A Unified Software-Generating Framework for Biological Data Analysis

Abstract

In recent decades, biological data analysis has increasingly relied on software. For example, complex analyses of the human brain or plant data have often been conducted with newly developed toolboxes, workflow engines, and graphical platforms. As the software ecosystem grows, however, it becomes harder to keep code, interfaces, and tests in step, and to reuse methods across studies, modalities, and domains without rewriting large parts of the stack.

This thesis proposes a software-generating framework Genesis, which is implemented within an in-house-developed software ecosystem BRAPH 2 (BRain Analysis using graPH theory, second edition). The framework has two stages. First, each analysis component is written once as a human-readable .gen.m description that combines MATLAB code with structured declarations of its inputs, parameters, documentation, and a graphical interface. Second, a compiler turns these descriptions into executable modules, graphical interfaces, unit tests, and complete, shareable software distributions. The core idea of Genesis is to shift effort from crafting one-off tools and graphical user interfaces to maintaining a centrally defined library of descriptions. Most importantly, these centrally defined descriptions can be recombined and regenerated to investigate a different research question, while the code, interfaces, and tests remain in sync.

This thesis presents four studies to demonstrate Genesis’s feasibility across different biological analyses. BRAPH 2, a distribution for network neuroscience, provides the initial infrastructure for graphical interfaces, pipeline orchestration, and core statistical and neural-network elements. Using the same .gen.m description language and compiler, GapNet extends this stack with an alternative training approach for incomplete neuroimaging cohorts. The human bone-marrow light sheet microscopy study adds graph-based microenvironment descriptors and variational autoencoder elements for unsupervised niche discovery. Finally, the plant Raman spectroscopy study reuses the same autoencoder machinery, with minimal spectral preprocessing, to characterise stress responses across species and stressors.

Taken together, these studies show that a single description-centred framework can span imaging and spectroscopy, minimise boilerplate, and make complex pipelines more portable, auditable, and extensible. By treating code, interfaces, and tests as generated views of shared .gen.m descriptions, Genesis supports more transparent and reproducible development of analysis software in computational biology.