A computational microscopy framework that pairs high-speed super-resolution imaging with convolutional neural networks has achieved high accuracy in discriminating distinct cellular epigenetic states while mapping the nanoscale spatial redistribution of critical chromatin regulators within single cell nuclei. Reported by a research team led by Yicheng Wang and Yutaka Saito, the imaging platform achieved 99.6 percent classification accuracy in distinguishing cultured human embryonic kidney cells exposed to histone deacetylase inhibition from untreated controls, alongside 96.1 percent accuracy in identifying induced pluripotent stem cells derived from patients with Rett syndrome harboring MECP2 mutations. Beyond categorical classification, class activation mapping and quantitative particle segmentation identified the nuclear periphery as a principal topological hotspot of chromatin remodeling. By linking nanoscale visual descriptors directly to sequencing-based profiles of spatial genome architecture and transcriptomic output, the investigation establishes an optical profiling methodology capable of resolving physical chromatin reorganization across intact individual nuclei.
Understanding how the genome is organized within the nucleus remains fundamental to cellular biology, as the physical compaction and three-dimensional arrangement of chromatin directly govern transcriptional activity and cell fate. Although genomic sequencing approaches such as high-throughput chromatin conformation capture, chromatin immunoprecipitation sequencing, and bulk RNA sequencing provide extensive genome-wide catalogs of chromatin accessibility, histone marks, and topological domain boundaries, these biochemical techniques typically require cell fixation and cellular lysis. Consequently, sequencing methods discard crucial spatial context, subnuclear localization, and intercellular morphological heterogeneity across individual cell populations. Conventional optical microscopy, while preserving intact cellular architecture, has historically been restricted by the optical diffraction limit of approximately 200 to 300 nanometers. Furthermore, standard super-resolution microscopy platforms that surpass this barrier often suffer from protracted acquisition times, severely constraining sample throughput and precluding the generation of the extensive image datasets required for training robust deep learning architectures.
To overcome these historical throughput barriers, the investigators implemented a high-speed spinning-disk super-resolution optical system, the Yokogawa CSU-W1 SoRa, equipped with dedicated microlenses on a pinhole disk that sharpen the effective point spread function prior to optical deconvolution. This optical configuration achieved an effective lateral resolution of approximately 120 nanometers with a physical pixel dimension of roughly 20 nanometers, capturing multi-channel image frames at rates reaching up to 20 frames per second. By imaging multi-cell fields across laser excitation wavelengths of 405, 488, and 561 nanometers, the team simultaneously recorded nuclear DNA stained with Hoechst 33342 alongside indirect immunofluorescence signals for two functional chromatin markers: histone H3 lysine 27 acetylation, an established epigenetic indicator of active promoters and enhancers, and CCCTC-binding factor, an architectural protein governing chromatin organization and long-range genomic interactions. The accelerated imaging pipeline enabled the rapid acquisition of more than 400 composite three-channel super-resolution images within a single one-hour recording session.
The experimental study interrogated two biologically distinct models of chromatin perturbation against matched control lines provided by the RIKEN BioResource Research Center. In the first perturbation paradigm, human embryonic kidney 293T cells were cultured in medium supplemented with 1 millimolar valproic acid for 24 hours, alongside parallel untreated controls. Valproic acid functions as a pharmacological histone deacetylase inhibitor targeting Class I and Class IIa deacetylases, blocking acetyl group removal and stimulating global hyperacetylation. In the second perturbation paradigm, the team analyzed human induced pluripotent stem cells derived from three Rett syndrome patients—designated HPS3042, harboring a mutation within the methyl-CpG-binding domain of the MECP2 gene, and lines HPS3049 and HPS3084, each carrying mutations in the transcriptional repression domain—benchmarked against healthy control stem cell line HPS0076. Because MECP2 binds methylated DNA and maintains higher-order chromatin structure, these genetic models enabled detailed evaluation of structural chromatin disruption arising from distinct neurodevelopmental mutations.
Individual cell nuclei were segmented from acquired multi-cell images using a Mask R-CNN model pre-trained on diverse cellular datasets from the 2018 Data Science Bowl. Super-resolution image fields were downsampled to 512 by 512 pixels to execute instance segmentation over the DNA channel, after which predicted nuclear boundary masks were restored to full resolution and filtered according to stringent morphometric criteria including aspect ratio, perimeter dimension, area, and roundness. This automated extraction yielded 3,046 untreated and 3,516 valproic acid-treated kidney cell nuclei, alongside 3,595 control and 1,723 Rett syndrome stem cell nuclei. Extracted nuclei were standardized via zero-padding into uniform 600 by 600 pixel inputs and evaluated across three convolutional neural network architectures implemented in PyTorch, with primary analyses utilizing a nine-layer residual network comprising nine convolutional layers, batch normalization, and a fully connected classification layer. Initial convolutional layers used a 7 by 7 kernel to span multiple optical resolution units, capturing contextual features across pre-trained representations fine-tuned using the Adam optimizer.
Classification experiments across single-channel and combined composite inputs revealed clear hierarchies in epigenetic information content. Among single-channel inputs, the H3K27ac-only model performed closest to the all-channel model with an accuracy of 0.994, whereas models trained exclusively on Hoechst-stained DNA reached an accuracy of only 0.896. This marked disparity demonstrated that coarse nuclear DNA counterstaining alone does not capture the subtle organizational features driven by chromatin remodeling. In the Rett syndrome paradigm, combining all three staining channels similarly produced peak classification performance, yet single-channel CCCTC-binding factor inputs outperformed H3K27ac inputs across accuracy metrics. This indicated that MECP2 deficiency exerts more uniform structural disruption across architectural boundary anchors than across active enhancer acetylation marks.
To interpret the internal representations learned by the convolutional networks, the researchers applied Score-CAM, generating normalized activation heatmaps that highlight specific subnuclear coordinates driving model classification decisions. Individual cell heatmaps were normalized and aligned along fitted nuclear ellipses, revealing that valproic acid-treated nuclei exhibited significant increases in global H3K27ac activation intensity alongside elevated signal concentration at the nuclear periphery. By partitioning each nucleus into five concentric elliptical zones extending from core to perimeter, quantitative scoring demonstrated that peripheral enrichment of both H3K27ac and CCCTC-binding factor was statistically elevated in treated cells. To verify that peripheral heatmaps reflected authentic chromatin redistribution rather than edge segmentation artifacts, the team evaluated raw pixel intensities across concentric zones. Untreated cells exhibited uniform radial puncta densities, whereas valproic acid exposure caused a pronounced radial density gradient, characterized by high central puncta concentrations transitioning to peripheral accumulation of diffuse epigenetic activation signals.
Detailed morphological profiling of individual chromatin foci was achieved through a multi-level Otsu thresholding and marker-based watershed segmentation pipeline executed in Python. In valproic acid-treated kidney cells, the total count of segmented H3K27ac and CCCTC-binding factor puncta increased significantly, while their mean spatial area contracted, pointing toward the structural fragmentation of regulatory chromatin domains into smaller discrete units. Segmented puncta also exhibited higher shape irregularity and marked elevations in average fluorescence intensity. Minimum Euclidean distance calculations between mark centroids revealed that active H3K27ac puncta converged into close physical proximity with CCCTC-binding factor puncta within a range of 0 to 20 pixels. Furthermore, geometric overlap analysis indicated that the proportion of H3K27ac puncta overlapping architectural anchors rose from 0.210 in control cells to 0.269 under valproic acid treatment, while the reciprocal overlap ratio of CCCTC-binding factor puncta expanded from 0.070 to 0.174, demonstrating enhanced physical coupling between active regulatory elements and structural domain boundaries.
Sequencing assays performed on matching cell populations corroborated the imaging observations and linked visual alterations to genomic coordinates. High-throughput chromatin conformation capture revealed that valproic acid induced extensive compartment switching, characterized by an expansion of transcriptionally active A compartments and a contraction of repressive B compartments. Differential compartments displaying an eigenvector difference greater than 0.5 significantly overlapped with heterochromatic lamina-associated domains identified through reported DamID profiling. Specifically, B-to-A compartment transitions were strongly enriched within lamina-associated domains, aligning with altered peripheral imaging signatures. Integration with RNA sequencing identified 2,037 up-regulated and 86 down-regulated genes, showing that roughly 30 percent of differentially expressed transcripts resided within lamina-associated domains. Notable up-regulated loci within switched peripheral domains included the nuclear envelope structural gene SYNE1, the transcriptional regulator SP110, the collagen assembly gene P3H2, and the tight junction component CLDN1, with P3H2 activation coinciding directly with a gained intronic H3K27ac ChIP-seq peak and enhanced long-range looping interactions.
Parallel evaluations in Rett syndrome stem cells demonstrated contrasting spatial dynamics, characterized by a loss of puncta abundance and area alongside an uncoupling of active enhancers from structural boundaries. The minimum distance separating H3K27ac from CCCTC-binding factor increased, with overlap ratios declining from 0.318 to 0.224 for acetylation and from 0.266 to 0.174 for the insulator mark. Cross-domain classification confirmed strong generalizability between patient lines HPS3042 and HPS3084, whereas line HPS3049 displayed divergent features. To verify the necessity of super-resolution optics, the researchers applied Gaussian blurring with an 11 by 11 kernel, simulating diffraction-limited conventional confocal microscopy covering 550 nanometers. Under blurred conditions, classification accuracy deteriorated across models, and statistical differences in puncta abundance between diseased and healthy states lost significance. The authors noted that future iterations of the platform must expand beyond two-dimensional imaging to resolve three-dimensional volumetric chromatin folding, incorporate internal landmarks such as nuclear speckles, and integrate nucleoporin mapping to separate active pore domains from repressive peripheral lamina.
Subject of Research: Deep learning and super-resolution microscopy for single-cell epigenetic profiling
Article Title: Image-based epigenetic profiling with deep learning and high-speed super-resolution microscopy
Article References: Wang, Y., Ab Ghani, N. S., Dutta, M., Adachi, S., Katoh, K., Namihira, M., Mitsuyama, T., & Saito, Y. (2026). Image-based epigenetic profiling with deep learning and high-speed super-resolution microscopy. Epigenetics & Chromatin, 19(1), Article 15. https://doi.org/10.1186/s13072-026-00662-5
Image Credits: AI Generated
DOI: 10.1186/s13072-026-00662-5
Keywords: Epigenetics, Super-Resolution Microscopy, Deep Learning, Chromatin Organization, Histone Acetylation, CTCF, H3K27ac, Valproic Acid, Rett Syndrome, MECP2, Lamina-Associated Domains, Hi-C, ChIP-Seq, RNA-Seq, Single-Cell Imaging, Convolutional Neural Networks, Score-CAM
News Source: Blake Davidson. (November 10, 2026). Super-Resolution Imaging and Deep Learning Accurately Map Subcellular Epigenetic Reorganization. Scienmag.
