All experimental procedures with animals were carried out in accordance with National Institute of Health Guide for Care and Use of Laboratory Animals and the project was approved by the Bioethics Committee of the Instituto de Neurobiología (INB) of the Universidad Nacional Autónoma de México (UNAM) (registration number 117). Male C57BL/6 mice, aged 10–12 weeks and with an initial body weight ranging from 20 to 30 g, were used for this study. The animals were sourced from the vivarium facility of the Institute of Neurobiology. They were housed in polycarbonate cages, with 2–3 mice per cage, and provided with water and chow (Purina^® 5001) ad libitum. The housing conditions were controlled, maintaining a temperature of 21–25°C, relative humidity of 45–65%, and a 12-h inverted dark/light cycle.

A solution of scopolamine hydrobromide trihydrate (SCOP, Sigma-Aldrich, #S0929) was prepared in sterile 0.9% saline solution (SS). The mice were divided into two groups: a control group (n = 4), which received SS (vehicle), and an experimental group (n = 4), which received 1.0 mg/kg SCOP. Both the SS and SCOP solutions were administered daily via intraperitoneal injection during the morning period (9:00–10:00 a.m.) for nineteen consecutive days.

Mice were euthanized by intraperitoneal injection of sodium pentobarbital (250 mg/kg). After anesthesia, the animals were intracardially perfused with 4% p-formaldehyde (PFA) in 0.1 M phosphate buffer saline pH 7.4 (PBS). Brains were collected, post-fixed in PFA for 48 h at 4°C, and cryoprotected in 35% sucrose in PBS for 7 days at 4°C. The brains were then bisected, and the right hemisphere was used to obtain 40 μm sagittal sections in a cryostat. Slices were stored in PBS with 0.1% sodium azide at 4°C until analysis.

Microglia were stained with antibodies against Iba-1 protein and an antibody signal enhancer to address over-fixation and non-specific binding (Rosas-Arellano et al., 2016). For analysis, 8–10 slices from the dorsal hippocampus (lateral 0.48 – 1.90 mm) were selected. Sections underwent antigen retrieval in citrate buffer (pH 6.0) at 80°C for 20 min, quenching in 3% H2O2 and 10% methanol in PBS for 30 min, and enhancement in 0.1% Triton X-100, 0.5% Tween-20 in PBS with 50 mM glycine for 45 min. They were incubated overnight at 4°C with anti-Iba-1 antibody (1:500, Abcam, #ab107159). The next day, sections were washed, incubated with secondary biotinylated antibodies (1:500, Vector Laboratories, #BA-9200) for 2 h, and processed with the VECTASTAIN^® ABC-HRP Kit (Vector Laboratories, #PK-4000). The signal was developed with a solution containing 0.0005 g/ml diaminobenzidine, 0.03 g/ml nickel ammonium sulfate, 0.006 g/ml dextrose, 0.0012 g/ml NH4Cl, and 0.0001 mg/ml glucose oxidase until the color changed to purple. The reaction was then stopped with water, and sections were stored in PBS at 4–8°C for 24 h. Finally, sections were mounted, air-dried, dehydrated, cleared in xylene, and covered with Entellan^® before being sealed with coverslips for image acquisition.

Digital images of Iba-1 + cells in the CA1 and dentate gyrus subregions of the hippocampus were captured using a Nikon Eclipse Ci-Li light microscope equipped with a Nikon Plan 40 × /0.65 air lens. Photomicrographs were taken from four equidistant slices per brain (n = 4).

Digital images were pre-processed under blinded conditions in Image J Fiji software based on a previous protocol (Young and Morrison, 2018; Supplementary Figures 4A, B). Briefly, each image was converted into an 8-bit image, and a lookup table of grays was applied. After that, brightness and contrast were adjusted using the Auto function, and an unsharp mask (Radius sigma = 3.0 pixels) and noise elimination (despeckle) were applied to the image. The threshold function was used for segmentation, and each Iba-1 + cell in the image was completely reconstructed manually using drawing tools (Pencil tool, size = 2.0) to finally obtain binary images. Afterward, binary images were noise-smoothed by applying despeckle, close (option for binary), and remove outliers (default options); cleaned by erasing all the background marks; and finally saved in .tiff format for further processing.

MorphoGlia has been developed with a focus on user-friendliness and accessibility, making it an ideal tool for the broader scientific community. The software is available in two main modes: a software mode and an interface mode, both designed to facilitate ease of use. The executable file was generated using PyInstaller,¹ while the interactive interface was built using the Tkinter Python library.² For advanced users, direct modification of the source code is recommended to tailor the application to the specific needs of individual experiments. This approach allows for greater flexibility and customization, ensuring that MorphoGlia can be adapted to a wide range of research scenarios. All components of MorphoGlia, including the executable file, interface mode, and code mode, are available for download from the following GitHub repository: https://github.com/Maya-Arteaga/MorphoGlia. Currently, the executable file is limited to use on macOS with M1/M2 processors.

Please refer to the video tutorial attached in supplementary information section for a practical demonstration of how to utilize the MorphoGlia software (Supplementary Video 1).

Each cell was identified in the complete binary photomicrographs, and classic morphometric features were computed using Python, primarily with the OpenCV library.³ For skeleton analysis, the total branch length (in pixels), number of initial points (cell processes emerging from the soma), number of junction points (branch subdivisions), and number of endpoints (ends of branches) were measured. Cell body analysis included calculating the area, perimeter, circularity (with 1 representing a perfect circle), Feret diameter (maximum caliper diameter), compactness (how closely an object packs its area), aspect ratio (width/height), orientation (angle in degrees), and eccentricity (major axis/minor axis). The same metrics used in cell body analysis were applied to the entire cell. Fractal analysis involved determining convex hulls (the smallest convex set of pixels enclosing a cell) and performing the same calculations as in cell body analysis, as well as calculating the fractal dimension. Sholl analysis consisted of identifying the number of Sholl circles (circles with increasing radii created around the centroid of the cell soma), counting crossing processes (intersections of cell processes with Sholl circles), and measuring the maximum distance (distance between the centroid and the four vertices of the image). These metrics allow researchers to analyze the biologically relevant characteristics of the cell. During morphometric data extraction, a directory is created for each image containing the segmented cells and corresponding analyses. This approach ensures rigorous quality control.

Selecting the most appropriate features to characterize microglia is challenging due to significant biological variability depending on the region and pathology. A fixed set of features that best differentiates morphological states cannot be universally applied. To address this, a dynamic feature selection approach is necessary to ensure relevance and mitigate noise. We employed the Recursive Feature Elimination (RFE) algorithm, a specialized technique for selecting crucial features by iteratively reducing the feature set and removing the least important ones. RFE uses a Random Forest engine as the underlying training model to determine feature importance. The Random Forest algorithm ensures robustness, avoids overfitting, and captures non-linear relationships between features and the target variable. In this study, the groups (SS-CA1, SCOP-CA1, SS-Hilus, and SCOP-Hilus) were used as the target variable. The features analyzed were those computed in the morphology analysis, totaling 32 variables. The RFE algorithm selected the most significant half of these features, enhancing the robustness of the model and enabling feature selection suited to the specificities of each study group. This approach was implemented using the scikit-learn package.⁴

Uniform Manifold Approximation and Projection (UMAP) is a technique designed for non-linear and non-parametric dimensionality reduction. This technique preserves both local and global structures of the data. It assumes that the data is uniformly distributed on a Riemannian manifold with a locally constant Riemannian metric and local connectivity (Taylor et al., 2014; Becht et al., 2018). Key UMAP parameters include the number of nearest neighbors (n_neighbors) and the minimum distance between points (min_dist). The n_neighbors parameter constructs the high-dimensional neighborhood graph, with lower values focusing on local structures and higher values capturing global structures. Recommended values range from 5 to 50. The min_dist parameter controls point clustering, with lower values resulting in tighter clustering and higher values in more dispersed points. The recommended min_dist value is 0.1. Additionally, the number of components (n_components) determines the dimensionality for data reduction. UMAP uses fuzzy set theory to represent the probability distribution in both high-dimensional and low-dimensional spaces, preserving complex data patterns through non-linear embedding (Taylor et al., 2014; Becht et al., 2018). This makes UMAP effective for capturing intricate relationships in the data. This approach was implemented using the scikit-learn package.⁵ For further details consult https://github.com/lmcinnes/umap.

It is essential to adjust UMAP hyperparameters (n_neighbors and min_dist) based on the data characteristics and experimental goals. Various hyperparameters were tested, demonstrating robust results (Supplementary Figure 1).

Following UMAP for dimensionality reduction, Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) was used for clustering. This non-parametric method constructs a cluster hierarchy based on the multivariate modes of the underlying distribution by transforming the space according to density, building the cluster hierarchy, and extracting the clusters. HDBSCAN’s density-based approach makes minimal assumptions about the clusters, identifying them as regions of high density separated by low-density regions, thus eliminating the need to specify the number of clusters beforehand. This method can identify clusters of varying shapes and sizes and creates a hierarchy based on different density levels. Key hyperparameters include the minimum cluster size (min_cluster_size) and minimum cluster samples (min_samples). The minimum cluster size determines the smallest number of points in a cluster for it to be considered valid; fewer points are treated as noise. Min_samples defines the minimum number of neighboring points required for a point to be considered a core point, which must include at least the specified number of sample points (including itself) in its neighborhood. HDBSCAN effectively handles data noise by excluding points that do not fall within high-density regions, making it robust to noise and outliers. This approach was implemented using the scikit-learn package.⁶ For further details consult https://github.com/scikit-learn-contrib/hdbscan.

Dimensionality reduction and clustering result in color-coded data points, each representing a cell. These color-coded cells are mapped back onto the tissue microphotograph to visualize their spatial arrangement (Figures 5A, B and Supplementary Figures 4C, D). This approach serves two main purposes: confirming similarities among cells within the same cluster and providing insights into the spatial distribution of each clustered cell. This visualization is particularly useful for spatial analyses, allowing for the identification of the most affected zones and uncovering previously unexplored patterns in disease physiopathology. In the original images, color-coded microglia were assessed within the CA1 hippocampal layers: stratum oriens (SO), stratum pyramidale (SPyr), and stratum radiatum (SR). The number of each cluster-type microglia was then estimated by manual cell counting in the different zones of the CA1 subregions for both the SS and SCOP groups. Additionally, we utilized the Allen Brain Explorer^® Beta viewer⁷ to obtain spatial references of the CA1 region from frontal and superior perspectives. This tool facilitated enhanced spatial context, enabling precise localization of the slice with the greatest lateral impact.

Statistical analysis was performed using R (v4.4.0). Correlation plots and chi-square tests were generated using the corrplot and gplots libraries, respectively, while heatmaps were created with the ComplexHeatmap library. The chi-square test determined significant associations between study groups (SS-CA1, SCOP-CA1, SS-Hilus, SCOP-Hilus) and clusters from the UMAP-HDBSCAN analysis (Clusters 0–4). Standardized residuals were calculated to identify the nature of these associations, with positive values indicating higher observed frequencies and negative values indicating lower observed frequencies compared to expected values. This test assesses the independence of categorical variables, and standardized residuals highlight specific associations contributing to the overall chi-square statistic, helping to link morphological states (clusters) with study groups. Heatmaps were used to visualize variable correlations with study groups or clusters, facilitating rapid condition comparisons. Spatial analysis and visualization data were analyzed using GraphPad Prism 8 (GraphPad Software, Inc.). The Mann–Whitney U test was performed to evaluate differences in the number of cells and relative frequency per cluster for each CA1 layer of each image, with a p-value < 0.05 considered significant. The sample size was calculated using the following parameters: P(X > Y): 0.8; Alpha two-sided: 0.05; Power: 0.8.

In this study, we evaluated microglial morphology in the hippocampus of C57BL/6 mice following chronic treatment with 1.0 mg/kg (i.p) SCOP (Figure 1A). To visualize microglia in the hippocampal subregions, we performed enzymatic immunohistochemistry using specific antibodies against Iba-1 (Figure 1B). We captured four images of each subregion, CA1 and Hilus, for both the SS and SCOP groups, resulting in a total of 64 images (SS-CA1 = 16; SCOP-CA1 = 16; SS-Hilus = 16; SCOP-Hilus = 16). These photomicrographs were preprocessed to obtain binarized images (Figure 1C).

Once the binary images are prepared, the MorphoGlia interface (Figures 1D, E) or code mode can be used to execute our proposed pipeline. This pipeline involves five steps (Figure 1D): (1) morphology analysis, which extracts 32 microglial morphology features; (2) feature selection, where the RFE algorithm selects the 16 best features that distinguish the study groups; (3) dimensionality reduction, which employs UMAP to project the 16 selected features into a two-dimensional space; (4) clustering, which uses HDBSCAN, a non-supervised algorithm with noise detection, providing a more objective clustering method; and finally, (5) spatial visualization, where cells are color-coded according to their cluster and mapped back onto the tissue microphotograph to visualize their spatial arrangement, facilitating the identification of the most affected zones.

From the 64 images, a total of 1,221 cells were identified, with an average of 19.09 cells per image (Supplementary Table 1). MorphoGlia’s Morphology Analysis extracted 32 features (Supplementary Table 2) from each cell based on five types of analyses: (1) Skeleton analysis, (2) Soma analysis, (3) Cell analysis, (4) Fractal (Convex Hull) analysis, and (5) Sholl analysis (Figures 2A–E).

Once the features were obtained (Figure 2F, bottom left corner correlation matrix), to address the significant heterogeneity in microglia morphology depending on the region and pathology, we used RFE with a Random Forest engine to select the most significant variables that distinguished our study groups (SS-CA1, SCOP-CA1, SS-Hilus, and SCOP-Hilus) as the target variable. This step reduces noise and enhances the performance of dimensionality reduction algorithms (Figure 2F, top right corner correlation matrix).

Once the most significant features were obtained, to reduce their dimensionality, we applied a non-linear dimensionality reduction technique called UMAP (Figure 3A), which converted the 16 features into two dimensions (n_components = 2). Notably, despite varying hyperparameters, the data structure was preserved (Supplementary Figure 1). The hyperparameters that best preserved both global and local structures were n_neighbors = 10 and min_dist = 0.1.

Subsequently, to perform an unbiased classification, we employed HDBSCAN clustering with parameters set to a minimum cluster size of 20 and a minimum sample size of 10. This analysis resulted in 5 distinct clusters (Figure 3B and Supplementary Table 1) and identified three cells as not belonging to any cluster (Supplementary Figure 2), leading to a total of 1,218 cells classified. To determine if our pipeline identified a pattern of microglial morphology in the obtained clusters, we created heatmaps contrasting the variables with the study groups and the obtained clusters (Figure 3C). Additionally, we generated confusion matrices targeting the study groups and the obtained clusters to evaluate whether clustering improved the algorithm’s performance by avoiding the mixing of morphological states within the study groups (Supplementary Figure 3).

First, we evaluated the distribution of the study groups within the manifold and obtained clusters, as well as the percentage of each cluster in the study groups (Figures 4A–D). We then quantified the total number of cells identified in each group (Figure 4E) and calculated the mean number of cells per image: SS-CA1 had 283 total cells (mean: 17.68 cells per image), SCOP-CA1 had 312 total cells (mean: 19.50 cells per image), SS-Hilus had 276 total cells (mean: 17.25 cells per image), and SCOP-Hilus had 347 total cells (mean: 21.68 cells per image) (Figure 3A, Supplementary Figure 5, and Supplementary Table 1).

Given the observed variations in the number of cells corresponding to each cell group among the study groups, we investigated the relationship between the obtained clusters and the study groups using a chi-square test. The test yielded a chi-square statistic of 148.84, p < 2.2e−16, indicating a significant relationship between the clusters and the study groups. To determine which clusters were most associated with each study group, we calculated standardized residuals. Clusters 0 and 1 exhibited significant negative associations with the SS groups (marked in red), suggesting that these clusters are less prevalent in the SS groups. In contrast, Clusters 3 and 4 were positively associated with the SS groups (marked in blue), indicating a higher prevalence of these clusters in the SS groups. The SCOP groups displayed the inverse pattern, with positive associations in Clusters 0 and 1 and negative associations in Clusters 3 and 4 (Figure 4F).

To investigate the spatial distribution of cells based on their cluster labels, we mapped the cells back onto the tissue microphotograph to visualize their arrangement (Figures 5A, B). To validate the strong associations of Cluster 0 (3.99 and 2.47) and the moderate associations of Cluster 1 (1.39 and 1.24) with the SCOP groups (Figure 4F), we examined their number and relative frequencies within the CA1 region. Our analysis confirmed a significant increase in both number and relative frequency of Cluster 0 (p < 0.001) and Cluster 1 (p < 0.05) cells in the CA1 region of SCOP-treated mice (Figures 5C–E).

Following this confirmation, we analyzed the cell count distribution of these clusters across different lateral coordinates (from 0.48 to 1.90 mm in 240 μm intervals) to identify specific locations with elevated cluster concentrations (Figures 5G–J). This analysis revealed a marked increase in Cluster 0 cells in the 0.48–0.96 mm interval (Figures 5G, H) and an increase in Cluster 1 cells in the 1.20–1.44 mm interval (Figures 5I, J) in the dorsal CA1 of SCOP treated mice compared with SS group.

Finally, to determine whether these clusters were predominantly located within specific layers of the CA1 region of the hippocampus, we further divided the CA1 subregion into its anatomical layers: SO, SPyr and SR (Figures 5A, B). We then quantified the number and relative frequencies of these clusters within each layer as observed in the microphotographs (Figures 5K–P). This analysis showed that Cluster 0 was significantly more prevalent in SO (p < 0.001), SPyr (p < 0.05), and SR (p < 0.001) of SCOP mice compared with SS group (Figures 5K, M). Conversely, number of Cluster 1 cells was significantly higher only in SR (p = 0.040), with no significant differences observed in SO (p = 0.265) or SPyr (p = 0.811) between the SCOP and SS groups (Figure 5L). However, no significant differences were found among groups for the relative frequencies of Cluster 1 cells of any hippocampal sublayers (Figure 5N). Therefore, the distribution analysis across different lateral coordinates for this data (Figures 5O, P) revealed that the specific locations of SR and SO with higher number of Cluster 0 cells are consistent with those found in the dorsal CA1 for SCOP animals.