Kersting, M., Patrikar, A., Schneider, E., Kusunoki-Martin, T., Drumm, A., O’Neil, C. (2023)
United States maritime services are looking to better integrate artificial intelligence (AI) into their maintenance procedures to improve readiness. Modern naval ships employ many onboard sensors to collect information about the ship’s systems, performance, and navigation parameters. The current method for analyzing this data and detecting possible issues is mostly manual. The abundance of sensor data enables one to apply machine learning (ML) techniques to continuously monitor and analyze the ship’s operations. ML systems can be applied to automatically detect or predict and report potential failures. Using ML to automate the analysis of shipboard sensors will allow the Navy to move from time-based maintenance to condition-based maintenance. In this paper, we describe experiments in which the sensor data is represented as multivariate time-series. In practice, the anomalies do not occur frequently, and the data associated with anomalies is scarce. Therefore, we apply unsupervised time-series anomaly detection (UTSAD) techniques to find anomalies. Initially, the experiments are carried out on a simulated time-series with artificially inserted anomalies, so that the ground truth is known, and quantitative results can be obtained. We experiment with point anomalies as well as subsequence anomalies such as shift, trend, and variance anomalies. For the detection of anomalies, we present results using a density-based spatial clustering of applications with noise (DBSCAN) method, a tree-based isolation forest (IF) method, and a reconstruction-based autoencoder (AE) method. Finally, we present results on the actual ship data and discuss the implementation of an onboard real-time hierarchical analytics system.
Spirou, G.A., Kersting, M., Carr, S., Razzaq, B., Yamamoto Alves Pinto, C., Dawson, M., Ellisman, M.K., Manis, P.B. (2023)
Globular bushy cells (GBCs) of the cochlear nucleus play central roles in the temporal processing of sound. Despite investigation over many decades, fundamental questions remain about their dendrite structure, afferent innervation, and integration of synaptic inputs. Here, we use volume electron microscopy (EM) of the mouse cochlear nucleus to construct synaptic maps that precisely specify convergence ratios and synaptic weights for auditory nerve innervation and accurate surface areas of all postsynaptic compartments. Detailed biophysically based compartmental models can help develop hypotheses regarding how GBCs integrate inputs to yield their recorded responses to sound. We established a pipeline to export a precise reconstruction of auditory nerve axons and their endbulb terminals together with high-resolution dendrite, soma, and axon reconstructions into biophysically detailed compartmental models that could be activated by a standard cochlear transduction model. With these constraints, the models predict auditory nerve input profiles whereby all endbulbs onto a GBC are subthreshold (coincidence detection mode), or one or two inputs are suprathreshold (mixed mode). The models also predict the relative importance of dendrite geometry, soma size, and axon initial segment length in setting action potential threshold and generating heterogeneity in sound-evoked responses, and thereby propose mechanisms by which GBCs may homeostatically adjust their excitability. Volume EM also reveals new dendritic structures and dendrites that lack innervation. This framework defines a pathway from subcellular morphology to synaptic connectivity, and facilitates investigation into the roles of specific cellular features in sound encoding. We also clarify the need for new experimental measurements to provide missing cellular parameters, and predict responses to sound for further in vivo studies, thereby serving as a template for investigation of other neuron classes.
Brandebura, A.N., Kolson, D.R., Amick, E.M., Ramadan, J., Kersting, M.C., Nichol, R.H., Holcomb, P.S., Mathers, P.H., Stoilov, P., Spirou, G.A. (2022)
Neural tissue maturation is a coordinated process under tight transcriptional control. We previously analyzed the kinetics of gene expression in the medial nucleus of the trapezoid body (MNTB) in the brainstem during the critical postnatal phase of its development. While this work revealed timed execution of transcriptional programs, it was blind to the specific cells where gene expression changes occurred. Here, we utilized single-cell RNA-Seq to determine transcriptional profiles of each major MNTB cell type. We discerned directional signaling patterns between neuronal, glial, and vascular-associated cells for VEGF, TGFβ, and Delta-Notch pathways during a robust period of vascular remodeling in the MNTB. Furthermore, we describe functional outcomes of the disruption of neuron-astrocyte fibroblast growth factor 9 (Fgf9) signaling. We used a conditional KO (cKO) approach to genetically delete Fgf9 from principal neurons in the MNTB, which led to an early onset of glial fibrillary acidic protein (Gfap) expression in astrocytes. In turn, Fgf9 cKO mice show increased levels of astrocyte-enriched brevican (Bcan), a component of the perineuronal net matrix that ensheaths principal neurons in the MNTB and the large calyx of Held terminal, while levels of the neuron-enriched hyaluronan and proteoglycan link protein 1 (Hapln1) were unchanged. Finally, volumetric analysis of vesicular glutamate transporters 1 and 2 (Vglut1/2), which serves as a proxy for terminal size, revealed an increase in calyx of Held volume in the Fgf9 cKO. Overall, we demonstrate a coordinated neuron-astrocyte Fgf9 signaling network that functions to regulate astrocyte maturation, perineuronal net structure, and synaptic refinement.
Kersting, Matthew. (2020)
Globular bushy cells (GBCs) of the cochlear nucleus are specialized neurons that encode the temporal features of sound. Multiple auditory nerve inputs are known to synapse onto a single GBC, but the exact number and sizes of these inputs have not been systematically investigated in adult mice. To gain a high-resolution and unbiased look at the auditory inputs contacting GBCs, our lab utilized Serial Block-Face Scanning Electron Microscopy. Specifically, 21 GBCs and all their large inputs were reconstructed at nanometer resolution. To produce the most precise results, we applied careful attention to the reconstruction and implemented cutting-edge meshing algorithms. We found that a range of 5 – 12 large auditory nerve terminals converge onto each GBC, which is higher than previously reported electrophysiological estimates. Interestingly, some GBCs were found to have a single large, dominant input, whereas others did not. Thus, we conclude that there are two models of GBC innervation, i.e., a mixed model (1 or 2 suprathreshold inputs and multiple subthreshold) and a coincidence detection model (all subthreshold inputs). The detailed reconstructions were then combined with a GBC computational model which confirmed the presence of two innervation models. We also present novel discoveries about the structure of GBCs that could only be seen in volume electron microscopy.
Spirou, G.A., Kersting, M., Ellisman, M., Manis, P.B. (2019)
Globular bushy cells demonstrate enhanced synchrony to periodicity in the acoustic waveform relative to auditory nerve fibers (ANF), which provide their driving activation. Two very different mechanistic models to account for this property are the convergence of multiple subthreshold inputs that sum in a short coincidence detection time window, and convergence of multiple suprathreshold inputs, whereby the shortest latency input on each stimulus cycle generates a postsynaptic spike. To provide an accurate count and size measurement of these inputs, we imaged a portion of the GBC region of the cochlear nucleus at ultrastructural resolution, using serial blockface scanning electron microscopy (SBEM). This volume contained 27 complete GBC somata, which is the location of innervation by ANFs via modified endbulb terminals. Minimal stimulation in brain slices indicated 4-6 ANFs (5.1 ± 0.64 sd; Cao and Oertel, 2010) converged on single GBCs. Anatomical reconstruction of nerve terminals indicated a greater number (8.0 ± 2.03) of large terminals per GBC. Two patterns of innervation were noted: pattern #1 had 1-2 large and several moderately sized inputs (9 cells) and pattern #2 had only moderately sized inputs (7 cells), which suggest a mixed model of 1-2 suprathreshold and several subthreshold inputs, and a model of only subthreshold inputs, respectively. To assay the functional significance of these models, we generated swc files from GBC reconstructions, converted them into hoc code, and tested the effects of these two input patterns on a compartmental GBC model. Activation of individual inputs revealed that large inputs were suprathreshold and moderately sized inputs were subthreshold. We next generated spike patterns on afferent ANFs using a cochlear transduction model (Zilaney et al. 2014), and showed that model GBCs had enhanced first spike precision over their individual ANF inputs. Furthermore, model GBCs had enhanced vector strength in response to amplitude modulated CF tones over their ANF inputs (vector strength 0.66). Pattern #2 (mixed model) generated more synchronous (0.89) spike trains than pattern #1 (only subthreshold; 0.84), raising questions about the functional utility of mixed supra and subthreshold inputs.
Spirou, G.A., Kersting, M., Bayliss, T., Razzaq, B., Holcomb, P., Morehead, M., Spencer, N., Ellisman, M., Manis, P.B. (2019)
Computational properties of neural circuits are determined by the integrative properties and synaptic maps of their cellular elements. This information is challenging to acquire. For example, in the auditory system it is incomplete for even the most studied cell type, the bushy cell (BC). BCs encode temporal fluctuations in sound amplitude, and are important for a range of perceptual abilities including sound localization and communication. Prevailing theory is that one of two subtypes of BCs, called globular BCs, accomplish their task by measuring coincidence of similarly weighted sub-threshold auditory nerve (AN) inputs using short integration time constants. However, the number and sizes of these inputs are not known, and the synaptic map of dendrites is woefully incomplete. We employed serial blockface scanning electron microscopy (SBEM) of mouse cochlear nucleus to acquire this information. This approach revealed more inputs (mean 7.8, mode 8, range 5–12, n = 15 cells) than estimated using physiological methods (mean 5.1, mode 5, range 4–6, n = 8 cells; Cao and Oertel 2010), and a large range of sizes (35 – 302 μm2), and likely influence on BC activity. We discovered a new dendrite structure, called a hub, which can yield up to 13 branches, and further document frequent and periodic swellings along dendritic processes. We completed the first synaptic map for a BC dendrite system, and revealed that entire branches were non-innervated. Dendrite branches from the same and adjacent BCs were intertwined, and formed direct membrane contacts. We established a pipeline, using 3D virtual reality software, called syGlass, developed in our laboratory, to skeletonize dendrites and export the structures for compartmental modeling in NEURON. These models permitted activation of AN inputs individually, which revealed that the largest input onto one-half of cells was supra-threshold. Therefore BCs do not operate via a simple coincidence model of uniformly weighted sub-threshold inputs. Dendrites served as a tuned current sink, since their removal rendered even the smallest AN input supra-threshold. Current work is exploring the presence of electrical contacts at apposition of dendrite membrane, and whether these contacts improve synchrony in activity of BC clusters.