Hematopoietic stem cells are the cornerstone of the blood and immune system, maintaining lifelong homeostasis of all blood lineages. As they age, their functional decline drives the deterioration of the entire hematopoietic system in the elderly, contributing to anemia, weakened immunity, and age-related blood disorders. Although biologists have catalogued twelve hallmarks of aging, including cellular senescence, stem cell exhaustion, and epigenetic alterations, the precise molecular processes that erode HSC function over time remain incompletely understood. Because HSCs are extraordinarily heterogeneous, with individual cells in the same aged animal showing very different degrees of decline, researchers have long needed tools that can assess age-associated chromatin states cell by cell rather than in bulk populations.

Existing biological age estimators rely on other molecular readouts. Epigenetic clocks based on DNA methylation, pioneered more than a decade ago, can predict mortality and disease risk and are used to evaluate rejuvenation therapies. Single-cell transcriptomic clocks, proteomic and metabolomic signatures, telomere length, and markers of genomic instability all capture complementary dimensions of aging. What has been missing is a way to exploit the spatial organization of chromatin inside the nucleus, which is known to deteriorate with age at every scale of the genome’s folding hierarchy, from chromosome territories and A/B compartments down to topologically associating domains and chromatin loops. During aging and senescence, cells lose facultative heterochromatin, chromatin redistributes from the nuclear periphery toward the interior, and lamin-associated domains erode, detaching peripheral heterochromatin from the nuclear lamina.

To capture these changes, the team collected 1,229 3D confocal microscopy images of HSC nuclei from young mice aged 10 to 16 weeks and aged mice older than 80 weeks. Because long-term HSCs are rarer in young animals, more young mice were needed to obtain comparable cell numbers. The researchers segmented each 3D stack, resized the images to a uniform isotropic resolution, and extracted 2D slices along the Z-axis, using interpolation to generate additional synthetic slices that preserved biologically meaningful intensity distributions. The result was a balanced training set of 81,643 2D slices, roughly 44 percent from 552 young HSCs and 56 percent from 551 aged HSCs, acquired by three independent operators on two different confocal microscopes.

ChromAgeNet itself is a compact convolutional network, a reduced adaptation of the Xception architecture with depthwise separable convolutions and residual connections, containing only about 350,000 trainable parameters. The network processes each 2D slice and outputs a probability that the chromatin pattern belongs to a young nucleus. Predictions from all slices of the same 3D nucleus are then aggregated through soft voting to produce a single nucleus-level score, a strategy that measurably improved performance. In fivefold cross-validation, with data split at the level of whole nuclei to prevent information leakage, the model achieved an average validation AUROC of 0.77 at the nucleus level, up from 0.72 for individual slices, with the best fold reaching 0.81. After post hoc beta calibration, the output scores aligned closely with true class frequencies, with an expected calibration error of just 0.02.

The researchers benchmarked this approach against classical machine learning pipelines trained on 70 handcrafted chromatin features, including intensity statistics, shape descriptors, and texture measures computed according to the Imaging Biomarker Standardization Initiative. A calibrated XGBoost classifier performed respectably, reaching an AUROC of 0.73, but fell short of the deep learning model. Importantly, the interpretable feature analysis revealed which chromatin properties changed with age: texture descriptors such as size zone nonuniformity and sum entropy, along with Wavelet median intensity and least axis length, consistently ranked highest. Features derived from Laplacian of Gaussian and Wavelet filtered images were especially informative, suggesting that fine-scale spatial variations in chromatin density carry much of the aging signal. A UMAP embedding built from these top features showed clear separation between young and aged cell populations.

Because convolutional networks are notoriously opaque, the team applied a battery of explainable AI techniques to peer inside ChromAgeNet’s decisions. Occlusion sensitivity, which measures how masking small image patches affects predictions, proved the most reliable attribution method in systematic benchmarking. The resulting attention maps told a coherent biological story. In young HSCs, model attention concentrated along the nuclear envelope, forming a continuous thin line of DAPI signal presumably corresponding to lamin-associated facultative heterochromatin, as well as at nuclear invaginations and regions near nucleoli. In aged HSCs, attention was more dispersed and entropy in the attribution maps was higher, highlighting large, irregular DAPI-dense blobs and a thickened, disrupted envelope signal, consistent with known age-related heterochromatin deterioration. Notably, the model’s scores showed no meaningful association with technical variables such as microscope operator, image intensity, noise, or Z-stack size, indicating that it captures biology rather than acquisition artifacts.

The most striking application came when the researchers treated aged HSCs with candidate rejuvenating compounds and asked whether ChromAgeNet could detect a shift toward a youthful chromatin state. They tested four drugs in two mechanistic classes: inhibitors of Rho GTPase signaling, including the Cdc42 inhibitor CASIN and the RhoA inhibitor Rhosin, and inhibitors of H3K9 methylation, including the G9a/GLP methyltransferase inhibitor UNC0646 and the broad-spectrum demethylase inhibitor IOX1. Baseline aged HSCs received a mean youthful score of 0.34. CASIN and Rhosin nudged scores modestly upward, to 0.40 and 0.48 respectively. Remarkably, UNC0646 and IOX1 raised scores to 0.56 and 0.55, essentially matching untreated young HSCs, which averaged 0.55. The finding suggests that pharmacological modulation of H3K9 methylation can restore a youthful nuclear architecture in aged stem cells, and it positions ChromAgeNet as a scalable phenotypic screening readout for rejuvenation drug discovery.

The authors are careful to note the limitations. Ground-truth labels came from the chronological age of the donor mice, a surrogate for biological age that introduces learning noise given the heterogeneity of aged HSCs. Batch effects, a known challenge for image-based profiling, were mitigated through standardized preprocessing, data augmentation, and cross-validation, but wider application will require validation on independent datasets from other institutions, microscopes, and imaging conditions. DAPI staining alone also cannot reliably separate true aging from cellular senescence, whose chromatin signatures overlap, and translation to human samples will demand model adaptation because chromatin organization differs between mouse and human cells. Future work could incorporate additional molecular layers, such as RNA polymerase II or specific histone modifications, and could adopt fully 3D convolutional architectures or visual transformers as larger datasets become available.

Even so, the study marks a conceptual advance: aging leaves a legible fingerprint in the three-dimensional texture of chromatin that a small, inexpensive neural network can decode from a single DNA stain. Because DAPI is affordable and compatible with standard imaging workflows, and because the model is light enough for high-throughput use, the approach could be extended to other stem cell compartments and cell types where chromatin remodeling is a conserved feature of aging. The team has released its curated 3D HSC microscopy dataset together with ChromAgeNet as an open resource, addressing the current scarcity of publicly available stem cell imaging data. If validated functionally and across laboratories, image-derived youthful scores could one day complement DNA methylation clocks, providing a rapid, single-cell view of biological age and accelerating the search for therapies that restore youthful architecture to aging tissues.

Subject of Research: Deep learning-based prediction of hematopoietic stem cell aging from 3D chromatin imaging

Article Title: Deep Learning Predicts Hematopoietic Stem Cell Aging From 3D Chromatin Images

Article References: Picazo, P. I., Mejía‐Ramírez, E., Di Bari, D., Vitali, E., Florian, M. C., & Petrone, P. (2026). Deep Learning Predicts Hematopoietic Stem Cell Aging From 3 D Chromatin Images. Aging Cell, 25(10), Article e70656. https://doi.org/10.1111/acel.70656

Image Credits: AI Generated

DOI: 10.1111/acel.70656

Keywords: hematopoietic stem cells, chromatin architecture, deep learning, convolutional neural networks, biological aging, 3D confocal microscopy, epigenetic clocks, nuclear organization, explainable AI, drug screening, rejuvenation, H3K9 methylation

Splicing is the molecular editing step that removes non-coding introns from freshly made RNA transcripts and joins the protein-coding exons together. The machinery responsible, the spliceosome, must identify precise boundaries between exons and introns, most critically the 3′ splice site that marks the end of an intron. In most human genes, this task is made easier by the underlying nucleotide composition: exons tend to be rich in guanine and cytosine, the G and C of the DNA alphabet, while the surrounding introns are rich in adenine and thymine. That contrast gives the splicing machinery a clear signal of where exons begin and end. But a distinct subset of genes breaks this rule. In these genes, both exons and introns are GC-rich, producing a GC-leveled architecture in which the usual compositional contrast between coding and non-coding regions largely disappears.

That architectural quirk has consequences. When exons and introns carry similar GC content, the spliceosome loses one of its strongest clues for distinguishing exon from intron, and the 3′ splice sites in these genes tend to be cytosine-rich and depleted of thymine, features that make them intrinsically weak substrates for recognition by the splicing machinery. The new study set out to determine how cells manage to splice these difficult genes efficiently despite their unfavorable sequence landscape. The answer, the researchers report, lies in a compartment that has fascinated cell biologists for decades: the nuclear speckle.

Nuclear speckles are membraneless organelles, dense droplet-like structures within the nucleus that are packed with RNA-processing factors, including many components of the spliceosome. Rather than being passive storage depots, speckles are now understood to function as dynamic hubs for RNA metabolism, exchanging factors with nearby genes and concentrating the molecular tools needed for transcript maturation. Recent work has shown that speckles are largely organized by two large scaffold proteins, SON and SRRM2, and that genes located in close spatial proximity to speckles tend to be spliced more efficiently than genes farther away. What remained unclear was the molecular mechanism connecting that spatial coupling to the actual chemistry of splicing.

The Zhejiang University team, led by corresponding author Yafei Yin with important contributions from Xiaoyu Li, approached the question by combining acute protein depletion with a battery of genome-wide assays. Using the dTAG degron system, a technique that allows rapid and targeted destruction of a chosen protein inside living cells, the researchers generated cell lines in which SON could be eliminated on demand. They then profiled the consequences using 4sU-seq, which captures newly synthesized RNA and therefore reports directly on splicing efficiency in real time, along with CLIP-seq and TurboRIP-seq to map where SON binds and which RNA-protein complexes it associates with, U2 RAP-seq to follow the recruitment of the U2 spliceosomal component to pre-messenger RNA, and Pol II ChIP-seq to track the behavior of the RNA polymerase II enzyme that transcribes genes into RNA. All of the resulting datasets have been deposited in the NCBI Gene Expression Omnibus under accession number GSE325871, allowing other researchers to examine the raw evidence.

The data converged on a clear conclusion. When SON was removed, splicing defects appeared preferentially in GC-rich genes, precisely the subset whose weak, cytosine-rich and thymine-poor 3′ splice sites make them dependent on extra help. The mechanism the authors propose is that SON stabilizes the recognition of those weak splice sites, effectively propping up the first and most error-prone step of spliceosome assembly at GC-rich intron boundaries. In other words, SON does not change the sequence of the gene, but it changes the probability that the splicing machinery correctly interprets a difficult sequence, converting what would otherwise be inefficient and error-prone processing into fast, accurate exon joining.

This mechanism elegantly explains why GC-leveled genes cluster near nuclear speckles in the first place. If a gene’s splice sites are inherently weak, the gene benefits from sitting close to a reservoir of concentrated splicing factors, and SON, as a core scaffold of the speckle, appears to be the factor that translates physical proximity into biochemical support. The finding also assigns a concrete molecular role to SON that goes beyond its established architectural function in organizing speckle structure alongside SRRM2. SON, the study suggests, is not merely a beam holding the nuclear compartment together; it is an active participant in the splicing reactions that the compartment enables.

The implications reach beyond basic cell biology. SON has previously been implicated in human disease, and genes with GC-rich, GC-leveled architecture include many regulators of development and cell growth, so understanding how their splicing is safeguarded could illuminate why nuclear organization fails in certain disorders. The work also connects to a broader theme in modern biology: the recognition that genome function depends on spatial organization, with the position of a gene relative to nuclear bodies influencing how efficiently it is expressed. By identifying the sequence feature, the weak 3′ splice site, that makes a gene dependent on speckle proximity, the study provides a concrete molecular bridge between nuclear geography and RNA chemistry, a bridge that has been hypothesized but never resolved in such detail.

Technically, the study stands out for its use of complementary high-throughput methods applied after acute, rather than chronic, protein loss. Chronic depletion experiments can be confounded by secondary effects as cells adapt over days, whereas the dTAG system destroys SON within hours, allowing the researchers to observe the primary consequences of its absence. The combination of nascent RNA sequencing with crosslinking-based binding maps and polymerase occupancy data allowed the team to distinguish direct effects on splice-site recognition from indirect effects on transcription or RNA stability, strengthening the causal chain from SON binding to efficient splicing of GC-rich transcripts.

Open questions remain. The precise molecular contacts through which SON stabilizes 3′ splice-site recognition, whether through direct binding to the weak cytosine-rich sites, through recruitment of U2 snRNP components, or through both routes acting in concert, will require structural and biochemical follow-up. It also remains to be seen whether other speckle components, including the co-scaffold SRRM2, contribute redundant or specialized support for different classes of weak splice sites. Nevertheless, by demonstrating that a single nuclear speckle protein safeguards the splicing of an entire architectural class of genes, the study transforms nuclear speckles from a correlated curiosity into a mechanistically understood support system, and it gives researchers a new lens for examining how the architecture of the genome and the architecture of the nucleus co-evolved to make complex gene expression possible.

Subject of Research: Role of the nuclear speckle protein SON in the splicing of GC-rich genes

Article Title: Nuclear speckle protein SON safeguards efficient splicing of GC-rich genes

Article References: Fan, W., Zhou, Y., Zhang, X., Tong, C., Li, X., & Yin, Y. (2026). Nuclear speckle protein SON safeguards efficient splicing of GC-rich genes. Cell Research. https://doi.org/10.1038/s41422-026-01303-y

Image Credits: AI Generated

DOI: 10.1038/s41422-026-01303-y

Keywords: nuclear speckles, SON, RNA splicing, GC-rich genes, spliceosome, 3′ splice site, dTAG degron, 4sU-seq, CLIP-seq, nuclear organization, pre-mRNA processing, Cell Research

The research, led by Taronish Biayna, Anton Baranovskii and colleagues, provides one of the most detailed portraits yet of how cells manage intron retention, a form of RNA processing control in which an intron, a segment of RNA that is normally spliced out and discarded, is deliberately kept within a mature transcript. Rather than representing a processing error, the study shows that intron retention operates as a genuine regulatory strategy, one that allows cells to pre-manufacture messenger RNAs and hold them in reserve until the precise moment their protein products are needed.

Nuclear speckles, also called interchromatin granule clusters, are dense assemblies of splicing factors and other RNA-processing proteins that punctuate the nucleoplasm. For decades they have been viewed primarily as storage and recycling depots for the molecular machinery that removes introns. The new work adds an unexpected twist: the periphery of these speckles, the thin shell of material surrounding each cluster, functions as a dedicated holding zone for a specific and functionally coherent set of long-lived intron-retained RNAs. The RNAs are not inside the speckles themselves but accumulate in a defined rim around them, suggesting that the boundary between the speckle interior and the surrounding nucleoplasm is an active regulatory interface rather than a passive border.

To reach this conclusion, the team combined super-resolution imaging with single-molecule RNA detection methods, allowing them to see precisely where individual RNA molecules reside relative to speckle markers. They complemented these spatial maps with metabolic labeling approaches that reveal RNA age, showing that the transcripts parked at the speckle periphery are remarkably stable, persisting far longer than typical messenger RNAs, which are usually made, used and degraded within hours. This longevity is a key feature of the compartment: it enables the cell to accumulate a ready supply of transcripts whose final activation can be deferred.

What makes the discovery particularly striking is the identity of the genes whose RNAs are stored in this way. The researchers found that the pooled intron-retained transcripts are disproportionately derived from genes involved in cell cycle regulation and genome maintenance, the very processes a cell must coordinate with exquisite timing as it prepares to divide. By retaining an intron, the cell prevents these RNAs from being translated into protein while keeping them intact and available. The speckle periphery thus acts as a staging ground where the molecular instructions for division and DNA repair are held in check until the proper moment.

That moment, the study shows, arrives at mitosis. When a cell commits to division, its nuclear envelope breaks down, chromosomes condense, and the ordinary architecture of the nucleus dissolves. The researchers observed that the intron-retained RNAs accumulated at the speckle periphery undergo synchronous splicing during this transition, with the retained introns removed en masse as the cell divides. In other words, the entire stockpile is resolved at once, releasing a coordinated wave of newly matured messenger RNAs precisely when the cell reorganizes its contents and re-establishes gene expression in two daughter cells. This synchronized resolution suggests a mechanism by which cells couple RNA maturation directly to the physical events of division.

The finding reframes intron retention from a curiosity of RNA biology into a temporally precise control point. In many previous studies, intron-retained transcripts were viewed mainly as substrates for surveillance pathways that degrade defective RNAs. The new evidence indicates that at least a substantial fraction of retained transcripts are instead functional intermediates, deliberately produced and stored. Their sequestration at the speckle periphery provides a physical address that separates them from the translation machinery and from the splicing factors that would otherwise complete their maturation, while keeping them close to the very machinery they will eventually need.

The spatial logic of the system is as important as the temporal one. By concentrating related RNAs in a shared compartment, the cell may achieve coordination that would be difficult to accomplish with freely diffusing molecules. Grouping cell cycle and genome maintenance transcripts at the speckle rim could allow a single regulatory event, such as the reactivation of splicing during mitosis, to act on many genes simultaneously, ensuring that their protein products appear together rather than in a stochastic trickle. This kind of collective behavior echoes themes familiar from other membrane-less organelles, where concentrated molecules can switch states rapidly in response to cellular cues.

The work also raises new questions that the field is now poised to explore. What molecular signals mark particular intron-retained RNAs for delivery to the speckle periphery, and what prevents their premature splicing during interphase? How is the compartment disassembled and reassembled across successive rounds of the cell cycle, and does the synchronous splicing at mitosis depend on the wholesale reorganization of the nucleus or on dedicated regulatory factors? Answers to these questions could illuminate how cells safeguard the fidelity of division, and how that safeguarding fails in diseases such as cancer, where both splicing patterns and cell cycle control are frequently disrupted.

For now, the study stands as a vivid demonstration that the nucleus is organized with far more purpose than a bag of freely mixing parts. At the edge of each nuclear speckle, cells maintain a carefully curated archive of unfinished messages, genes for division and genome care held in a long-lived, spatially defined waiting room. When mitosis arrives, the doors open all at once, and the archive is converted, in a single synchronized act of splicing, into the working instructions for building two new cells.

Subject of Research: Spatial and temporal regulation of long-lived intron-retained RNAs at nuclear speckle peripheries and their synchronous splicing during mitosis

Article Title: The periphery of nuclear speckles defines a spatially and temporally regulated compartment of long-lived intron-retained RNAs that resolves during mitosis

Article References: Biayna, J., Baranovskii, A., Chaudhuri, A., Paladin, M., Erdem, B., Keller, L.-E., Barutcu, A. R., Dimmeler, S., Marsico, A., & Dumbović, G. (2026). The periphery of nuclear speckles defines a spatially and temporally regulated compartment of long-lived intron-retained RNAs that resolves during mitosis. Nature Cell Biology, 28(9), 1857-1874. https://doi.org/10.1038/s41556-026-02040-5

Image Credits: AI Generated

DOI: 10.1038/s41556-026-02040-5

Keywords: nuclear speckles, intron retention, RNA splicing, mitosis, cell cycle, genome maintenance, RNA processing, nuclear organization, long-lived RNAs, speckle periphery, gene expression, Nature Cell Biology