Decoding computational neuroscience: accessible frameworks for sophisticated neural analyses
OA Version
Citation
Abstract
Advancements in neural recording technologies have led to an explosion in the scale and complexity of neuroscientific datasets, demanding analytical frameworks that are not only computationally powerful, but also interpretable and accessible. This dissertation introduces a cohesive suite of computational methodologies, validations, and tools to address these demands, focusing on the application of both unsupervised and supervised machine learning techniques to extract meaningful neural dynamics and predict emergent network behavior, or clinically significant events, from a high volume of recordings. We first focus on the application and validation of dimensionality reduction (DR), rigorously evaluating Nonnegative Matrix Factorization (NMF) for analyzing state-dependent neuronal network dynamics in calcium imaging data. We demonstrate its superiority over commonly used DR methods, specifically PCA, ICA, and UMAP, in recovering biologically plausible, interpretable subnetwork activity.
Building on this validation, we present CaNetiCs, an open-source software toolbox that integrates standardized implementations of NMF and complementary DR methods, alongside modules for geometric low-dimensional space modeling, and synthetic neuronal simulations. We apply this framework to datasets from C. elegans and murine recordings under graded anesthesia, recapitulating known physiological trends while uncovering novel insights into neural population dynamics and state transitions.
Finally, we extend this analytical ideology to the supervised learning domain by introducing a data augmentation strategy for enhancing epileptic spike ripple detection in human electrophysiological recordings. Employing LSTM-based neural networks, we show that training on a combination of empirically grounded synthetic and in vivo data improves classifier performance, offering a scalable solution to training-data scarcity in clinical neurodiagnostics.
Together, these contributions provide accessible, interpretable, and rigorously validated frameworks for dimensionality reduction and machine learning in neuroscience, advancing both understanding and translational utility in the analysis of complex neural systems. By integrating biologically grounded decomposition methods, open-source analytical software, and data-driven neural network architectures, this work bridges methodological innovation with practical implementation. The tools and techniques developed herein not only facilitate the discovery of latent structure in large-scale neuronal recordings but also enable the automated identification of clinically relevant biomarkers. In doing so, this dissertation empowers researchers to more effectively characterize dynamic brain activity across species, brain states, and experimental paradigms, adding to a computational foundation for reproducible, scalable, and hypothesis-driven neuroscience.
Description
2025