PI: Mishne, Gal
Institution: UC San Diego
Title: Data-driven analysis for neuronal dynamic modeling
Grant #: EB026936
Abstract: Our main goal is to unravel dynamics in the brain, as they relate to various sensory-motor actions and to the learning process. Two photon calcium imaging has revolutionized experimental capabilities to measure large-scale neuronal activity, but poses a significant challenge in terms of massive dynamical data analysis. We intend to confront these challenges face-on, to significantly boost the quality and relevance of experimental data collected during the process of animal learning and execution of motor functions. Our goal is to build an end-to-end modular platform to organize automatically (in a data agnostic way) the dynamical observation space into dynamic scenarios corresponding to contextual groups of neuronal dynamics and to specific motor activity in different related trials. We aim to extend prior geometric dynamics analysis methods for nonlinear empirical modeling to the complexity of the large-scale neuronal data. Our methodology leads to the determination of low-dimensional intrinsic dynamical sub-processes that provides a coherent explanation of the observed data, and to testable experimental predictions. Unlike conventional neuronal data processing postulating a-priori specific structural models, we rely only on general data-agnostic coherence assumptions. These settings remove bias due to a-priori modeling and enable developing tools that are independent of the acquisition modality, simplifying data fusion (such as neuronal and behavioral observations).
To accomplish these aims we are developing a framework to analyze time-series from neuro-imaging data acquired over multiple weeks as an animal learns a task. This framework will address imaging challenges (extracting neuronal structures from videos, alignment of videos across days, handling multiple acquisition modalities) as well as unsupervised analysis of the activity of hundreds of neurons jointly with the observed motor behavior, across multiple time-scales: from immediate execution of behavior through a single trial to multiple days of learning. Our end-to-end modeling and analysis framework for multi-trial neuronal activity across multiple modalities and spatiotemporal scales includes: 1) low-level processing of raw calcium imaging data , 2) mid-level organization of extracted interconnected neuronal time-traces, 3) high-level analysis of evolving network of neurons and behavior over long term learning.
Link to Data/Model Reuse abstract
- Rapid fluctuations in functional connectivity of cortical networks encode spontaneous behavior (Benisty et al., Nature Neuroscience, in press)
- Visualizing the PHATE of Neural Networks (Gigant et al., NeurIPS 2019, Cosyne 2020): Visualization of learning in artificial and biological networks.
- LDLE: Low Distortion Local Eigenmaps (Kohli, Cloninger and Mishne, Journal of Machine Learning Research, 2021) Bottom-up manifold learning with low distortion.
Functional imaging analysis
- Learning Spatially-correlated Temporal Dictionaries for Calcium Imaging (Mishne and Charles, ICASSP 2019).
- GraFT: Graph Filtered Temporal Dictionary Learning for Functional Neural Imaging (Charles et al., IEEE Transactions on Image Processing, 2022). Morphology agnostic calcium imaging analysis,
Code: https://github.com/adamshch/GraFT-analysis, Matlab GUI: https://github.com/stradaa/GraFT-Application-Dev
- Spectral embedding norm. Visualization of functional imaging data - separating neurons from background activity
- CIDAN (Calcium Imaging Data ANalysis): Code: https://github.com/Mishne-Lab/cidan
- Data Processing of Functional Optical Microscopy for Neuroscience (Benisty et al. 2022) Review of the data processing pipeline of functional optical microscopy for neuroscience and ongoing and emerging challenges
- Learning Disentangled Behavior Embeddings (Shi et al., NeurIPS 2021, spotlight) Robust behavioral embeddings from unlabeled, multi-view, high-resolution behavioral videos across different animals and multiple sessions.
Analyzing multiway tensor data with manifold embeddings and multiway clustering
- Co-manifold learning with missing data (Mishne, Chi and Coifman, ICML 2019) coupled manifold learning on multiway 2d data
- Multiway Graph Signal Processing on Tensors: Integrative Analysis of Irregular Geometries (Stanley, Chi and Mishne, IEEE Signal Processing Magazine) review of graph signal processing for tensors
- Scalable Algorithms for Convex Clustering (Zhou et al. IEEE Data Science and Learning Workshop 2021)
- COBRAC: a fast implementation of convex biclustering with compression (Yi et al. Bioinformatics, 2021)
demo: https://cvxbiclustr.rice.edu/, code: https://github.com/haidyi/cvxbiclustr
- Multi-scale affinities with missing data: Estimation and applications (Zhang, Mishne and Chi, Statistical Analysis and Data Mining, 2021)
2023 Brain PI Meeting Posters:
- GraFT: Graph Filtered Temporal dictionary learning for functional neural imaging
- Neural manifold decoding using low-distortion Riemannian Alignment of Tangent Spaces
2021 Brain PI Meeting Update:
* Calcium imaging analysis:
- GraFT: Graph Filtered Temporal Dictionary Learning for Functional Neural Imaging. Code: https://github.com/adamshch/GraFT-analysis
* Multiway analysis of tensor data.
- Multiway Graph Signal Processing on Tensors: Integrative Analysis of Irregular Geometries
- COBRAC: a fast implementation of convex biclustering with compression demo: https://cvxbiclustr.rice.edu/, code: https://github.com/haidyi/cvxbiclustr
* Low distortion embedding of manifold data: LDLE: Low Distortion Local Eigenmaps