Periodic Reporting for period 1 - SICNET (Statistical Inference of the Cerebellar Network)
Berichtszeitraum: 2020-04-01 bis 2022-03-31
Several actions we make in our every-day life, from speaking to playing musical instruments, require timed execution. The stricking accuracy of skilled movements raises questions about how neural circuits can achieve precise behavioral control.
We are at the crest of technological advances where we can directly test hypotheses and models about neural function by monitoring the activity of large-scale neuronal populations. However, in order to make sense of such complex data, we need statistical methodologies able to identify patterns of neuronal activity with high temporal resolution. The scope of this action was to leverage recent methodological advances in statistics to develop algorithms able to infer activity patterns from calcium imaging data, determine network motifs and use information theory to understand their functional role.
Understanding the mechanisms involved in neural computation is challenging, however by studying well characterized brain areas such as the cerebellar cortex we can hope to shed lights onto the underlying principles of how the brain encode temporally structured sensory information.
We are at the crest of technological advances where we can directly test hypotheses and models about neural function by monitoring the activity of large-scale neuronal populations. However, in order to make sense of such complex data, we need statistical methodologies able to identify patterns of neuronal activity with high temporal resolution. The scope of this action was to leverage recent methodological advances in statistics to develop algorithms able to infer activity patterns from calcium imaging data, determine network motifs and use information theory to understand their functional role.
Understanding the mechanisms involved in neural computation is challenging, however by studying well characterized brain areas such as the cerebellar cortex we can hope to shed lights onto the underlying principles of how the brain encode temporally structured sensory information.
Work Package 1: inference of firing rates from functional imaging in the cerebellum:
The initial six months period of the action has been devoted to the development of analysis pipelines and software to visualize and process fluorescence imaging data of the cerebellar cortex alongside behavioral recordings. I have developed:
1. a Graph-Based SEGmentation (GBSEG) method to extract region of interests (ROI) from fluorescence recordings. The software implementing this algorithm is now an established tool in the lab used by PhD students and postdocs to analyze their data.
2. A graphical user interface to visualize imaging and behavioral data acquired in the lab. By enabling simultaneous inspection of imaging and behavior we could formulate hypotheses on how neuronal activity drives behavior.
3. A graphical interface to curate and manage imaging dataset. This tool enabled to standardize the data organization, to facilitate the systematic analysis of the data.
To extract firing patterns of cerebellar granule cells I developed a Bayesian inference method to estimate spike times from calcium imaging data. Based on previous works I have developed a generative model of fluorescence time series accounting for baseline fluorescence modulation and bursting firing statistics. Then, as described in the DoA, we employed novel sequential Monte Carlo algorithms to estimate static and dynamic model parameters. Our method allows us to quantify the statistical uncertainties associated to these estimates.
Our inference method allows us to resolve inter-spike intervals below five milliseconds, enabling the analysis of high firing neurons.
To accelerate the diffusion of our method we posted our results as a preprint on bioRxiv (G. Diana, B. Semihcan Sermet, D. A. DiGregorio. High frequency spike inference with particle Gibbs sampling. https://www.biorxiv.org/content/10.1101/2022.04.05.487201v1)
Our work provides a statistical framework to reliably identify firing patterns from fluorescence data. Our approach requires a model describing the dynamics of the calcium indicator in response to action potentials. Therefore, in collaboration with the Wang lab in Princeton, we designed a biophysical model of the ultra-fast calcium indicator GCaMP8f. This model takes into account the calcium binding states of GCaMP8f indicator and all processes of calcium buffering. We are currently working on a second paper where we embed such model into our statistical inference framework to identify spike times.
Work Package 2: Inference of MF-GC connectivity, short-term plasticity features and network motifs
In order to characterize the dynamics of synapses between mossy fibers (MF) and granule cells (GC) I have analyzed electrophysiological recordings of cerebellar granule cells stimulated with different frequencies. In particular:
1. I have established an analysis pipeline in collaboration with the experimentalists in the lab (Francisco Urra, B.Semihcan Sermet) and developed a data visualization software of linescan recordings (Fig. 1A).
2. We considered vesicle release models proposed by Hallermann and colleagues 2 to describe short-term synaptic plasticity (STP) and developed Monte Carlo samplers to infer model parameters. This work has highlighted issues related to sloppy direction in the parametric landscape which impact model identifiability. To address this issue we designed “diagonal” moves in the Monte Carlo samplers (Fig. 3), following the correlation structure of some of the parameters. This approach however was not sufficiently general to be carried out across all the possible models of STP, therefore we are currently working on a different approach based on Hamiltonian Monte Carlo, which is a more efficient sampling strategy in the presence of sloppy directions.
Work Package 3: Information-theoretic analysis of the MF-GC network
Information theory provides a foundamental language to quantify the amount information conveyed from mossy fibers to cerebellar granule cells. Thanks to previous theoretical work, the host lab has developed a two-layer perceptron model to describe the cerebellar network, however to carry out an information-theoretic analysis of the circuit we needed to constrain the network model with electrophysiological recordings and fluorescence imaging data. Due to the delays introduced during the course of this action we could not conclude such analysis, however we established the data and the tools to meet the aim of this work package within this year.
The initial six months period of the action has been devoted to the development of analysis pipelines and software to visualize and process fluorescence imaging data of the cerebellar cortex alongside behavioral recordings. I have developed:
1. a Graph-Based SEGmentation (GBSEG) method to extract region of interests (ROI) from fluorescence recordings. The software implementing this algorithm is now an established tool in the lab used by PhD students and postdocs to analyze their data.
2. A graphical user interface to visualize imaging and behavioral data acquired in the lab. By enabling simultaneous inspection of imaging and behavior we could formulate hypotheses on how neuronal activity drives behavior.
3. A graphical interface to curate and manage imaging dataset. This tool enabled to standardize the data organization, to facilitate the systematic analysis of the data.
To extract firing patterns of cerebellar granule cells I developed a Bayesian inference method to estimate spike times from calcium imaging data. Based on previous works I have developed a generative model of fluorescence time series accounting for baseline fluorescence modulation and bursting firing statistics. Then, as described in the DoA, we employed novel sequential Monte Carlo algorithms to estimate static and dynamic model parameters. Our method allows us to quantify the statistical uncertainties associated to these estimates.
Our inference method allows us to resolve inter-spike intervals below five milliseconds, enabling the analysis of high firing neurons.
To accelerate the diffusion of our method we posted our results as a preprint on bioRxiv (G. Diana, B. Semihcan Sermet, D. A. DiGregorio. High frequency spike inference with particle Gibbs sampling. https://www.biorxiv.org/content/10.1101/2022.04.05.487201v1)
Our work provides a statistical framework to reliably identify firing patterns from fluorescence data. Our approach requires a model describing the dynamics of the calcium indicator in response to action potentials. Therefore, in collaboration with the Wang lab in Princeton, we designed a biophysical model of the ultra-fast calcium indicator GCaMP8f. This model takes into account the calcium binding states of GCaMP8f indicator and all processes of calcium buffering. We are currently working on a second paper where we embed such model into our statistical inference framework to identify spike times.
Work Package 2: Inference of MF-GC connectivity, short-term plasticity features and network motifs
In order to characterize the dynamics of synapses between mossy fibers (MF) and granule cells (GC) I have analyzed electrophysiological recordings of cerebellar granule cells stimulated with different frequencies. In particular:
1. I have established an analysis pipeline in collaboration with the experimentalists in the lab (Francisco Urra, B.Semihcan Sermet) and developed a data visualization software of linescan recordings (Fig. 1A).
2. We considered vesicle release models proposed by Hallermann and colleagues 2 to describe short-term synaptic plasticity (STP) and developed Monte Carlo samplers to infer model parameters. This work has highlighted issues related to sloppy direction in the parametric landscape which impact model identifiability. To address this issue we designed “diagonal” moves in the Monte Carlo samplers (Fig. 3), following the correlation structure of some of the parameters. This approach however was not sufficiently general to be carried out across all the possible models of STP, therefore we are currently working on a different approach based on Hamiltonian Monte Carlo, which is a more efficient sampling strategy in the presence of sloppy directions.
Work Package 3: Information-theoretic analysis of the MF-GC network
Information theory provides a foundamental language to quantify the amount information conveyed from mossy fibers to cerebellar granule cells. Thanks to previous theoretical work, the host lab has developed a two-layer perceptron model to describe the cerebellar network, however to carry out an information-theoretic analysis of the circuit we needed to constrain the network model with electrophysiological recordings and fluorescence imaging data. Due to the delays introduced during the course of this action we could not conclude such analysis, however we established the data and the tools to meet the aim of this work package within this year.
This action had an important impact within the host lab, the institut Pasteur and my career development. In particular,
1. as an expert in statistical methods applied to neuroscience I have supervised 1 undergraduate student, 3 master students and 1 PhD student.
2. I have designed analysis pipelines for imaging and behavioral data used in the host lab, and I have designed softwares to enable experimentalists in the lab to manage and visualize their data.
3. we have submitted our work on spike inference using novel statistical methods (doi:10.1101/2022.04.05.487201).
1. as an expert in statistical methods applied to neuroscience I have supervised 1 undergraduate student, 3 master students and 1 PhD student.
2. I have designed analysis pipelines for imaging and behavioral data used in the host lab, and I have designed softwares to enable experimentalists in the lab to manage and visualize their data.
3. we have submitted our work on spike inference using novel statistical methods (doi:10.1101/2022.04.05.487201).