amplitude. Finally, regardless of reduced frequency variations, the compressibility tends to stabilise the system by reducing the vibration as the Ma increases at m*=10
Daylight Photoluminescence Imaging – Quantifying Power Losses Caused by Series Resistance Defects in Solar Modules
No full textFast and accurate performance analysis of fielded solar modules is essential for the reliable, long-term operation of large-scale solar farms. Daylight photoluminescence imaging has emerged as a promising inspection method, providing quantitative information while circumventing many logistical constraints associated with alternative methods. Luminescence images of modules acquired with partial current extraction reveal series resistance defects, a key contributor to cell and module degradation. In this thesis, a novel method is presented to estimate the reduction in output power caused by series resistance defects, based purely on daylight photoluminescence image data. This automated process generates electrical models to match series resistance related intensity variations observed in daylight photoluminescence images, which are used to quantify performance losses. Cell-level simulations and experimental results are presented, yielding excellent results. Promising proof-of-concept demonstrations on full modules show that the method has significant potential for routine application on operational solar farms
Propolis compound inhibits profibrotic TGF-β1/SMAD signalling in human fibroblasts
No full textHypertrophic scarring of the skin is a cause of pain, disfigurement, and restricted mobility. Excessive TGF-β1 signalling leads to SMAD3 phosphorylation, which is implicated in hypertrophic scarring. In this study, we examined the mechanism of action of tomentosenol A, a small compound that we isolated from the propolis of the Australian stingless bee Tetragonula carbonaria. Cultured adult human dermal fibroblasts and HEK293 cells were stimulated with TGF-β1, with or without tomentosenol A, and were assessed for phosphorylation of SMADs 2/3 (Western blot, AlphaLISA assay), SMAD signalling (HEK293 cells expressing a SMAD3 reporter gene), and profibrotic gene transcription using RTqPCR for ACTA2 (smooth muscle α-actin), COL1A1 and COL3A (collagens), CCN2 (connective tissue growth factor) and FN1 (fibronectin). Protein expression was measured using ELISA (fibronectin) and visualised via confocal microscopy (smooth muscle α-actin). TGF-β1 increased SMAD3 phosphorylation by 44.3-fold above baseline levels, and this effect was inhibited by tomentosenol A in a concentration-dependent manner (IC50, 99.0 nM). TGF-β1 stimulated SMAD3 reporter gene expression and upregulated ACTA2, COL1A1, COL3A1, FN1 and CCN2 transcription; fibronectin protein expression; and smooth muscle α-actin filament formation in fibroblasts. These responses were inhibited by 6.25 μM tomentosenol A. These findings indicate that tomentosenol A inhibits TGF-β1/SMAD3 signalling and downstream profibrotic gene transcription and protein expression. As this pathway is implicated in hypertrophic scarring of the skin, tomentosenol A can be developed as a novel therapy for the management of scars caused by deep dermal injuries that are associated with surgery, trauma and burns
Advanced Carbon-Based Electrocatalysts for Li-CO2, Li-O2, and Li-N2 Batteries
No full textMetal-gas batteries have been recognized as an advanced energy storage technology, offering superior capacity and high energy density. Among them, Li-gas batteries (LGBs), which utilize metallic lithium anodes and gaseous cathodes, exhibit exceptionally high theoretical energy density. Notably, Li-O2 batteries (LOBs) feature a theoretical energy density of 3,460 Wh kg-1, approximately ten times that of typical Li-ion batteries. However, their practical application remains significantly constrained by sluggish gas conversion kinetics, leading to high overpotentials, poor rate performance, and limited electrochemical reversibility, which severely affect cycling stability. Overcoming these limitations necessitates the innovation of bifunctional electrocatalysts that enhance both discharge and charge reactions while promoting efficient gas conversion.
Carbon-based metal-free electrocatalysts (C-MFECs) and carbon-supported single-atom catalysts (CS-SACs) are increasingly recognized as viable options for LGBs, benefiting from their high catalytic efficiency, tuneable structures, and remarkable stability, paving the way for improved electrochemical performance. These catalysts exhibit high efficiency in both discharge and charge processes, ensuring effective activation of reactants and decomposition of products.
The review (Chapter I) provides a comprehensive analysis of LGB systems, identifying the critical challenges limiting gas conversion reversibility and elucidating the bifunctional catalytic mechanisms of C-MFECs and CS-SACs. Furthermore, we highlight recent advancements in catalyst development for LCOBs, LOBs, and LNBs, with a particular focus on strategies to enhance catalytic efficiency and electrochemical stability. This thesis also addresses these challenges by developing advanced C-MFECs and CS-SACs cathodes to improve the kinetics of key reactions in LGBs. An integrated approach involving advanced characterization methods and theoretical modelling is utilized to uncover the catalytic mechanisms of these newly developed catalysts.
Experimentally, a metal-free, dopant-free, and carbon-based electrocatalyst was synthesized by combining single-walled carbon nanotubes (SWCNTs) with fullerene (C60). This SWCNTs/C60 hybrid catalyst, when used in Li-CO2 batteries (LCOBs), delivers an exceptional discharge capacity of 56,729 mA h g-1 at 0.5 A g-1, sustains 266 cycles at 1.0 A g-1, and maintains a low overpotential of 1.13V at 100 mA g-1, outperforming most previously reported counterparts. Density functional theory (DFT) calculations demonstrate that the charge transfer occurring between SWCNTs and C60 causes uneven charge redistribution, creating highly accessible catalytic active sites that enhance Li+/CO2 adsorption and facilitate product decomposition. These results establish a basis for optimizing the design of C-MFECs for LCOBs.
In LOBs, the plentiful adsorption sites and large surface area promote Li+/O2 diffusion, leading to the generation of extensive amorphous Li2O2 films, achieving to a capacity of 44,602 mAh g-1 at 0.5 A g-1 with a low potential gap of 0.78 V and prolonged cycling stability for 234 cycles at 0.5 A g-1. DFT calculations confirm that C60 incorporation enhances intermediate adsorption, stabilizing reaction pathways during oxygen reduction/evolution reaction (ORR/OER). This study establishes a strategic approach for developing efficient metal-free electrocatalysts for LOBs, offering broader insights into advancing next-generation energy storage technologies.
An axial Cl-coordinated Ni single-atom catalyst (Cl-Ni SAs-NC) is introduced as high-performance electrocatalyst for the cathode in Li-N2 batteries (LNBs). Trace H2O acts as a proton source, facilitating the discharge reaction: 6Li + 6H2O + N2 → 6LiOH + 2NH3 with theoretical potential of 2.24 V. Experimentally, a 1.69 V discharge plateau is observed, surpassing the 0.44 V reported for the reaction 6Li + N2 → 2Li3N. The Cl-Ni SAs-NC delivers high discharge/charge capacities (4.22/2.93 mAh cm-2 at 0.05 mA cm-2) and cycling life for 131 cycles at 0.05 mA cm-2). It achieves an NH3 yield rate of 1.15 μg h-1 cm-2 at 0.05 mA cm-2 with 12.28% Faradaic efficiency, while LiOH decomposition enables lithium recovery. DFT calculations confirm that pre-lithiation lowers reaction barriers and Cl coordination enhances nitrogen reduction reaction (NRR) kinetics. This study provides a new approach to nitrogen fixation in LNB systems, broadening the potential of LGB technology.
In summary, the successful application of SWCNTs/C60 hybrid and Cl-Ni SAs-NC as bifunctional electrocatalysts in LGBs underscores their potential for next-generation energy storage. These materials provide valuable design principles for metal-gas batteries, offering a framework for C-MFECs and CS-SACs catalyst development and performance enhancement. However, despite their remarkable bifunctional activity, ensuring long-term stability without compromising catalytic efficiency remains a major challenge. The practical deployment of these catalysts in commercial-scale LGBs is still in its early stages, necessitating further advancements in catalyst durability, electrode architecture, and electrolyte optimization to improve efficiency and cycling stability.
To transition from fundamental research to practical applications, future efforts should integrate computational modelling with advanced in situ characterization to elucidate reaction mechanisms and guide rational catalyst design. Additionally, structural engineering and material optimization will be crucial in achieving a balance between catalytic activity and stability, ensuring their viability for large-scale implementation. Addressing these challenges will accelerate the commercialization of LGBs, positioning them as scalable, high-performance energy storage systems capable of supporting sustainable energy solutions
Enhancing Quantification and Classification for Multi-Class Learning Under Label Shift
No full textWhile the independent and identically distributed (i.i.d.) assumption is common in machine learning, real-world applications often involve data distribution shifts. This work focuses on label shifts, where class prevalences change significantly between training and test time.
This thesis makes two main scientific contributions. First, it introduces a new class of ensemble-based methods for multi-class quantification, which estimates class prevalence in a sample. Quantification is particularly relevant in applications focused on groups rather than individuals, such as sentiment analysis, epidemiology, and ecological surveillance.
Most existing methods rely on one-versus-all (OVA) approaches to handle multi-class problems, but these perform poorly in such settings. This thesis identifies OVA’s fundamental shortcomings and proposes MC-SQ and MC-MQ, two ensemble-based quantifiers that directly model multi-class prevalence shifts. Extensive experiments validate their superior performance, including top-ranking results in a recent quantification competition.
The second contribution addresses classification under label prevalence shift as the next step after quantification. While quantifiers estimate class prevalences in the test distribution, integrating these estimates into classification remains a challenge. Existing methods fail to enforce the predicted prevalences on adjusted classifiers, limiting their effectiveness. To solve this, we introduce \match, a novel adjustment method that optimally aligns predicted probabilities with quantifier-estimated class prevalences. Unlike instance re-weighting or Bayes updates, \match formulates adjustment as a linear optimisation problem, ensuring that classification fully incorporates the prevalence information from quantification. This approach achieves state-of-the-art performance, demonstrating the importance of incorporating quantification predictions directly into classification.
Together, these contributions advance quantification and classification under class prevalence shifts, providing more accurate solutions for multi-class settings
A Comprehensive Investigation of the Impact of Copy Number Variants on Neurocognitive Disorders
No full textNeurocognitive disorders (NDs), including Autism Spectrum Disorder (ASD), Intellectual Disability (ID), Developmental Delay (DD), and Schizophrenia (SCZ), are early-onset conditions that pose significant challenges in diagnostics and treatment. Genetic factors account for approximately 50% to 80% of ND etiologies, with Copy Number Variations (CNVs) identified through Whole Genome Sequencing (WGS) being a predominant causative element. Establishing clear genotype-phenotype associations is crucial but challenging due to the difficulty in precisely linking specific CNV loci to corresponding disease phenotypes. This thesis aims to provide a comprehensive landscape of NDs by identifying novel phenotypic-genotypic connections, particularly focusing on the role of CNVs.
We identified 47 significant CNV regions from 6,479 control samples and 19,663 ASD samples, characterising commonly deleted or duplicated genomic regions by identifying key genes involved. By analysing case-specific CNVs from 40,000 healthy individuals and 75,000 individuals with ND phenotypes categorised into ASD, ASD with ID (ASD_ID), ID with DD (ID_DD), and SCZ groups—we found 9 CNV regions significantly associated with ASD, 69 with ASD_ID, 137 specific to ID_DD, and 15 linked to SCZ. Within these CNV regions, we identified both common and unique Gene Ontology (GO) terms and
biological pathways, such as the association of 22q11.2 deletion syndrome with Prader-Willi and Angelman syndromes across multiple ND phenotypic groups. Tissue enrichment analysis, conducted by mapping to FANTOM5 CAGE gene expression profiles, revealed that ND-specific genes are predominantly expressed in the brain, immune system, T-cells, reproductive systems, and liver. Cell-to-cell communication analysis using single-cell data from ND patients identified important ligand-receptor pairs, including SLURP2-CHRNA7 and RTN4R-ADGRB1, which are enriched across different ND categories.
Additionally, we investigated CNVs occurring within regulatory regions. Utilising Hi-C data, we examined how CNVs can affect gene dosage through chromatin interactions and identified two non-coding CNV regions that potentially disrupt gene expression via promoter-enhancer loops.
Focusing on ASD, we analysed a cohort of 437 autistic families, including 529 autistic children sourced from the Autism CRC Biobank. Applying criteria of less than 1% population frequency and less than 0.05% of 6,453 healthy control CNVs, and observed in the DECIPHER rare CNVs, we identified eight de novo protein-coding genes. Four of these genes are absent in both parents and healthy siblings and feature novel genes GAS2,
FARP2, and RGS3, which may contribute to ASD phenotypes.
This thesis underscores the genetic heterogeneity of NDs and illuminates potential biomarkers for diagnosis and targeted therapeutic strategies, including personalised treatment options. This research significantly contributes to a more comprehensive understanding of how CNVs, within both coding and non-coding regions, impact the etiology of NDs