Computational Approaches to Bridge Experimental and Simulation Viewpoints in Neuroscience

Yves Bernaerts

This repo contains raw data, preprocessed data, code and figures belonging to my PhD thesis. Full thesis is published with the public library of the university of Tübingen and can be downloaded here. It is based on the kaobook template.

Papers

This thesis capitulates a couple of works, published and unpublished, and with much of the code available on GitHub.

1. Most notably for this thesis: Hybrid statistical-mechanistic modeling links ion channel genes to physiology of cortical neuron types which can be found on BioArxiv and with code here.
2. Sparse bottleneck neural networks for exploratory non-linear visualization of Patch-seq data that can be found on ArXiv with code here.
3. Sparse reduced-rank regression for exploratory visualization of paired multivariate data that can be found here with code here.
4. Phenotypic variation of transcriptomic cell types in mouse motor cortex that can found in Nature with code here.
5. Efficient identification of informative features in simulation-based inference that can found as a NeurIps conference paper and with code here.

Data

Some raw electrophysiological recordings in .nwb format can be found in data/raw_data/. They serve as exemplary cells to illustrate some of the results. All raw electrophysiological recordings that have been used in multiple studies that make up the bulk of this thesis can be found on DANDI. Please follow the instructions there to download all the raw data. You might need it depending on which notebook you would want to turn, but the few raw data examples given could be sufficient too.
Most of the preprocessed data is in data/. Yet some files are too big for upload on GitHub. These include:

1. full_batch.npz that can be downloaded from Zenodo.
2. SmartSeq_cells_AIBS.pickle that also can be downloaded from Zenodo. Please put those 2 files in data/ if you consider to work with the code.

Code

Small analyses and figures have been produced with code in the main directory.

1. ephys_extraction.ipynb shows how to use the automated electrophysiological (ephys) feature extraction pipeline. It also produces a sketch illustrating its functionalities. Also check ephys_extractor.py and ephys_features.py, code worked with more extensively here and based on the Allen SDK.
2. hodgkin-huxley.ipynb illustrates the Hodgkin-Huxley-based model, implemented with the Brian2 simulator. Also check simulator.py, ephys_extractor_hh.py and ephys_features_hh.py for the implemented models.
3. sbnn.ipynb produces figures illustrating results from using sparse bottleneck neural networks or sBNNs for the nonlinear prediction of ephys measurements based on gene expression levels. Check bottleneck_neural_networks.py for TensorFlow2 model implementations and check out this GitHub repo to see how we run these models on real data.
4. npe-n.ipynb introduces Neural Posterior Estimation with Noise or NPE-N and shows it performs better on real-world experimental data, such as electrophysiological recordings, than its baseline counterpart NPE. We produce figures showing this for very specific cells. For the inference results used here, we capitalize on the simulation-based inference package sbi (also check Usage below).
5. bridge_genotype_to_phenotype.ipynb shows how sparse reduced-rank regression can be utilized to bridge the genotype of a neuron --- in the form of its gene expression levels --- to Hodgkin-Huxley model parameters, and specificially those that can generate simulations that reproduce the raw ephys recording, both qualitately and quantitatively.
6. bridge_and_embed.ipynb produces embeddings showing average performance of applying NPE-N to a full Patch-seq data set.

Regarding 4-6 it definitely makes sense to check out [this](https://github.com/berenslab/hh_sbi) GitHub repo.

Figures

In figures/, you can find most figures (not including sketches) that are included in the thesis.

Usage

Install the following dependencies before running any notebook and you should be good to go:
pip install -r requirements.txt

Supervision

A great deal of gratitude goes to Prof. Dr. Philipp Berens ([email protected]) and Dr. Dmitry Kobak ([email protected]) for their continued supervision and support.

Contact

Contact me at [email protected]