Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Neural Engineering

Neural Engineering - Poster Session C

Poster K2 - Hierarchical decomposition of calcium imaging into multi-level ensembles

Friday, October 25, 2024
10:00 AM - 11:00 AM EST
Location: Exhibit Hall E, F & G

Introduction:  In the brain, information is processed by the cooperative operation of countless neurons. In numerous experiments, spatiotemporal patterns in which multiple cells are active in a specific order have been observed to be repeated. These spatiotemporal patterns, called ensembles, are expected to be the basic units of information processing in the brain, but to know what information they represent, it is necessary to intervene in brain activity when motifs occur and examine their effects on behavior. With the recent development of optogenetics, it is now possible to induce or suppress the activity of specific types of neurons by irradiating the brains of experimental animals with laser light. In this study, we develop a method to identify a hierarchy of ensembles from calcium imaging using a variant of non-negative matrix factorization, namely Neural NMF. Detected ensembles can be put into a feedback system that discovers, defines, and detects motifs from brain activity in real time and causally clarifies their roles.

Our analysis uses calcium imaging videos recorded in one of the coauthor's laboratories that are already made publicly available (CaliAli source data: https://data.mendeley.com/preview/mvg4w89s4s?a=febfa476-d9ef-4e98-abe4-24113a24549e). The videos were recorded from the hippocampus of one mouse for multiple sessions. Each recording session lasts for about 10 minutes. From each video, we obtained a raster plot of calcium events using CaliAli (https://github.com/CaliAli-PV/CaliAli).

Materials and

Methods:  Each raster plot can be represented as a matrix X in R^{NxT} where N is the number of neurons, and T is the number of time points. Neural Non-negative Matrix Factorization (Neural NMF) library (https://github.com/jvendrow/NeuralNMF) was used to factor raster plot matrices into a product of non-negative factor matrices. Neural NMF approximates the raster plot as X ~ (\prod_{j=1}^{L}W^j)H, where L is the number of layers in the hierarchy. All of the factor matrices are constrained to be non-negative. A column of W^1 in R^{NxK_1} represents which neurons make up a lowest-level pattern. For 1 < j <= L, a column of W^j in R^{K_{j-1} x K_j} represents which lower-level patterns are combined to make a higher-level pattern. {K_j}_{j in [1,L]} are hyperparameters that must be set prior to decomposition.

The pattern matrices {W^{j}}_{j in [1,L]} reveal distinct neural activity patterns, while the activation matrix H depicted temporal dynamics and pattern contributions. We set K_1 = 10 and K_2 = 2. From the H matrix, strips of 20 consecutive columns were cut out where the amplitude exceeded a threshold. The threshold was obtained by visually inspecting the histogram of the amplitudes and finding the midpoint between two clusters, one around zero and another far from zero. Five such strips were extracted and then concatenated. The heatmap was generated from the resulting 2x100 matrix.

Results, Conclusions, and Discussions: Neural NMF effectively decomposed high-dimensional calcium imaging data into lower-dimensional representations. In hierarchical decomposition, we set the number of layers to two, that is, L = 2. The activation matrix H depicts temporal activation profiles of neural patterns. The analysis included cutting the data where the total amplitude exceeded the threshold and observing neural activity in 20 subsequent frames. The result revealed consistent patterns in neural activation, validating the capability of our method to find patterns from a raster plot effectively. To illustrate the findings, we generated heatmaps and temporal plots of the matrices W^j and H. Heatmap visualization of H shows patterns often appearing consecutively for several time points. The heatmaps of W^j displayed the spatial distribution of the neural patterns, while the temporal plots of H showed the activation strength of each pattern over time. Each column of W^j represents a unique neural pattern, providing insights into brain connectivity and spatial organization. Since the model is unsupervised, evaluating pattern matrices relies on biological experiments, which we plan to carry out in future work.

The hierarchical decomposition of calcium imaging provided valuable insights into the neural activity patterns of the mouse brain. By decomposing high-dimensional data into interpretable spatial and temporal patterns, the model effectively captured the complex dynamics of neural activity. The pattern matrices {W^j}_{j in [1,L]} revealed distinct spatial patterns of neural activation, highlighting spatial organization within the brain. Concurrently, the activation matrix H illustrated the temporal dynamics of these patterns, demonstrating their roles in responding to stimuli and potential involvement in cognitive functions. The analysis underscored the model's utility in uncovering hidden patterns in neural data, which are essential for understanding brain function. In the figures, we presented factor matrices W^1, W^2, H, and H^1 = W^{2}H using heatmaps. In future work, we plan to investigate the functional role of the patterns by a closed-loop intervention system that triggers optogenetic stimulation to the brain when a pattern is detected.

Acknowledgements (Optional):  This work was supported by grants JP23wm0525003h0003, JP23zf0127005, JP24K10470 and G-7 Scholarship Foundation.