Scalable semi-supervised cell identification reveals canonical swim and preparatory networks
Figure 1: Voltage-sensitive dye imaging during multiple behaviors.
(A) The isolated nerve cord of the leech was dissected and ganglion 10 was prepared for VSD imaging. Anterior or posterior nerves were be stimulated with an extracellular suction-electrode to elicit shortening or swimming, respectively. (B) Schematic of leech ganglion. Cells that have been previously identified are shown in color. Local bending is activated by targeted stimulation of a P cell. (C) Voltage-Sensitive Dye imaging of cells. A simultaneous intracellular (black) and optical (green) recording is shown (top), which illustrates clear signals of oscillations less than 5mV. Three other example traces show optically recorded action-potentials.
Figure 2: Semi-supervised signal extraction with concatenated-trial ICA (ctICA).
(A) PCA is performed on concatenated trials. Four example principal components are shown. Each principal component has a PC mask (left) and multiple traces (right). Two example traces of each PC are shown from different trials. (B) ICA is performed on the top 120-150 principal components. The IC masks (left) and traces (right) are shown for four examples. The top two components are cellular signals. The bottom two components are examples of background neuropil response and motion artifact. These components are discarded. (C) The independent component masks (left) are then segmented using a threshold (middle). An example is shown where two cells are part of the same independent component, but can be separated through spatial segmentation. These cells are strongly electrically coupled to each other and ICA does not separate them because they have similar activity (right). (D) Regions-of-interest are automatically generated from the segmentation of the components. (E) Three components that are from a single source cell are shown. The algorithm occasionally splits single sources (especially larger cells), and components with similar locations and correlated traces can be clustered using the GUI.
(A) Shortening is activated by stimulating an anterior DP nerve (gray bar). Other DP nerves show motor neuron spikes and are used to read behavioral output (top). Several example VSD traces from identified cells are shown in color. (B) Cell traces are decomposed into two factors, a slow factor that follows the behavior (Factor 1, green) and a fast factor (Factor 2, red). (C) The factors are fit to each trace and the coefficients of the factor fits are plotted. The components are given colors based on these coefficients. The coefficients of the examples from A are shown as large circles with white labels. (D) Activity map of shortening. Each component's map is colored based on the shortening factor coefficients and overlaid onto an image of the ganglion. The example cells are circled. The top right letter, G, indicates that this data is from Animal G, and the bottom left number, 3, indicates which ganglion was stimulated.
(E) A burst of 6 action-potentials is given every 2-seconds to sensory P cells (gray bars) to repetitively elicit the local bend response. The motor neuron spikes of the behavior can be seen in the DP nerve (top). The VSD traces of the same example cells are shown in color. (F) The coherence of the sources was calculated against the stimulus at 0.5 Hz (the stimulus frequency), and the coherence phase and magnitude for each trace are plotted as a polar plot. The components are colored based on their phase and magnitude. (G) Activity map of local bending. Colors are based on the phase and magnitude of the coherence analysis. The same example cells are circled. White circle indicates the stimulated P cell.
Figure 4: Analysis of within animal and across animal functional variability.
(A) The shortening feature is extracted from multiple trials of shortening within the same animal. The same cell's response during the two trials of shortening is plotted as a line segment, where the ends of the line segment are the shortening factor coefficients from each of the trials. The activity maps of the two trials are shown next to the coefficient plot. Four different animals are shown for comparative purposes. Well isolated and short line segments indicates that the feature is useful for identifying that cell. Further the common patterns of the behavior can be seen across animals, indicating that the shortening response (especially factor 2) is a good indicator of cell identity. (B) The same is done for the local-bend feature. A clear set of cells can be isolated that are colored pink and green/cyan. In these examples different P-cells were stimulated, showing that many cells respond consistently to stimulation of either P cell. (C) The same is done for the swim oscillation feature. Many of the cells show consistent phases of oscillations during swimming, making them easily identifiable.
Figure 5: Canonical learning of and visualization of medium-dimensional feature space.
(A and B) Three example activity maps of the three behaviors are shown from two example experiments (H and C) for visual comparison (top: shortening, middle: local bending, bottom: swimming). (C) Three cells are selected (green: 155, blue: 208, red: 264) by the user indicated by the three colored ROIs. Swim activity map is shown in background. (D) For each of the three selected cells in Animal H (panel C), ICM colors components from Animal C based on how close each component is to the selected component in the warped distance space. The three selections are visualized through the different R,G,B color channels. ICM uses these distances to compute the most likely homologs across animals, and the homolog that is predicted is shown as an ROI of the corresponding color. (E and F) The matches across animals can then be given labels and the components that are common matches for multiple animals are accumulated into a label category. The red colored components indicate components that were put into an identified cell category and have been given their category name or number, and the black components were components that were not identified and have arbitrary numbers.
Figure 6: Table of cells in canonical swim oscillator and preparatory network.
Through homologous feature matching, we identified 21 canonical swim oscillators, and 9 preparatory neurons. The cells are given numbers based on their locations and best match to other cells in the literature. For each entry, the experiment indicator (8 small colored circles, left), the ganglion position (ganglion), the shortening factors (cross), the local-bend coherence (circle) and the swim coherece (hexagon) are shown. The experiment indicator describes the data sets in which the cell is identified, a full-circle means that both bilateral pairs of cells are present in the ganglion. A semi-circle means only one bilateral pair was seen or the cell was not paired, and an empty circle means the cell was not identified in that experiment. The colors of the experiment indicators are the same colors as the ganglion position ROIs. The position ROIs (ganglion) show the position and size features for each identified cell, and the across animal variability can be seen by the distribution of ROIs. The shortening factors (cross), local-bend coherence (circle) and swim coherence (hexagon) summarize the activity of each cell during the behaviors (see Fig. 3) and show the within and across animal variability. For each of these plots, a single connected line or polygon corresponds to a single component across multiple trials. Each corner of the polygon are the feature coefficients from a single trial of a behavior. Each individual polygon is from a different animal. The cells boxed in yellow are part of the preparatory network, and cells that do not show a significant swim oscillation have an asterisk above the swim coherence hexagon. The entries in the table are organized based on their anatomical position.
Figure 7: Map of canonical swim and preparatory networks.
The identified cells that oscillate with swimming are grouped into one of six phases and colored by phase (left). The cells in the preparatory network are shown on the right.
Figure 8: Electrophysiological verification of targeted swim oscillator neurons.
(A) An activity map was generated from a VSD imaging trial of swimming and three cells (208, 152, 154) were identified and targeted for electrophysiological recordings. (B) Swimming was activated while intracellular recordings were made to targeted cells. The coherence of the intracellular recordings was computed and the traces are colored by the coherence phase in the same fashion as the VSD recordings. (C-F) The same verification experiment was carried out in two more animals targeting different swim oscillators. The electrophysiological recordings of these sample cells are in agreement with the phases predicted by the VSD mapping.
Figure 9: The preparatory network shows rapid responses to the stimulus regardless of behavior.
(A and B) Example VSD traces from three identified cells within the preparatory network (A) and three that are not (B). Each trace is sorted based on the distance the stimulus was from ganglion 10. The ganglion stimulated is shown to the left of each trace. The response latency is shown as a black circle, and this determines the color of the trace (color bar legend, above traces). (C) Activity maps of the response delays for each cell for several examples of shortening and swimming. Each component is colored by the response delay from one trial. Examples are shown for both shortening and swimming and for different stimulated ganglia. The letters at the top right indicate which experiment, and the numbers to the bottom right indicate which ganglion was stimulated. (D) The response latency of the example cells from all experiments are plotted against the stimulus distance from the imaged ganglion. The red line is a linear fit to the response latency. (E) The linear fits of the response latency for all 9 cells in the preparatory network are plotted in yellow. The fits of the three examples cells that are not part of the preparatory network are plotted in blue.
(A)The S cell was stimulated with a 0.5Hz square wave of current (black is optical trace of S cell). Several traces from identified cells are shown as examples. AP, 155, 161, and 212 are part of the preparatory network. (B) The coherence of all components was computed against the S cell, indicating cells that are functionally excited (red dots) and inhibited (cyan dots) by S cell stimulation. (C) S coupling activity map. Several cells are first identified based on their activity during swimming and shortening (colored text) and the S cell is targeted for stimulation (white circle). Components are colored based on their coherence in panel B.
Figure 11: Electrophysiological validation of predicted S cell connections.
(A) In order to identify preparatory network cells, a P cell is activated with a burst of spikes every 2 seconds and the ganglion is monitored with VSDs. Preparatory cells S, 155, 153, and AP can be identified due to their size, position and the fact that they receive P cell input. (B) The coherence of the components was calculated to quickly identify P cell followers. (C) The activity map highlights cells that respond to the P cell stimuli, allowing rapid identification of preparatory cells of interest. This map was used to then target cells with electrophysiology. (D) The S cell and AP cell were targeted with sharp electrodes, the AP cell was slightly hyperpolarized to below spike threshold. Two example traces show one-for-one spiking to large 2-3 mV EPSPs in the post-synaptic AP cell. (E) Cell 155 was then targeted for a paired physiology experiment. Exciting the S cell with current showed a clear excitatory response in the post-synaptic 155, but the one-for-one spiking is unclear. Further, the strength of the synapse is significantly smaller (only about 0.5 mV).(F) Another preparatory network cell, 153. A small excitatory response can be seen. (G and H) Two other cells nearby to 153 and 155 were targeted, these cells were unidentified. There is very little to no response in the other neighboring cells.
Shunting and inhibitory currents co-regulate the input-output function of an identified leech neuron
Figure 1: Modulation of DE-3's input-output function.
(A) We recorded intracellularly from an inhibitory neuron (cell 116 or cell DI-1) and cell DE-3 simultaneously while injecting current pulses into them. Extracellular electrodes recorded the action potentials of DE-3 from the DP nerve. (B) Different levels of depolarizing current steps (0 to +1nA) were passed into cell DE-3 (top) to measure its input-output function. Cell 116 was simultaneously either hyperpolarized (left) or depolarized (right). The largest class of spikes in the DP nerve are from cell DE-3 (bottom); each trace corresponds with a different current step into DE-3. (C) The firing rate of cell DE-3 is plotted against the current step size when cell 116 was silenced by hyperpolarization (black line) or when cell 116 was activated by depolarization (red line). The curves are fit with a straight line, ignoring firing rates less than 1 Hz or greater than 80 Hz. Three examples from different animals are shown. (D) The slope of the linear fits from C are plotted against cell116 stimulus. In all three examples shown, the slope decreases when 116 is depolarized. (E and F) The same experiments were done with a different inhibitory neuron, cell DI-1. The slope of the IO function decreases in only one of the examples shown. (G) Several experiments were combined by normalizing the slope of each cell DE-3's IO function by its slope when the inhibitory neuron was hyperpolarized. The numbers indicate the number of different animals and the total number of experiments. (H) The offsets of the linear fits were normalized by the DE-3 slope and show that 116 and DI-1 have similar effects of the offset. (I) Trial-by-trial comparison of the relative effects of slope and offset from activation of the inhibitory neurons. Cell 116 exhibits a greater effect on the slope relative to the offset than does cell DI-1.
Figure 2: Shunting inhibition predicts synaptic structure of DE-3.
(A) The structure of cell DE-3 after being filled with a fluorescent dye, Alexa 488. The components of the model are overlain on the predicted anatomical regions of cell DE-3. Cell DI-1 synapses are typically located on the first contralateral secondary dendrite (blue box). Previous models (Vu \& Krasne, 1992) predict that cell 116 synapses would be on the main process of DE-3 (red box). (B) Cell DE-3 was filled with a red-fluorescing dye and cell 116 was filled with a green-fluorescing dye. The predicted site of 116 synapses from the model is shown as the red box. A projection at low magnification (top) and 4 consecutive slices of the image stack zoomed at the predicted location (bottom) are shown.
Figure 3: Multiplication via shunting in two-compartment model.
(A) The dendrites transform synaptic activity into a current that reaches the ShC as $I_D$. This current is scaled multiplicatively by inhibitory shunting conductances ($g_\gamma$) before reaching the axon as $I_{Sh}$. (B) The response of the model to step current injection at four different levels of shunting inhibition. (C) The steady firing rate of the model is plotted versus the injected current for each level of shunting as solid lines, and the linear approximation of the model is plotted as dashed lines. Dots correspond to data in B. The linear fit is an excellent approximation except at low firing rates.
Approximate linear equations for the shunting inhibition model:
No comments:
Post a Comment