A data-driven framework for modeling the dendritic spine continuum using dimensionality reduction and clustering toward understanding synaptic plasticity

RESEARCH ARTICLE A data-driven framework for modeling the dendritic spine continuum using dimensionality reduction and clustering toward understanding synaptic plasticity Uma Shashi Sharma ☯ *, Philip R. LeDuc☯, Yongjie Jessica Zhang ☯ Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, Pennsylvania, United States of America ☯ These authors contributed equally to this work. * Abstract OPEN ACCESS Citation: Sharma US, LeDuc PR, Zhang YJ (2026) A data-driven framework for modeling the dendritic spine continuum using dimensionality reduction and clustering toward understanding synaptic plasticity. PLoS One 21(6): e0349775. https://doi.org/10.1371/ journal.pone.0349775 Editor: Rakesh Karmacharya, Harvard University, UNITED STATES OF AMERICA Received: December 19, 2025 Accepted: May 5, 2026 Published: June 2, 2026 Copyright: © 2026 Sharma et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data availability statement: The dendritic spine datasets analyzed in this study are publicly available from Ghani et al. (2017) and Smirnov et al. (2018). Processed data and code used in this study are Dendritic spines are dynamic extensions of dendrites that change in shape and distribution in response to neuronal activity, playing central roles in memory and learning. Computational methods are widely used to characterize spine morphology, yet feature selection, dimensionality reduction, and clustering choices are often made a priori and evaluated independently, and as a result it remains unclear how analysis decisions influence low-dimensional representations of spine shape and the biological interpretations drawn from them. We present a decision-based visual characterization framework that systematically evaluates dimensionality reduction and probabilistic clustering strategies for dendritic spine morphometry. Using a labeled two-photon laser scanning microscopy (2PLSM) dataset and a secondary dataset with differing imaging conditions to assess generalization, we compare PCA, ISOMAP, t-SNE, UMAP, and PCUMAP alongside hierarchical clustering, Fuzzy C-Means, and Gaussian Mixture Models. We additionally introduce a Biological Transition Score (BTS) to quantify how well low-dimensional embeddings reflect known developmental and functional relationships among spine types. Across datasets, dimensionality reduction methods capture complementary aspects of spine morphology. On the primary dataset, nonlinear approaches better preserve fine-scale structure, with PCUMAP providing a favorable balance between local structure preservation and global continuity. In contrast, analysis of a lower-resolution secondary dataset shows that PCA is more robust under increased feature-level noise. These findings demonstrate that the optimal dimensionality reduction strategy is dataset-dependent, underscoring the importance of systematic, data-driven method selection. When paired with probabilistic clustering, these representations reveal a morphological continuum that bridges classical “mushroom,” “stubby,” and “thin” spine categories. Increasing the number of identified sub-groups preserves or strengthens structural organization PLOS One | https://doi.org/10.1371/journal.pone.0349775 June 2, 2026 1 / 25 available at: https://github.com/sharmauma1/ DenSpineContinuumModeling.git. Funding: The author(s) received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. relative to expert-labeled classes, demonstrating that weakly supervised representations can resolve intra-class heterogeneity beyond discrete manual classifications. This framework provides a structured, quantitative approach for selecting dimensionality reduction and clustering strategies, enabling more consistent and biologically grounded interpretations of dendritic spine morphology. Introduction The ability of the brain to receive, process, and transmit information relies on dynamic connections between neurons. Dendritic spines, the microscopic protrusions along the branched extensions of nerve cells, are essential to this process, serving as key sites for receiving synaptic information by forming functional connections with neighboring neurons’ axons [1,2]. These tiny structures, ranging from only 0.2–2 µ m in length [3], exhibit significant morphological diversity and plasticity, constantly changing shape in response to neural activity. Furthermore, the complexity, shape distributions and density of dendritic spines are associated with the strength and function of synaptic connections, indicative of calcium dynamics, receptor location, and in turn, the probability of postsynaptic firing [1]. Thus, modeling dendritic spines has tremendous promise for progressing our understanding of brain development and cognitive flexibility [2]. Furthermore, deficits in spine shape and density have been observed in various cognitive disorders and intellectual disabilities, including Traumatic Brain Injury (TBI), Schizophrenia, Alzheimer’s Disease, Down Syndrome, Rett Syndrome, and even chronic stress and anxiety [4–6]. For example, individuals with Down Syndrome, a genetic condition resulting in intellectual disability, present with significant reductions in dendritic spine length and density [4]. Because these smaller-scale structural changes appear long before visible large-scale brain damage, dendritic spines serve as early markers of disease, and studying them can reveal how these conditions progress, enabling better diagnosis and treatment. Traditional dendritic spine characterization methods classify spines into three to five discrete groups (“mushroom”, “thin”, “stubby”, and sometimes “branched” and “filopodia”) based on morphological geometries [2,7]. These geometries are associated with function: mushroom spines have large heads and small necks, and are associated with strong, long-term synaptic connections (lasting weeks to months). Thin and stubby spines are more dynamic; thin spines have small heads with long necks and are involved in synaptic plasticity. Stubby spines lack a distinct neck and are common in early development, both potentially transitioning into mushroom spines over time [1,2]. However, multiple reports have challenged traditional classification methods: As spines develop, they transition from one shape category to another, existing on a continuum rather than in discrete categories [8–12]. Furthermore, the classification approach for dendritic spine analysis may obscure relevant biological features of the data, resulting in a loss of information about feature correlations and spine shape variations [11,13]. PLOS One | https://doi.org/10.1371/journal.pone.0349775 June 2, 2026 2 / 25 Computational approaches, including classification and (...truncated)