Computational Approaches to Bridge Experimental and Simulation Viewpoints in Neuroscience

Abstract

The nervous system constitutes the highway of communication in living animals. Information propagates in both bottom-up fashion, that is from the periphery of the body to the brain, and top-down, from the brain back to the body. It is the nervous system's foundational units, the neurons, that serve as computational nodes in this bidirectional sophisticated design.

Neurons reside in biological organisms, yet the nature of their electrical responses are reminiscent of systems in engineering and physics, their communication through neurotransmitters of diffusion models in chemistry, and their interconnectivity patterns of neural networks in machine learning. The field of neuroscience develops increasingly interdisciplinary, yet we still lack methods that meaningfully bridge them.

To appreciate the importance of a bridge between different viewpoints in neuroscience, we can be inspired by the `cortical tree of neural types`. Even though the notion of neural types in neuroscience is fairly established, neurons have been observed to vary both discretely and continuously in both genetic and phenotypic landscapes, making their classification in `types`, as discrete entities conforming to both modalities, difficult.

Statistical models that bridge the genetic landscape to phenotypic modalities including electrophysiology and morphology could help us build a multimodal neuronal taxonomy. Yet, even though many exist for the unidimensional (`one-view`) analysis of cells in the nervous system, we currently still lack `joint-view` statistical tools that can for instance predict one modality from another or produce joint-view two-dimensional embeddings for the exploratory analysis of neuronal data sets.

Biophysical models move one step further. They do not merely describe the quantitative relationship between modalities, but allow for a peek into the black-box transformation from one to the other. They can describe \emph{how} the electrophysiology comes to be in neurons based on ion channel abundance in the cell's membrane, the latter determined by the expression of specific genes. Reliably inferring probability distributions over parameters of biophysical models that reproduce real-world observations in electrophysiology, has however also proved challenging. While recent advances in inference methods have demonstratively shown useful for synthetic simulated data, they remain difficult to generalize to experimental data.

In this thesis, we show how recent advances in machine learning can bring both statistical as well as biophysical models, to the next level. We present sparse bottleneck neural networks that select specific genes to nonlinearly predict electrophysiological measurements and outperform linear methods on the prediction task. They furthermore produce two-dimensional joint-view embeddings that are directly interpretable in biology.

We furthermore introduce a slight manipulation in training strategy for neural density estimators, causing the inference of biophysical model parameters to be much more reliable when it is applied to real-world electrophysiological data.

Finally, as a first attempt to bridge the genetic landscape with electrophysiological behavior in neurons, we show how sparse reduced-rank regression intuitively selects specific genes to predict fitted parameter values of biophysical models that replicate real-world neuronal electrophysiology.

Die Dissertation ist gesperrt bis zum 20. Dezember 2024 !