Article Figures & Data
Figures
- Figure 1.
Schematic representation of the workflow for generating an objective classification of F-actin organization in dendritic spines. The super-resolution image of F-actin generated using dSTORM microscopy is considered as the input. (1) Using the TWS on input, a segmented image was created. (2) The segments of interest were color coded and a binary image was obtained for F-actin-enriched regions (mask). (3) The super-resolution image of Homer 1c was generated for the same region of interest as that of input. (4) The segmented image of input was spatially correlated with the postsynaptic marker Homer 1c to select for dendritic spines; please refer to Extended Data Fig. 1-1 for a detailed work flow for steps 1-4. (5) The spines obtained from step 4 were further categorized as mushroom, stubby, and thin using supervised learning. (6) The final data were categorized and plotted into different classes as output 1. (7) The tubular model of the input image was generated using ANNA-PALM. (8, 9) Two processing steps were converged to understand the nanoscale distribution of F-actin in dendritic spines generated from the tubular model, which was spatially correlated with Homer 1c-positive regions obtained in step 4. (10) Spine-specific ridges were extracted in the regions identified positive for excitatory synapses; please refer to Extended Data Fig. 1-2 for a detailed work flow for steps 8-10. (11) The spine-specific parameters of the ridges were measured and plotted as output 2.
- Figure 2.
Supervised learning algorithm for morphologic characterization of spines from primary rat hippocampal neurons. A, A gallery of different morphologies of F-actin-enriched compartments in primary rat hippocampal neurons identified as spines. Scale bar = 1 μm. B, A matrix which depicts pair-wise agreement between different experts to classify spines into distinct morphologic classes. The pseudocolor bar depicting the pairwise agreement is shown below. C, A 2-dimensional representation of the classification using two principal components showing that the morphologic characterization of spines forms three nonoverlapping regions. The morphologic features were used for cataloging F-actin structure into a distinct spine category.
- Figure 3.
Analysis of nano-organization of F-actin at 1-nm/px sampling. A, PREM image of cytoskeletal distribution within a spine. Scale bar = 200 nm. B, The segmented image selecting only the thin filaments in PREM indicate the F-actin distribution. C, ANNA-PALM simulation of the F-actin network using tubular model. D, Extraction of ridges by skeletonizing the ANNA-PALM image. E, Overlay of an image obtained by PREM (green) and ridges that mark the F-actin network (red) of the spine. Scale bar = 200 nm. F, G, Magnified views of sections within the spine; please refer to Extended Data Fig. 3-1 for a comparative analysis between PREM and simulated super resolved images. The ridges overlapped with the PREM images with a correlation of >89%. Scale bar = 50 nm.
- Figure 4.
Tubular model of F-actin represents its actual distribution in spines. A, Super-resolution image of F-actin in neurons obtained by dSTORM. Scale bar = 1 μm. B, Mask of F-actin rich compartments in neuronal processes. C, Tubular model of F-actin obtained by ANNA-PALM. D, RSE maps indicating the correlation between dSTORM image and F-actin mask. E, RSE maps of the dSTORM image with a tubular model of F-actin. Scale bar = 1 µm. The pseudocolor bar ranging from purple to yellow indicates low to high error. G, RSP of dSTORM image with the F-actin mask (red) and with the tubular model of F-actin from ANNA-PALM (blue). I, RSE of the dSTORM image with the F-actin mask (red) and with the tubular model of F-actin from ANNA-PALM (blue).
- Figure 5.
Comparison of morphologic features of spines obtained by supervised learning algorithm from WT and APP/PS1 primary mice cortical neurons. A, A gallery of different morphologies of F-actin-enriched compartments in primary mice cortical cultures identified as spines. Scale bar = 1 μm; please refer to Extended Data Fig. 5-1 for the validation of supervised learning algorithm for morphological characterization of spines from primary mice cortical neurons. B, A pie-chart representing proportion of mushroom, stubby and thin spines in WT. C, A pie-chart representing proportion of mushroom, stubby and thin spines in the entire population of dendritic spines in APP/PS1; please refer to Extended Data Fig. 5-2 to view a set of morphologies of spines characterized as mushroom and thin spines.
- Figure 6.
Objective paradigm for segmentation and feature detection in dendritic spines. A, Representative gallery of different classes of dendritic spines are depicted with each class containing six representative spines. Scale bar = 500 nm. We found that the cumulative length of F-actin filaments in mushroom spines were significantly higher in WT spines compared to APP/PS1 spines (average actin filament length: WT mushroom = 5634.5 ± 2034 nm; and APP/PS1 mushroom = 3665.1 ± 1299.2 nm; p < 0.005 for a rank sum test on cumulative F-actin filament lengths for individual spines of WT and APP/PS1 groups), while there was no significant difference in the lengths of the F-actin networks in stubby and thin spines (WT stubby = 2288.5 ± 982.6 nm; APP/PS1 stubby = 2045.4 ± 763.9 nm; WT thin = 2927.3 ± 2023.5 nm; APP/PS1 thin = 3098.9 ± 1439.9 nm; p = 0.12 and p = 0.42 for a rank sum test on cumulative F-actin lengths of stubby and thin spines, respectively). B, The paradigm for feature extraction was performed in two steps. (1) The branch endpoints of the detected ridge of the spine were compared to the centroid of the Homer puncta to define the neck (yellow) and head regions (cyan) of the spine. (2) The length of the ridges was plotted for analysis. B, The difference in the cumulative F-actin filament lengths in mushroom spines was due to difference in their lengths in the head region, rather than the neck (average actin filament length: 5075.7 ± 2048.6 nm and 3126.2 ± 1284.3 nm for WT and APP/PS1 head regions, respectively, p < 0.005 for a rank sum test; 558.7 ± 331.7 and 538.9 ± 404.5 nm for WT and APP/PS1 neck regions, respectively, p = 0.31 for a rank sum test).
Tables
Cumulative length of F-actin (nm) Spine type Subspine compartment WT APP/PS1 Significance Mushroom - 5634.5 ± 2034 3665.1 ± 1299.2 <0.005, yes Spine head 5075.7 ± 2048.6 3126.2 ± 1284.3 <0.005, yes Spine neck 558.7 ± 331.7 538.9 ± 404.5 0.31, no Stubby - 2288.5 ± 982.6 2045.4 ± 763.9 0.12, no Thin - 2927.3 ± 2023.5 3098.9 ± 1439.9 0.42, no
Extended Data 1
On the GitHub repository, there are two folders titled rat and mice. Folder rat has a subfolder xls. Contents of xls are: (1) shape_info.xlsx, (2) class_01.xlsx, (3) class_02.xlsx, (4) class_03.xlsx, and (5) class_04.xlsx. Shape_info.xlsx contains the 22 features identified using Shape Filter plugin in ImageJ class_01.xlsx, class_02.xlsx, class_03.xlsx, and class_04.xlsx contain annotations of spines from four human experts, respectively, of all the 1056 spines. MATLAB code files: shapeinfo_cluster.m, reduces shape information to five dimensions using PCA. These five dimensions are used for training a SVM using MATLAB function fitecoc to classify the spines into three categories. get_head_neck_regions.m, computes cumulative branch lengths for head and neck regions separately from the image input. compare_head_neck_len.m, this code plots the lengths of the branches from head and neck regions as a histogram using nhist.m function. The F-actin images of dendritic spines from rat neuronal cultures is in the folder spines.rar. Folder mice has a subfolder xls. Contents of xls are: (1) shape_info_mice.xlsx, (2) class_01.xlsx, (3) class_02.xlsx, (4) class_03.xlsx, and (5) class_04.xlsx. Shape_info_mice.xlsx contain the 22 features identified using Shape Filter plugin in ImageJ class_01.xlsx, class_02.xlsx, class_03.xlsx, and class_04.xlsx contain annotations of spines from four human experts, respectively, of all the 249 spines. MATLAB code files: shapeinfo_cluster.m, reduces shape information to five dimensions using PCA. These five dimensions are used for training a SVM using MATLAB function fitecoc to classify the spines into three categories. get_head_neck_regions.m, computes cumulative branch lengths for head and neck regions separately from the image input. compare_head_neck_wt_tg.m, this code plots the lengths of the branches from head and neck regions as a histogram using nhist.m function. cumlen_wt_tg_stubbythin.m, computes cumulative F-actin lengths for stubby and thin. The F-actin images of dendritic spines from mice neuronal cultures are in the folder spines.rar. Download Extended Data 1, ZIP file.
Extended Data Figure 1-1
Feature-based supervised learning approach for structure identification. A, A dSTORM image of F-actin from neuronal culture. B, Feature-based segmentation of the dSTORM signal of F-actin and segregation into class 1 (green), class 2 (purple), and class 3 (red). C, Mask of segmented F-actin signal which contains putative spines. D, SRRF image of the postsynaptic marker Homer 1c. E, Feature-based segmentation of the SRRF signal of Homer 1c and segregation into class 1 (green), class 2 (purple), and class 3 (red). F, Mask of a segmented signal indicating the nanoscale localization of postsynaptic density. Scale bar = 500 nm. G, Colocalization of the mask of segmented F-actin with that of Homer 1c. H, Categorization of F-actin-enriched compartments with PSD as spines (green), which were exported for further shape-based analysis. Scale bar = 500 nm. Download Figure 1-1, EPS file.
Extended Data Figure 1-2
Identifying F-actin organization using ridge detection in single spines. A, Feature-based segmentation of the dSTORM signal of F-actin and segregation into class 1 (green), class 2 (purple), and class 3 (red). B, The input dSTORM images were transformed into the tubular model using ANNA-PALM. C, The ANNA-PALM image was transformed and skeletonized using ridge detection module to represent the F-actin ridges. D, The segmented regions colocalizing with the postsynaptic marker Homer 1c were extracted. E, The F-actin mask was used to selectively filter spine-specific F-actin ridges (black) in C. F, The selected ridges (black) depicted bundled F-actin within each spine (red). Scale bar = 500 nm. Download Figure 1-2, EPS file.
Extended Data Figure 3-1
Simulation of dSTORM like images of F-actin from PREM images. A–C, Examples of PREM images with 1-nm/px sampling of a subsection of a neuronal process, where the red region indicates the presence of a spine. D–F, Simulation of single molecule based super-resolution images using SuReSim, with 20-nm/px sampling, of F-actin cytoskeleton in spines identified by PREM. G–I, Approximation of tubular rod-like distribution of F-actin nanoscale images using ANNA-PALM. J–L, Error of mismatch between the tubular model and the simulated single molecule based super-resolution image. The mean RSP between the model and the simulated dSTORM image was 0.89 ± 0.03. The pseudocolor bar ranging from purple to yellow indicates low to high error. Scale bar = 200 nm Download Figure 3-1, EPS file.
Extended Data Figure 5-1
Supervised learning algorithm for morphological characterization of spines from primary mice cortical neurons. A, A matrix which depicts pair-wise agreement between different experts to classify spines into distinct morphological classes. The pseudocolor bar depicting the pairwise agreement is shown below. B, A 2-dimensional representation of the classification using two principal components showing that the morphological characterization of spines forms three non-overlapping regions. A 2-dimensional representation of the classification using two principal components shows that there exist three categories of spines in both WT and APP/PS1, and thus can be used for predicting if a given structure belongs to any of the three categories (red o, WT mushroom; maroon o, APP/PS1 mushroom; dark blue +, WT stubby; light blue +, APP/PS1 stubby; dark green o, WT thin; light green o, APP/PS1 thin). Download Figure 5-1, EPS file.
Extended Data Figure 5-2
A gallery of super-resolution images of mushroom and thin spines. The top panel depicts mushroom spines with very short necks, which would be classified as a different morphological entity by conventional light microscopy. The bottom panel depicts oddly oriented thin spines, which would be characterized as stubby spines by conventional microscopy. Scale bar = 500 nm Download Figure 5-2, EPS file.
In this issue
