Spatial mapping of RNA turnover kinetics in the mouse brain

Read post on nature.com

MainThe complex functions of the mammalian brain arise from diverse cell types organized into spatially distinct circuits and regions1. This intricate architecture requires precise spatiotemporal control of gene expression at both transcriptional and post-transcriptional levels2. RNA turnover kinetics, which governs how rapidly transcripts are produced and degraded, shapes steady-state transcriptomes and determines the temporal precision of gene expression responses3,4. Rapid RNA turnover enables cells to adapt to developmental cues and environmental stimuli5,6,7, whereas slow turnover supports long-range RNA transport for localized protein synthesis in distal neuronal compartments8 and stabilizes long-lived noncoding neuronal RNAs involved in maintaining chromatin integrity9. Dysregulation of RNA turnover and stability has been implicated in neurodevelopmental, neuropsychiatric and neurodegenerative diseases2,6,10,11,12,13, highlighting the need to map RNA turnover kinetics and stability regulation in vivo across the spatial and cellular contexts of the brain.Standard single-cell and spatial transcriptomics technologies measure total RNA abundance, a static snapshot reflecting the balance between RNA synthesis and degradation, thereby obscuring underlying kinetic regulation. Metabolic RNA labeling can distinguish newly transcribed RNAs from pre-existing pools, enabling quantitative measurement of RNA turnover through ‘single-pulse’ or ‘pulse-chase’ labeling strategies14,15. Recent approaches combine metabolic labeling with single-cell RNA sequencing (scRNA-seq) or imaging to measure cell-type-specific RNA turnover kinetics, but they remain largely restricted to analysis of in vitro cultured cells15,16,17. Thus, the spatial landscape of in vivo RNA kinetics, and how sequence features or post-transcriptional regulators shape RNA stability, remain largely unexplored.Cell-permeable nucleoside analogs such as 4-thiouracil (4tU)18, 4-thiouridine (4sU)19 and 5-ethynyluridine (5-EU)20 can label newly transcribed RNAs in vivo, but their utility for studying RNA turnover and stability is not well established. The most widely used in vivo RNA labeling method, ‘TU tagging’21,22, employs transgenic expression of Toxoplasma gondii uracil phosphoribosyltransferase (UPRT) to convert 4tU to 4-thiouridine monophosphate (4-thio-UMP), enabling its incorporation into nascent RNAs in specific cell-types. However, the requirement for transgenic models limits its scalability and broader application.Here, we optimized and validated a transgenesis-free 4sU metabolic RNA labeling strategy to quantify in vivo RNA turnover in the mouse brain and integrated it with in situ chemical recoding to develop spatial new RNA tagging sequencing (spatial NT-seq). This method enables simultaneous spatial mapping of RNA abundance and turnover kinetics in intact tissues. Applied to mouse brains after electroconvulsive stimulation (ECS), a clinically effective treatment for refractory depression, spatial NT-seq disentangled the contributions of RNA synthesis and degradation to activity-induced gene expression changes across brain regions. It further revealed striking regional heterogeneity in RNA turnover and identified the dentate gyrus (DG) as a hotspot of RNA turnover marked by coordinated upregulation of RNA synthesis and decay, suggesting a RNA kinetics scaling mechanism that may support functional plasticity of DG cells during continuous neurogenesis and external stimuli. Together, our ‘in vivo timescope’ framework integrates transgenesis-free RNA labeling, spatial multimodal transcriptomics (that is, co-detection of newly synthesized and pre-existing RNAs) and computational modeling to dissect how transcriptional and post-transcriptional regulation jointly shape gene expression landscapes.ResultsBenchmarking transgenesis-free metabolic RNA labeling in mouse brainsThe lack of systematic validation has limited the application of in vivo metabolic RNA labeling for studying RNA turnover and stability in tissues. Although 4sU enables direct RNA labeling in non-neural tissues23, its labeling efficiency and potential toxicity in the brain remain uncertain. To address these gaps and enable analysis of challenging brain cell types, we optimized droplet-based single-cell metabolically labeled new RNA tagging sequencing (scNT-seq)15 to improve chemical conversion efficiency and library complexity (scNT-seq2 in Fig. 1a, Extended Data Fig. 1a–c, and Supplementary Fig. 1). Using scNT-seq2, we systematically validated transgenesis-free 4sU labeling for profiling RNA turnover in embryonic and postnatal mouse brains.Fig. 1: Benchmark and validation of in vivo transgenesis-free metabolic RNA labeling for cell-type specific RNA turnover analysis in mouse brains.Full size imagea, A schematic depiction of the integrated workflow of in vivo metabolic RNA labeling followed by scNT-seq2. b, UMAP visualization of cortical cells from four intraperitoneally injected UPRT transgenic mice colored by annotated cell types. DMSO injection serves as the control for the 4tU-labeled mouse, while saline injection is the control for the 4sU-labeled mouse. The cell number for each sample is indicated on the top. astro, astrocytes; CR, Cajal–Retzius cells; EC, endothelial cells; Ex-L2/3/4/5/6, cortical-layer-specific Ex neurons; fib, meningeal fibroblast; Inh-CGE, caudal ganglionic eminence-derived interneurons; Inh-MGE, medial ganglionic eminence-derived interneurons; MG, microglia; oligo, oligodendrocytes; RG, radial glial cells. c, A box plot showing the 4sU- or 4tU-labeled new RNA fraction per cell for both Neurod6-positive and -negative cell-types (as defined in Fig. 1b). For box plots, it shows the median as the center line, the 25th and 75th percentiles as box bounds and whiskers extending to the smallest and largest values within 1.5× the interquartile range. The group mean is shown as a dot, with its value indicated above the box. Cell numbers for each condition across two cell-type groups (Neurod6 positive/Neurod6 negative): saline, 557/1,502 cells; DMSO, 477/1,750 cells; 4sU labeled, 422/1,765 cells; 4tU labeled, 721/1,737 cells. d, Scatter plots comparing transcriptome-wide gene-level new RNA fractions between 4sU- (y axis) and 4tU-labeled (x axis) UPRT mice (P12 CTX) in Neurod6/NEX-cre targeted (Neurod6-positive cortical Ex neurons on the left) or nontargeted (Neurod6-negative cortical cells on the right) cell types. Pearson’s correlation coefficient R and the corresponding two-sided P value are shown (not adjusted for multiple comparisons). e, Violin plot illustrating the proportion of 4sU-labeled new transcripts per cell from in vivo-labeled embryonic day 16.5 (E16.5) mouse cortical tissues compared to in vitro-labeled cultured E16.5 cortical cells15. The mean new transcript fractions for each sample is presented by a white dot and the value is displayed above. Cell numbers for the conditions: control (media for in vitro culture), 1,999; in vitro-labeled (rep1), 2,450; in vitro-labeled (rep2), 4,295; control (saline injection for in vivo labeling), 3,129; in vivo labeled (rep1), 3,823; in vivo labeled (rep2), 3,305. See ‘Data visualization’ in the Methods for definitions of box plot elements. f, A box plot (top) showing the 4sU-labeled new transcript fraction per cell for wild-type postnatal day 7 (P7) mouse cortical samples in labeling time evaluation experiments. A schematic depiction of 4sU labeling time experiment is shown (bottom). Cell numbers for the samples (see Methods for cell filtering): saline control, 2,904; 4sU-2h-rep1, 4,476; 4sU-2h-rep2, 5,002; 4sU-4h-rep1, 4,154; 4sU-4h-rep2, 4,858. For box plots, it shows the median as the center line, the 25th and 75th percentiles as box bounds, and whiskers extending to the smallest and largest values within 1.5× the interquartile range. The group mean is shown as a dot, with its value indicated above the box. g, Scatter plots showing NTRs (x axis, from RNA labeling) and USTRs (y axis, from RNA splicing) for gene groups with distinct gene lengths (short – blue, 100 kb) in cortical Ex neurons. Only expressed genes (transcript counts >5, CPM) are shown and the number of genes for each group is indicated. The marginal histogram plots depict the distributions of NTRs (x axis) and USTRs (y axis) for each gene group. h, A scatter plot showing the NTR derived from RNA labeling and USTR from RNA splicing for cell-type-specific TFs (red, n = 124) and broadly expressed slow turnover genes (blue, n = 269) in cortical Ex neurons. The size of the circles represents total RNA levels for each gene (CPM). The marginal density plots depict the distributions of NTRs (x axis) and USTRs (y axis) for two gene groups.Source dataFirst, using the ‘TU tagging’ as a benchmark, we evaluated the in vivo RNA labeling efficiency with 4sU in the brain. We generated ‘Ex-UPRT’ mice by crossing a Nex-Cre (driving Cre expression in Neurod6+ forebrain excitatory (Ex) neurons)24 with a conditional UPRT line (CA>GFPstop>UPRT)21, achieving Ex neuron-specific Uprt expression. Postnatal day 12 (P12) ‘Ex-UPRT mice’ were injected intraperitoneally with 4tU or 4sU at equimolar doses, and scNT-seq2 identified 15 major cortical cell types (Fig. 1b) and confirmed Uprt transgene expression restricted to Neurod6-expressing Ex neurons (Extended Data Fig. 1d,e). Both 4sU and 4tU labeling minimally altered cell-type-specific transcriptomic states, as shown by closely matched distributions of cortical cell types in treated and control mice (Fig. 1b). Quantification of labeled new RNA fractions indicated robust labeling in Neurod6-positive Ex neurons with both 4sU and 4tU (Fig. 1c), with highly concordant transcriptome-wide gene-level measurements from 4sU and 4tU labeling in cortical Ex neurons (Fig. 1d). Notably, 4sU also efficiently labeled new RNAs in Neurod6-negative cell types independent of UPRT expression, whereas 4tU labeling in these cell types was substantially lower (Fig. 1c,d), probably due to low endogenous UPRT-like enzymatic activity25. Thus, 4sU enables efficient and broader RNA labeling in the brain, matching 4tU labeling efficiency in UPRT-expressing cells.Second, we evaluated transgenesis-free 4sU RNA labeling in wild-type animals. Time-pregnant females were injected 4sU, and embryonic day 16.5 (E16.5) embryonic cortices were compared to primary E16.5 cortical cells labeled in vitro. This comparison showed similar increases in 4sU-labeled new RNA fractions (in vitro, 16.5%; in vivo, 15.8%) compared to saline controls (2 in Extended Data Fig. 5d; adjusted P UPRT females with Nex-Cre/+ males and genotyping F2 offspring. Genotyping of the UPRT transgene was performed as previously described21 using the following primers: 5′-AGTGACAACCCCTCTGGATG-3′ and 5′-CATCGGATCTAGCAGCACA-3′. Nex-Cre insertion was genotyped using the following previously reported primers24: 5′-GAGTCCTGGAATCAGTCTTTTTC-3′ and 5′-CCGCATAACCAGTGAAACAG-3′.All animal procedures were conducted in accordance with the ethical guidelines of the National Institutes of Health (NIH) and were approved by the Institutional Animal Care and Use Committee of the University of Pennsylvania. Mice were group-housed (three to five per cage) in a controlled environment under a 12-h light/dark cycle (light on at 7:00 am), 55% relative air humidity and 22 °C room temperature, with ad libitum access to food and water.ECSBefore stimulation, both ECS-treated and control mice underwent a 3-h in vivo metabolic RNA labeling via intraperitoneal injections of 4sU (1.2 mg g−1 of body weight). For ECS treatment, 8-week-old C57BL/6J male mice were subjected to a 1.0 s electrical stimulus (total duration) consisting of 100-Hz, 18-mA, 0.3-ms pulses, delivered using a Ugo Basile ECT unit (model 57800) as previously described41. Control mice were handled identically in parallel but did not receive the electrical stimulation. Mice were euthanized either 30 min or 120 min following ECS. For each of the three time points, including non-ECS (control), 30 min (ECS 30-min) and 120 min post-ECS (ECS 120-min), biologically independent samples were collected from two adult mice.Whole brains were dissected and embedded in optimal cutting temperature (OCT) compound as described for the spatial NT-seq. OCT-embedded tissue blocks were stored at −80 °C in a sealed container. For the non-ECS mice, four coronal brain sections were collected from each mouse: two for the in situ SLAM reaction and the other two as nonreaction controls. For ECS-treated mice, two coronal sections per mouse were collected for spatial NT-seq experiments.In vivo metabolic RNA labeling in embryonic and postnatal animalsTo compare TU-tagging (4tU in UPRT transgenic mice) with direct 4sU RNA labeling, littermate CA>GFPstop>UPRT/+; Nex:Cre/+ mice were injected intraperitoneally with either 4sU (0.8 mg g−1 of body weight) or 4tU (0.4 mg g−1 of body weight) solution for 4-h labeling. Notably, dosages were equimolar, accounting for the molecular weight difference between 4sU (260 g mol−1) and 4tU (128 g mol−1). In brief, 4tU (Sigma, 440736) was dissolved in DMSO (250 mg ml−1 stock) and diluted in corn oil to 50 mg ml−1 immediately before injection, as previously described70. For transgenesis-free in vivo metabolic RNA labeling using 4sU (Sigma, T4509), a 100 mg ml−1 stock solution was prepared in 0.9% sterile saline and stored in aliquots at −20 °C.To optimize 4sU-based transgenesis-free in vivo metabolic RNA labeling, dosage was adjusted by body weight: 0.8 mg g−1 for early postnatal mice (P7–12) and 1.2 mg g−1 for pregnant females (E16.5 embryos) and adult mice (ECS experiments). Intraperitoneal injections of 4sU or saline (control) were administered accordingly. For dose–response experiments in P7 pups, 4sU was diluted to 50, 100 or 150 mg ml−1 to deliver 0.4, 0.8 and 1.2 mg g−1 at 8 μl g−1 injection volume. Labeling durations ranged from 2 to 4 h. After labeling, mouse cortex (CTX) or whole brains were collected for scNT-seq2 or spatial NT-seq. For scNT-seq2, tissues were dissociated into a single-cell suspension; for spatial NT-seq, whole brains were dissected, embedded in OCT compound, snap-frozen and stored at −80 °C following the Visium Tissue Preparation Guide (10x Genomics, CG000240, Rev D).Quantification of 4sU in mouse brain by LC–MS/MSWild-type P7 pups received intraperitoneal injection of 4sU at 0.8 mg g−1 body weight and were analyzed after 1, 2 or 4 h. Cortical tissues were immediately dissected, weighed and transferred to Eppendorf tubes for LC–MS/MS analysis in collaboration with Creative Proteomics. In brief, an internal standard solution of uridine-d2 in water (2 μl mg−1 tissue) was added to each sample, followed by homogenization using an MM 400 mixer mill at 30 Hz for 3 min. Methanol (18 μl mg−1 tissue) was then added, and samples were homogenized for an additional 2 min. After centrifugation at 21,000g for 10-min, the supernatant was diluted fivefold with water.For LC–MS/MS analysis with multiple reaction monitoring mode, 10 μl of each sample and calibration standard solution were injected into a C18 column (2.1 × 100 mm, 1.8 μm) using an Agilent 1290 UHPLC system coupled to an Agilent 6495 C triple-quadrupole mass spectrometer with positive-ion detection. In brief, gradient elution was performed using 0.1% formic acid in water and methanol as the binary solvents. Quantification was based on a standard curve prepared from serial dilution of 4sU calibration solutions (ranging from 0.0001 to 10 nmol ml−1) in a concentration-matched internal standard-containing solvent. The concentrations of 4sU in tissue samples (nmol g−1 tissue) were calculated using internal standard calibration by interpolating the calibration curve with the respective analyte-to-internal standard peak area ratios measured from the samples.Tissue dissection and sample processing for scNT-seq2For embryonic CTX samples (E16.5), brains were dissected in cold DPBS under a microscope as previously described71, keeping all tissue samples in cold buffers on ice throughout. Cortices from two embryos were pooled and resuspended in 0.5 ml DPBS with 0.01% BSA and 30 μM of transcription inhibitor actinomycin D (ActD, Sigma, A9415). Tissue was dissociated by pipetting (~20 times with a P1000 tip, followed by two times using a P200 nested in the P1000 tip) and filtering through a 40-μm cell strainer. Cells were pelleted (300g, 5 min) and resuspended in 0.5 ml DPBS with 0.01% BSA (without ActD). Cell numbers were determined using the Countess II system, and cells were split into two 1.5-ml LoBind tubes (Eppendorf) for methanol fixation.For postnatal CTX samples (P7–12), mice were euthanized by cervical dislocation and brains were dissected on an ice-cold platform72. Cortical tissues were minced with a razor blade to small pieces and maintained in cold DPBS with 0.01% BSA and 30 μM of ActD. Tissue dissociation was performed using the Worthington Papain Dissociation System with a 30-min enzymatic digestion, following the manufacturer’s protocol with the following modifications. In brief, transcription inhibitor (15 μM ActD) was added to both the digestion (with Papain/DNase) and the density gradient (with albumin-ovomucoid inhibitor) solutions. Cells were pelleted (100g, 6 min), resuspended in 0.5 ml DPBS with 0.01% BSA (without ActD), counted using the Countess II system, and divided into two 1.5-ml LoBind tubes (Eppendorf) for methanol fixation.Following tissue dissociation, cells were fixed as previously described15,73. Single-cell suspensions were prepared as outlined above and resuspended in 0.4 ml DPBS with 0.01% BSA. The suspension was split into two 1.5-ml LoBind tubes (Eppendorf) and 0.8 ml of methanol was added dropwise to achieve a final concentration of 80% methanol in DPBS. Cells were incubated on ice for 1 h and then stored (in LoBind tubes) at −80 °C for up to a month. For rehydration, cells were centrifuged at 1,000g for 5 min at 4 °C, methanol–DPBS solution was carefully removed and the cell pellet was resuspended in 1 ml DPBS with 0.01% BSA and 0.5% RNase inhibitor (Lucigen, 30281-2). After cell counting with the Countess II system, the cell suspension was diluted to 100 cells μl−1 and used immediately for the scNT-seq2 experiment.Development and benchmarking of scNT-seq2 using in vitro-labeled K562 cellsHuman K562 cells (ATCC, CCL-243) were cultured in RPMI media supplemented with 10% fetal bovine serum (Sigma, F6178). For in vitro metabolic RNA labeling, the culture media was replaced with fresh medium containing 100 μM 4sU. After 4-h labeling, K562 cells were then collected by centrifugation, rinsed once with DPBS and fixed in methanol as described above.While metabolic labeling-based time-resolved scRNA-seq approaches enable cell-type-specific RNA kinetics measurements15,17,57,74,75,76, they remain largely limited to in vitro or ex vivo systems. To benchmark in vivo metabolic RNA labeling of intact mouse brains, we used scNT-seq15, a droplet microfluidics-based, time-resolved scRNA-seq method optimized for capturing diverse brain cell types77 and achieving efficient on-bead conversion of 4sU-labeled new transcripts. By capturing both nuclear and cytoplasmic RNAs, scNT-seq jointly profiles ‘new’ RNAs marked by T-to-C substitutions and ‘old’ RNAs in whole cells. As in vivo brain cells may have lower RNA content and labeling efficiency than in vitro cell lines, we further developed and benchmarked scNT-seq2 in K562 cells to improve its performance for in vivo applications.To enhance the efficiency of second-strand synthesis in scNT-seq, we first designed and tested a panel of random oligos (Supplementary Fig. 1a and Supplementary Table 3), guided by previous studies15,78. All primers were synthesized by Integrated DNA Technologies. In the random primer comparison experiment (Supplementary Fig. 1b,c), each oligo was added to a 200-μl Klenow-based reaction mixture (Klenow buffer A: 1× Maxima RT buffer (Thermo Fisher), 12% PEG-8000, 1 mM dNTPs (Clontech), 5 μM template switch oligo (TSO)-GAATG (TSO-GAATG: /5SpC3/AAGCAGTGGTATCAACGCAGAGTGAATG), 10 μM TSO-random primer and 1.25 U μl−1 Klenow exo- (Enzymatics)). Reactions were incubated with bead-bound samples at 37 °C for 60 min with end-over-end rotation78.To further optimize reaction conditions, we compared two different DNA polymerases (Supplementary Fig. 1a and Supplementary Table 4). First, we tested the Klenow exo- enzyme (Enzymatics) in two different reaction buffers (Klenow Buffer A: see above; 200 μl Buffer B: 1× Blue buffer (Enzymatics), 4% Ficoll PM-400, 1 mM dNTPs (Clontech), 10 μM TSO-random primer and 1.25 U μl−1 Klenow exo- (Enzymatics)). For the Klenow-based system, reactions were incubated for 60 min at 37 °C with rotation. Second, we tested Bst 3.0 DNA polymerase (Bst3, NEB, M0374), which offers strong strand displacement activity and high fidelity (Supplementary Fig. 1d–f). The Bst3 reaction mixture included 1× Isothermal Amplification Buffer II (NEB), 6 mM MgSO4, 4% Ficoll PM-400, 1.4 mM dNTPs (Clontech), 10 μM TSO-random primer and 0.4 U μl−1Bst 3.0. The Bst3-based reactions were incubated on ice for 2 min for primer annealing, followed by incubation at 60 °C for 15 min with rotation.Overall, systematic evaluation identified the combination of TSO-N3G2N4B random primer and Bst 3.0 DNA polymerase as the optimal condition for second-strand synthesis reaction in scNT-seq2. Specifically, compared to the original scNT-seq (TimeLapse + second strand synthesis (2nd SS)/Klenow with N9 random primers), these enhancements in scNT-seq2 substantially improved read alignment rates (Extended Data Fig. 1a), library complexity (Extended Data Fig. 1b) and chemical conversion of 4sU to cytosine analogs (T-to-C mutations in Extended Data Fig. 1c), and reduced background mutations (non-T-to-C mutations in Supplementary Fig. 1f).In vivo scNT-seq2 library preparation and sequencingThe scNT-seq2 library was generated using the custom-built droplet microfluidics-based scRNA-seq platform15, incorporating an extensively optimized second-strand synthesis reaction (see above). The scNT-seq2 workflow consists of the following steps (Fig. 1a):Step 1: droplet-based co-encapsulation of cells and barcodes beads. Single-cell suspensions were diluted to 100 cells μl−1 in DPBS containing 0.01% BSA, and approximately 1.5 ml of this suspension was loaded for each scNT-Seq2 run. Cells were then co-encapsulated with barcoded beads (ChemGenes) using an Aquapel-coated PDMS microfluidic device (μFluidix) connected to syringe pumps (KD Scientific) via polyethylene tubing (inner diameter, 0.38 mm; Scientific Commodities). Barcoded beads were resuspended at 120 beads μl−1 in lysis buffer composed of 200 mM Tris–HCl (pH 8.0), 20 mM EDTA, 6% Ficoll PM-400 (GE Healthcare), 0.2% Sarkosyl (Sigma-Aldrich) and 50 mM DTT (freshly prepared on the day of experiment). Flow rates for both cells and beads were set to 3,200 μl h−1, while QX200 droplet generation oil (Bio-Rad) was run at 12,500 μl h−1.Step 2: TimeLapse chemistry-based conversion of 4sU-labeled new RNAs on pooled beads. Droplets were broken using Perfluoro-1-octanol (Sigma-Aldrich). After droplet breakage, beads were subjected to the on-bead TimeLapse reaction to convert 4sU to cytidine analog51. In brief, 50,000–100,000 beads were washed once with 450 μl wash buffer (1 mM EDTA, 100 mM sodium acetate, pH 5.2), then resuspended in a reaction mixture containing TFEA (600 mM), EDTA (1 mM) and sodium acetate (100 mM, pH 5.2) in water. Sodium periodate (NaIO4) was added to a final concentration of 10 mM, and the reaction was incubated at 45 °C for 1 h with rotation. Beads were washed once with 1 ml TE buffer, followed by incubation in 0.5 ml of reducing buffer (10 mM DTT, 100 mM NaCl, 10 mM Tris pH 7.4, 1 mM EDTA) at 37 °C for 30 min with rotation. Beads were then washed once with 0.3 ml 2× RT buffer.Step 3: RT and Bst3-based second-strand synthesis. For the RT reaction, up to 120,000 beads were resuspended in 200 μl of RT mix consisting of 1× Maxima RT buffer (Thermo Fisher), 4% Ficoll PM-400, 1 mM dNTPs (Clontech), 1 U μl−1 RNase inhibitor, 2.5 μM TSO (AAGCAGTGGTATCAACGCAGAGTGAATrGrGrG)79 and 10 U μl−1 Maxima H Minus Reverse Transcriptase (Thermo Fisher). The RT reaction was incubated at room temperature for 30 min, followed by incubation at 42 °C for 120 min. After exonuclease I treatment, pooled beads were washed once with TE-SDS buffer and twice with TE-TW buffer. Beads were then resuspended in 500 μl of 0.1 M NaOH and incubated at room temperature for 5 min with rotation. The solution was neutralized with 500 μl of 0.2 M Tris–HCl (pH 7.5), followed by additional washes once with TE-TW buffer and once with 10 mM Tris–HCl (pH 8.0).Second-strand cDNA synthesis was carried out using the optimized Bst3-based reaction. Specifically, beads were resuspended in 200 μl of Bst3-based reaction mixture containing 1× Isothermal Amplification Buffer II (NEB), 6 mM MgSO4, 4% Ficoll PM-400, 1.4 mM dNTPs (Clontech), 10 μM TSO-N3G2N4B random primer (AAGCAGTGGTATCAACGCAGAGTGA (N1:25252525)(N1)(N1)GG(N1)(N1)(N1) (N1)(N2: 00333433); N1 represents a mixture of A, C, G and T at a 25:25:25:25 ratio; N2 represents a mixture of A, C, G and T at a 0:33:34:33 ratio) and 0.4 U μl−1Bst3 DNA polymerase (NEB, M0374). Reactions were incubated on ice for 2 min for primer annealing, followed by incubation at 60 °C for 15 min with rotation. Reactions were stopped by washing the beads once with TE-SDS buffer and twice with TE-TW buffer. Downstream steps including cDNA amplification, sample tagmentation and indexing PCR were performed as previously described15.Step 4: cDNA amplification. To determine the optimal number of PCR cycles for cDNA amplification, an aliquot of 6,000 beads was initially amplified by PCR in a 50-μl reaction (25 μl of 2× KAPA HiFi hotstart readymix (KAPA biosystems), 0.4 μl of 100 μM TSO-PCR primer (AAGCAGTGGTATCAACGCAGAGT), 24.6 μl of nuclease-free water) using the following thermal cycling parameter: 95 °C for 3 min; 4 cycles of 98 °C for 20 s, 65 °C for 45 s, 72 °C for 3 min; 9 cycles of 98 °C for 20 s, 67 °C for 45 s, 72 °C for 3 min; a final extension at 72 °C for 5 min, hold at 4 °C. After two rounds of purification with 0.6× SPRISelect beads (Beckman Coulter), amplified cDNA was eluted with 10 μl water. To assess amplification efficiency, 10% of amplified cDNA was analyzed via real-time qPCR using Applied Biosystems QuantStudio 7 Flex with a 10 μl qPCR reaction (1 μl of purified cDNA, 0.2 μl of 25 μM TSO-PCR primer, 5 μl of 2× KAPA FAST qPCR readymix and 3.8 μl of water). Based on this, the remaining beads from each sample were divided into multiple cDNA amplification reactions (~6,000 beads per 50 μl reaction) and amplified for (4 + 9–12) cycles to enrich cDNA.Step 5: final library construction and sequencing. After cDNA amplification, 1 ng of cDNA (pooled in equal amounts from all reactions) was tagmented using the Nextera XT DNA Sample preparation kit (Illumina, FC-131-1096). The tagmented samples were enriched with 12 PCR cycles using the Nextera XT i7 primers and the P5-TSO hybrid primer79. Final libraries were assessed using a Bioanalyzer (Agilent) and sequenced on a NextSeq 500 using the 75-cycle High Output v2.5 Kit (Illumina). Specifically, libraries were loaded at 2.0 pM with 0.3 μM Custom Read1 Primer (GCCTGTCCGCGGAAGCAGTGGTATCAACGCAGAGTAC) added to position 7 of the reagent cartridge. Sequencing was performed in the following configuration: 20 bp (Read1), 8 bp (Index1) and 60 bp (Read2).Optimization of in situ chemical reaction for spatial NT-seq using human K562 cellsAs the SLAM reaction requires organic solvents53, we reasoned that it could be adapted for in situ chemical conversion in methanol-fixed cells or tissue sections and made compatible with various sequencing-based spatial transcriptomics platforms. Since previous applications of SLAM reaction reported relatively low conversion efficiency (~0.5–1% T-to-C substitutions) in cultured cells76,80, we first systematically optimized key parameters of in situ SLAM reaction (for example, iodoacetamide (IAA) concentration and reaction temperature/time) using in vitro cultured cells (Supplementary Fig. 5a).To this end, human K562 cells were metabolically labeled with 4sU and subsequently fixed as described above. We first optimized in situ SLAM reaction based on previously reported protocols53,81, benchmarking its performance against the established on-bead TimeLapse chemistry. Specifically, chemical conversion was tested directly in the fixation solution (80% methanol and 20% DPBS) using varying concentrations of IAA (ranging from 10 to 40 mM), applied either overnight at room temperature or at 50 °C. For the 50 °C condition, the reaction was pre-incubated on ice for 15 min before incubation at 50 °C with rotation. Following the reaction, cells were rinsed with 1 ml cold fixation solution (80% methanol and 20% DPBS) and incubated for 5 min at room temperature in 1 ml quenching buffer (0.1% BSA, 1 U µl−1 RNase-Inhibitor, 100 mM DTT in DPBS). Library preparation proceeded using the scNT-seq2 protocol (omitting the on-bead conversion step). For samples treated at 50 °C, the Bst3-based second-strand synthesis reaction was incorporated following RT and exonuclease I treatment.Using on-bead TimeLapse chemistry used in scNT-seq2 as a benchmark, we systematically optimized key parameters of the in situ SLAM reaction (for example, IAA concentration and reaction temperature/time), achieving markedly higher conversion rates (in situ SLAM versus on-bead Timelapse based 4sU conversion rates −5.35% versus 5.71%) without compromising performance (Supplementary Fig. 5b–e).In situ chemical reaction and spatial NT-seq library preparationTo develop spatial NT-seq on a widely adopted spatial transcriptomics platform, we integrated the optimized in situ chemical conversion reaction with the Visium platform from 10x Genomics (CG000239, Rev D). Two major modifications were made to the standard Visium workflow: (1) incorporation of the optimized in situ SLAM reaction before the ‘tissue permeabilization’ step and (2) replacement of the original ‘reverse transcription master mix’ and ‘second-strand synthesis’ reagents with the ‘Maxima H minus reverse transcriptase mix’ and optimized Bst3-based second-strand synthesis reaction mix, respectively, as used in scNT-seq2 (see below for details). While this study focused on the Visium platform, these modifications are platform agnostic and can be adapted to other sequencing-based spatial transcriptomics platforms. The spatial NT-seq workflow consists of the following steps (Fig. 2a):Step 1: tissue sample preparation. Following in vivo metabolic RNA labeling (see above), mice were euthanized and brains were dissected, embedded in OCT compound (Tissue-Tek) and stored at −80 °C.Step 2: cryosection followed by H&E staining. OCT-embedded frozen brain hemispheres were cryo-sectioned (10-μm thickness) at −20 °C using a cryostat (Leica, CM3050S). Methanol fixation, H&E staining and tissue imaging were performed according to the Visium user guide (10x Genomics, CG000160, Rev C) using the EVOS FL Auto 2 Cell Imaging System to generate tissue image files for downstream analysis.Step 3: in situ chemical conversion. The Visium Gene Expression Slide containing tissue sections were placed into the slide cassette and washed once with 100 μl fixation buffer (80% methanol, 20% DPBS) per well. The fixation buffer was then replaced with 80 μl of SLAM reaction mix (fixation buffer supplemented with 10 mM IAA), followed by incubation on ice for 15 min and then at 50 °C for 15 min in a thermal cycler. The reaction was quenched by removing the IAA solution, washing once with 100 μl fixation buffer, and adding 100 μl quenching buffer (DPBS with 0.1% BSA, 1 U µl−1 RNase Inhibitor and 100 mM DTT), followed by incubation at room temperature for 5 min. Although the TimeLapse reaction was also tested, it resulted in markedly lower conversion efficiency and library complexity compared to SLAM.Step 4: permeabilization and RT. Permeabilization and RT steps were performed immediately following the in situ SLAM reaction. Permeabilization time was first optimized with the Visium Tissue Optimization kit (12–18 min, 37 °C). To enhance the T-to-C conversion efficiency, the original ‘RT master mix’ was replaced with the scNT-seq2 RT mix: 1× Maxima RT buffer (Thermo Fisher), 4% Ficoll PM-400, 1 mM dNTPs (Clontech), 1 U μl−1 RNase inhibitor, 2.5 μM TSO (AAGCAGTGGTATCAACGCAGAGTGAATrGrGrG) and 10 U μl−1 Maxima H minus reverse transcriptase (Thermo Fisher). The RT reaction was incubated at room temperature for 30 min, followed by incubation at 42 °C for 120 min.Step 5: second-strand synthesis. The standard second-strand synthesis reaction mix was replaced with a Bst3.0-based reaction mixture: 1× Isothermal Amplification Buffer II (NEB), 6 mM MgSO4, 4% Ficoll PM-400, 1.4 mM dNTPs (Clontech), 10 μM random primer (TSO-N3G2N4B: AAGCAGTGGTATCAACGCAGAGTGA(N1:25252525)(N1)(N1)GG(N1)(N1)(N1)(N1)(N2: 00333433); N1 represents a mixture of A, T, C and G at a 25:25:25:25 ratio, N2 represents a mixture of A, T, C and G at a 0:33:34:33 ratio), and 0.4 U μl−1Bst 3.0 DNA polymerase (NEB, M0374). After adding 75 μl reaction mix per well, the slide was incubated on ice for 2 min, followed by incubation at 60 °C for 30 min in a thermal cycler. The second-strand cDNA molecules were eluted using 35 μl 0.1 M NaOH and neutralized with 5 μl 1 M Tris–HCl (pH 7.0) before cDNA amplification.Step 6: cDNA amplification. To determine the optimal number of PCR cycles for amplification, 1 μl of cDNA was used to perform real-time qPCR analysis (1 μl of cDNA, 0.2 μl of 25 μM TSO-PCR primer (AAGCAGTGGTATCAACGCAGAGT), 0.3 μl of cDNA primers (10x Genomics), 5 μl of 2× KAPA FAST qPCR readymix and 3.5 μl of water) using the QuantStudio 7 Flex (Applied Biosystems). The remaining cDNA was amplified in a 100 μl reaction containing 35 μl of cDNA, 50 μl of 2× Amp Mix (10x Genomics), 3 μl of 25 μM TSO-PCR primer and 12 μl of cDNA primers (10x Genomics). Thermal cycling conditions were 98 °C for 3 min; 12–13 cycles of 98 °C for 15 s, 63 °C for 20 s, 72 °C for 1 min; 72 °C for 1 min, hold at 4 °C.Step 7: library construction and sequencing. Final library construction was performed following the standard Visium workflow, including fragmentation, end repair and A-tailing, adapter ligation and index PCR. Libraries were quality checked using an Agilent Bioanalyzer and sequenced on an Illumina NextSeq 500 instrument using the 150-cycle High Output v2 or v2.5 Kit (Illumina). Libraries were loaded at 2.0 pM and sequencing was performed with the following configuration: 28 bp (Read1), 10 bp (Index1), 10 bp (Index2) and 110 bp (Read2).scNT-seq2 data preprocessingPreprocessing of scNT-seq2 datasets, including read alignment, quality filtering and count matrix generation, was performed utilizing the ‘Dynast’ pipeline (v1.0.1 (ref. 82)), an inclusive and efficient toolkit for processing metabolic labeling-based scRNA-seq datasets. Modifications were made to accommodate the droplet-based scNT-seq2 workflow. The pipeline comprises four main steps: ‘dynast align’, ‘dynast consensus’, ‘dynast count’ and ‘dynast estimate’.First, raw sequencing reads were aligned to the mouse reference genome (GRCm39) using ‘dynast align’ with paired-end mode (that is, ‘--strand forward -x dropseq’). High-quality cells were selected with the parameter ‘--soloCellFilter TopCells’ from ‘STARsolo’ to generate a cell-barcode whitelist for downstream analysis. Second, ‘dynast consensus’ constructed a consensus sequence for each transcript by pooling reads with the same UMI index and selecting the most frequent variant at each position as previously described15. ‘dynast count’ was then used to quantify unspliced, spliced, newly synthesized and total RNAs for each cell. Moreover, sites with background T-to-C substitutions were identified from control samples without chemical conversion and excluded for further analysis. Finally, ‘dynast estimate’ applied a UMI-based binomial mixture model to statistically infer the new RNA fractions, correcting for incomplete new RNA labeling. To improve computational efficiency, we used the alpha correction model (that is, ‘--method alpha’), as previously described15. The corrected count matrices of new and old RNAs were generated by ‘dynast estimate’.Cell-type clustering and marker gene identification for scNT-seq2All downstream data preprocessing, clustering and differential expression analyses were performed using Scanpy (v1.9.1). The total RNA count matrices were imported into Scanpy as AnnData objects. The analysis pipeline for scNT-seq2 datasets consists of the following main steps.Quality controlFor postnatal (P7–12) CTX datasets, cells with fewer than 400 or more than 8,000 genes, or with >10% mitochondrial gene content, were excluded. For the E16.5 dataset, cells with fewer than 400 or more than 6,000 genes, or >10% mitochondrial gene content were removed. In all scNT-seq2 datasets, genes expressed in 500 UMIs) were retained, and the new RNA fraction was calculated as the ratio of newly synthesized RNA counts to total RNA counts. The resulting distributions were visualized using box plots (Fig. 1f and Extended Data Fig. 2f).To quantify cell-type-specific new RNA fractions (Fig. 1c and Extended Data Fig. 2e), cells underwent additional filtering and clustering as described above, after which the percentage of newly synthesized RNA was calculated and visualized for each annotated cell type.Spatial NT-seq data preprocessingPreprocessing of Visium platform-based spatial NT-seq data was first conducted using ‘Space Ranger’ (v2.0.0, 10x Genomics), including read alignment, quality filtering and generation of total RNA count matrices for each spatially barcoded spot. Reads were aligned to the mouse reference genome (GRCm39), and both exonic and intronic reads were incorporated into count matrices. Histological H&E images were processed using ‘Space Ranger’ to identify regions covered by tissue sections, by aligning prerecorded spot locations with fiducial border spots in the image.Subsequent preprocessing steps of spatial NT-seq datasets were performed using the ‘Dynast’ pipeline (v1.0.1) as described for the scNT-seq2 dataset. Tailored for the 10x Genomics platform, raw sequencing reads were aligned to the reference genome GRCm39 using ‘dynast align’ in paired-end mode (that is, ‘-x 10xv3’). Spatial spots were filtered using a barcode whitelist generated by ‘Space Ranger’ to retain high-quality spatial spots for downstream analyses. Estimations of new and old RNAs were obtained as described in the scNT-seq2 preprocessing workflow.Identification and annotation of spatial clusters for spatial NT-seqFor spatial NT-seq data analysis, total RNA count matrices generated by ‘Space Ranger’ were first imported into Scanpy (v1.9.1) as AnnData objects. All downstream data preprocessing, clustering and differential expression analyses were performed using Scanpy. The preprocessing of Spatial NT-seq datasets consists of the following main steps.For P7 mouse brains, spatial spots with less than 200 UMIs or >15% mitochondrial gene content were filtered out. Genes expressed in fewer than ten spots were also removed. For P56 (ECS) samples, spatial spots with less than 100 UMIs or >30% mitochondrial gene content were removed, and genes expressed in fewer than three spots were also removed. Highly variable genes were identified using default settings in Scanpy, with the top 3,000 selected for P7 samples and the top 2,000 selected for P56 ECS samples. Batch corrections across multiple samples were performed using the bbknn algorithm in Scanpy. Dimensionality reduction was carried out via PCA analysis, retaining the top 50 components for downstream clustering and visualization using UMAP plots.Spatial clusters identified through this workflow were mapped back to their original anatomical locations using Visium spatial barcodes and annotated based on the Allen common coordinate framework. The anatomical identity of each spatial cluster was further validated by comparison with a published snRNA-seq whole mouse brain reference dataset using the Cell2location algorithm (see below).Spatial mapping and validation of cell types in spatial NT-seq data with Cell2locationTo validate the annotation of spatial clusters identified in the spatial NT-seq datasets (Supplementary Fig. 6a, P7 and Extended Data Fig. 5b, P56), we employed ‘cell2location’ (v0.1.3 (ref. 83)) to spatially map cell types in the postnatal mouse brains based on gene expression profiles derived from a published snRNA-seq dataset of adult wild-type whole brains (P56, C57BL/6 mice)30. Cell-type reference signatures from the snRNA-seq dataset were computed using regularized negative binomial regressions, with estimated expression levels calculated as the average value in each cluster. For integrated analysis, raw counts from the spatial NT-seq data were used as input, and genes were filtered to exclude mitochondrial genes and to retain only genes shared between the snRNA-seq reference and spatial NT-seq datasets. The cell2location spatial mapping model was trained to infer cell type proportions and spatial gene expression patterns for each section. The model was configured with the default settings except for two hyperparameters: ‘N_cells_per_location=5’ and ‘detection_alpha=20’ as specified in the ‘cell2location.models.Cell2location’ function. This allowed for decomposition of the multicellular RNA count matrix in each spot into reference-derived cell-type signatures, enabling spatially resolved mapping of cell types in mouse brain sections. For visualization, we used the top 5% quantile of the posterior distribution as the high-confidence cell abundance score.Transcriptome-wide analysis of the impact of metabolic RNA labeling on global gene expressionTo assess the global effects of 4sU-mediated metabolic labeling on gene expression, we used a previously established approach26. First, we calculated the NTR using pseudo-aggregated new and total RNA counts from each experiment. When both the labeled and control groups included only a single replicate, the NTR was computed directly as the ratio of new to total RNA for that sample. In cases with multiple replicates in the labeling group, the mean NTR was computed across replicates. To ensure robust downstream analysis, genes with low expression levels (10% of cells/spots within a specific cell or spatial cluster, RNA half-life imputation was performed using a kNN approach (that is, the ‘KNNImputer’ function from scikit-learn, v1.1.1) to fill in missing values. Specifically, missing half-life values were imputed by considering the group mean derived from five nearest neighbors with available estimates.Estimation of RNA synthesis and degradation rates at the single-cell/-spot levelUsing the RNA half-life and cell/spot matrix as the input, we calculated the RNA synthesis rate (α, CP10k min−1) and RNA degradation rate (γ, 1 min−1) with the following ordinary differential equation:The degradation rate constant (γ, units min−1) can be calculated from RNA half-life (t1/2) using$${{\gamma }}=\frac{\mathrm{ln}\left(2\right)}{{{{t}}}_{1/2}}$$Then, we assumed the gene-specific RNA biogenesis rate (α, molecules min−1) is a constant for all cells from each cell state, which can then be calculated using$$\alpha =\frac{n\gamma }{1-{{\rm{e}}}^{-\gamma t}},$$where n is the average labeled RNA abundance for each gene in each single cell/spot, γ is the degradation rate constant in each single cell/spot and t (min) is the metabolic labeling time.Bootstrap estimation of RNA half-life uncertaintyTo quantify uncertainty in half-life estimates (Supplementary Fig. 11f), we applied a parametric bootstrap procedure to simulate repeated experiments. For each gene, labeled (L) and unlabeled (U) read counts were observed, with total counts n = L + U and fraction labeled f = L/n. We modeled the number of labeled reads as a binomial random variable, L* ~ binomial(n, f), and defined U* = n − L*. Using this model, we generated 2,000 bootstrap replicates of (L*, U*), each representing a plausible experimental outcome under the same sampling conditions. For each replicate, the half-life was recalculated using the same formula as for the observed data. The distribution of bootstrap half-lives was then used to obtain the point estimate (from the observed counts) and a 95% confidence interval (2.5–97.5th percentile). This simulation-based approach captures the stochastic sampling of labeled and unlabeled reads and provides an approximate measure of uncertainty in half-life estimates.Predictive modeling of the post-transcriptional regulome of RNA stabilityRecent advances in machine learning have enabled the predictive modeling analysis of mRNA stability regulatory landscapes using transcriptome-wide RNA half-life data58,59 (Fig. 5a), but previous studies have been limited to in vitro cultured cells. To model the regulatory effects of sequence-dependent features and post-transcriptional regulators (for example, miRNAs and RBPs) on in vivo RNA stability, we implemented a multistep machine learning pipeline, as previously described with modifications58. This modified pipeline enabled predictive modeling of RNA half-life at both the cluster (cell-type or spatial domain) level and the single-cell/spatial-spot level.To systematically investigate how various genomic feature sets relate to transcriptome-wide RNA half-life values across cell types, we employed a LASSO regression model using the glmnet (v4.1-4) package in R. LASSO regression applies an L1 regularization penalty, which is known for selecting a minimal set of features that maximally explain the observed data. All input features considered in the model were z-score normalized and concatenated together. Gene-level feature sets were curated as previously reported58. The regularization parameter (denoted as λ) in the LASSO regression model controls the strength of the penalty applied to the coefficients. To identify the optimal λ value, tenfold cross-validation was performed with different λ values. Model performance was evaluated by calculating Pearson’s correlation coefficients between predicted and observed RNA half-life values across folds. We also used paired t-tests to determine significant improvements in performance attributable to specific feature subsets. After identifying the best feature sets, the LASSO regression model was retrained on the full dataset, and nonzero coefficients were extracted to identify the most informative features. These features were organized into a data frame with corresponding effect sizes, associated cell types/spatial regions and feature categories. The top-ranked coefficients within each feature class were visualized to highlight major contributions to RNA stability across cellular and spatial contexts.First, we established the best performing model for collectively explaining in vivo cell-type-specific RNA half-life data with feature sets consisting of basic mRNA features, codon frequencies, 3′ UTR k-mer motifs, miRNA target scores87 and RBP binding scores11,88. This analysis established the following feature set as the optimal configuration for collectively explaining in vivo RNA half-life: basic mRNA features (‘B’, n = 8), codon frequencies (‘C’, n = 61), 3′ UTR k-mer motifs (‘3’, n = 21,844), predicted repression scores of miRNA (‘M’, n = 315) and predicted RBP binding scores by SeqWeaver (‘S’, n = 780) or DeepRiPE (‘D’, n = 177). The choice of feature sets in our final model achieved an optimal balance between model performance and feature complexity, aligning with previous machine learning modeling analysis of ‘ensemble’ RNA half-life values from in vitro cultured mouse cells58.Second, the performance of trained models with various feature sets (that is, measured by Pearson’s correlation coefficients between predicted and observed RNA half-life) was systematically validated using held-out data and the tenfold cross-validation strategy in both single-cell (Supplementary Fig. 15b) and spatial NT-seq (Extended Data Fig. 9a) datasets. Specifically, we applied the model to cell-type-specific measurements of mRNA half-life across 15 major cortical neuronal and non-neuronal cell-types (Pearson’s correlation coefficients: mean ± s.d. of 0.46 ± 0.08; Supplementary Fig. 15c). In addition, we evaluated the performance of the model in predicting in vivo RNA half-life of 22 spatial brain regions (0.44 ± 0.02; Extended Data Fig. 9b), which is comparable to that of single-cell analysis. Importantly, the regulatory effects of top-ranked features are concordant between two replicates across 22 spatial brain regions (Supplementary Fig. 16a).Finally, to assess the robustness of our model in RNA half-life prediction, we systematically evaluated it across various metabolic labeling-based RNA half-life analysis experiments (Supplementary Fig. 15d–g). Specifically, the trained model performs comparably across labeling strategies (one-shot versus pulse-chase; Supplementary Fig. 15d), sequencing methods (single-cell versus bulk RNA-seq; Supplementary Fig. 15d), sample origin (in vitro versus in vivo 4sU labeling; Supplementary Fig. 15e), labeling durations (2 h versus 4 h; Supplementary Fig. 15f) and is highly reproducible across biological replicates (Supplementary Fig. 15g).For cluster-level modeling analysis, we used RNA half-life values derived from pseudo-bulk samples aggregated across all cells or spatial spots for each cell type or spatial domain. LASSO regression models were trained independently for each cluster using matched gene-wide feature and RNA half-life data. Genes with infinite or missing half-life values were excluded, and RNA half-life values between the 5th and 95th percentile range were retained to mitigate the influence of outliers. RNA half-life values in each cell-type or spatial domain were z-score normalized.To extend this approach to single-cell/-spot resolution, we applied the same modeling pipeline, using neighborhood-level RNA half-life values estimated from the RNAKinetoScope as input. Feature sets and preprocessing steps were identical to those used for cluster-level analysis. This enabled predictive modeling of RNA stability at the level of individual cells or spatial spots and identification of local regulatory influences on RNA stability.Predictive modeling of RBP-mediated regulation of in vivo RNA stabilityUtilizing LASSO regression-based predictions to assess the regulatory effects of a diverse array of RBPs on mRNA stability, we uncovered a total of 122 distinct RBP binding features, exhibiting cell-type specific patterns (Supplementary Fig. 18a). Specifically, we examined the effect coefficients pertaining to context-specific RBP bindings localized in the 5′ UTR, ORF and 3′ UTR regions, as curated by deep learning-based computational models trained on experimental RBP binding datasets11,88. In contrast to miRNAs, which predominantly exert a negative influence on mRNA stability, our analysis revealed that 55 RBP binding features (constituting 45% of context-specific RBP binding features) exhibited positive effects on mRNA half-life (indicated by red dots in the left panel in Supplementary Fig. 18a), while 67 features (55%) showed negative impacts on mRNA half-life (depicted by blue dots in the left panel in Supplementary Fig. 18a). This is consistent with the notion that RBPs can elicit both stabilization and destabilization of target mRNAs, with these specific interactions influenced by various signaling cascades35,58,64. Moreover, spatially resolved RNA half-life analysis showed that, among 167 RBP binding features, 46% are associated with mRNA stabilization (red dots in Extended Data Fig. 10a), while the remainder exhibited negative impacts on mRNA stability (blue dots in Extended Data Fig. 10a).To validate the predicted regulatory effects of RBPs, we examined RBPs with known roles in regulation of mRNA stability. Predicted effects of well-characterized RBPs aligned with findings from in vitro cultured cells, including mRNA stabilizing RBPs such as PCBP289, hnRNP-K90, hnRNP-U91 and destabilizing RBPs including PUM292, QKI93 and TIA1/TIAL194. Among top-ranked predicted RBP features, eight out of nine features, encompassing either destabilizing (FMRiso1-3′UTR, PUM2-3′UTR, srsf4-ORF, rbm15-3′UTR, mbnl-5′UTR and FXR2-3′UTR) or stabilizing (safb2-5′UTR, IGF2BP3-3′UTR) RBPs were consistent with prior data from cultured mouse cells58. Furthermore, predicted in vivo regulatory effects of five out of seven overlapping RBPs (destabilizing, DDX6, LARP4 and RBM15; stabilizing, CPSF6 and IGF2BP3) aligned with results derived from RBP knockdown experiments in cultured HepG2 and K562 cells64. The discrepancies probably reflect cell-type-specific regulation or methodological differences (bulk RNA-seq analysis of knockdown cell lines versus single-cell metabolic labeling RNA-seq). Thus, these results demonstrate the validity of our integrated experimental and computational approach in mapping the landscape of RBP-mediated post-transcriptional regulation of in vivo mRNA stability.Differential gene expression analysis for ECSTo identify DEGs across ECS time points, we performed pairwise comparisons for each spatial cluster using the Wilcoxon rank-sum test implemented in Scanpy (v1.9.1), based on both newly synthesized (‘new’) and pre-existing (‘old’) RNAs. Distinct thresholds were applied: for Fig. 3d, DEGs were defined as those with an adjusted P value 4; for Extended Data Fig. 5d, the fold change cutoff was relaxed to >2. To capture temporal gene expression dynamics across the ECS time course, we applied unsupervised clustering analysis of new and old RNAs (CPM) using the TCseq R package (v1.22.6). Gene expression patterns of new and old RNAs were grouped into six clusters (k = 6) by soft clustering (fuzzy c-means). Our analysis focused on gene clusters exhibiting transient upregulation or downregulation at the level of either new or old transcripts.Benchmarking RNA labeling- versus splicing-based analysis of spatial IEG dynamicsTo quantitatively benchmark RNA labeling- versus splicing-based detection of neuronal activity-induced IEG signatures across spatial spots in spatial NT-seq datasets (activated by ECS at 30 min), we adapted ‘in silico trapping’, a computational approach designed to identify cells probably activated by a stimulus from single-cell RNA-seq snapshot data44. Per-spot IEG scores were calculated from a curated set of genes (Arc, Bdnf, Btg2, Fos, Fosl2, Homer1 and Npas4). For each spatial spot, counts from total, unspliced and new RNA matrices were normalized to CP10k, and the IEG score was defined as the mean normalized expression of the eight-gene panel. Spatial spots with the top 5% of total RNA IEG scores were designated as ‘trapped’ (that is, IEGhigh spots) and used as the reference for benchmarking. Predictor scores were similarly defined from unspliced and new RNA IEG data. The ability of each predictor to recover reference-trapped spatial spots was quantified using precision–recall curves (precision_recall_curve) and summarized by average precision (average_precision_score), with a random baseline defined by the fraction of reference-trapped spots. To assess overlap and divergence across modalities, the 5% threshold was applied independently to total, unspliced and new RNA IEG scores. Logical combinations were categorized as ‘total RNA only’, ‘unspliced RNA only’, ‘new RNA only’, ‘total & unspliced RNA (overlapping)’ and ‘total & new RNA (overlapping)’. These categories were visualized on UMAP embeddings and summarized in bar plots to evaluate how well new and unspliced RNA recapitulate the ECS-activated spots defined by total RNA (Supplementary Fig. 8b–e).Comparison of RNA and protein turnover in neurons and gliaTo better understand whether and how RNA and protein turnover are coordinated in specific cell types, we performed additional analyses comparing RNA and protein half-lives in neurons and glia using data from a published study95.First, we demonstrated that RNA half-lives in Ex neurons and astrocytes (from P7 mouse CTX) are moderately correlated (Pearson’s correlation coefficient R = 0.54, P  4). Differentially expressed genes were identified using a two-sided Wilcoxon rank-sum test. P values were adjusted for multiple comparisons.Source dataExtended Data Fig. 6 High temporal and spatial resolution profiling of ECS-induced ARG dynamics using spatial NT-seq and computational modeling.a. Panels displaying the normalized spatial expression (in CP10K) of new and old RNAs for one representative early ARG gene, Fos, across three time points (ECS control, 30-, and 120-min) from two sections (#1-2) of two biologically independent cohorts of mice (top: mouse #1/3/5; bottom: mouse #2/4/6). b. UMAP visualization (top) of metabolic labeling-based RNA velocity analysis of early ARG activation in response to acute ECS (n = 227 spatial spots in DG). Shown in the bottom are normalized expression (total RNA levels, in CP10K) of two early ARGs, Fos and Egr1. UMAP visualization of early ARG activation in response to acute ECS (n = 227 spatial spots in DG from two biologically independent cohort of mice) that are colored by experimental samples. c. Model summary of modified iStar (inferring Super-resolution Tissue Architecture)46 workflow for near single-cell resolution (pixel size: 8 μm) analysis of newly-synthesized and pre-existing RNA expression derived from spot-level spatial NT-seq data (spot diameter: 55 μm). d. Panels displaying the normalized near single-cell resolution (pixel size: 8 μm) spatial expression (in CP10K for input spot-level new and old RNA expression data) of new (upper panel) and old (lower panel) RNAs for four representative genes (same as in Fig. 3g) from each ARG group using a representative spatial NT-seq dataset (section #2 from mouse #2/4/6).Extended Data Fig. 7 In vivo profiling of spatial RNA kinetics landscapes in the mouse brain.a. Scatter plot showing transcriptome-wide RNA synthesis (x-axis: CPM/min) versus degradation (y-axis: min−1) rates in DG (left) versus non-DG regions (right) of P7 mouse brains. Total RNA levels are represented by dot size and colors (log2-scaled). b. Scatter plot showing RNA synthesis (x-axis: CPM/min) versus degradation (y-axis: min-1) rates in three brain regions (P7) for genes in kinetics modules #3 (red: Stxbp1, Nrgn, Map4, Snca, and Cd24a) and #13 (blue: Pou3f1, Plppr4, Peg3, Satb1, and Rorb). Five representative neurite-localized/enriched genes (black: Actb, Mapt, Rps14, Tpt1, and Ybx1) are also highlighted. Total RNA levels are represented by dot size. c. Scatter plot showing RNA synthesis (x-axis: CPM/min) versus degradation (y-axis: min−1) rates for transcriptional regulators (yellow; CPM > 20) and neurite-localized/enriched genes56 (black) in three representative brain regions (P7). Total RNA levels are represented by dot size. d. Box plots comparing spatial region-specific RNA synthesis rates (alpha, left), degradation rates (gamma, middle) and total RNA abundance (right) for transcriptional regulators (yellow; CPM > 20) and neurite-localized/enriched genes (black) in five representative brain regions (P7). The mean for each is shown as a white dot, with its value indicated above the box. P-values from two-sided Welch’s two-sample t-test comparing transcriptional regulators and neurite-localized genes are reported above each boxplot. Numbers of genes: neurite-localized genes −101 genes; TF −659 genes. The same gene lists were used for all spatial regions. Exact P values are provided in Source Data for Extended Data Fig. 7e (not adjusted for multiple comparisons). For boxplots, it shows the median as the centre line, the 25th and 75th percentiles as box bounds, and whiskers extending to the smallest and largest values within 1.5× the interquartile range. The group mean is shown as a dot, with its value indicated above the box. e. Heatmaps showing statistical analysis of spatial region-specific total RNA abundance, RNA degradation rates (gamma), and RNA synthesis rates (alpha) between the DG versus representative non-DG brain regions of P7 mouse brains for neurite-localized genes and transcriptional regulators. Statistical significance was assessed using two-sided Welch’s two-sample t-test comparing DG with other spatial regions (not adjusted for multiple comparisons). In the heatmap, significance is indicated as follows: ***, P