Spatially resolved single-cell translatomics at molecular resolution

RIBOmap design and validation

RIBOmap is built upon a targeted-sequencing strategy in which a specific design of tri-probes selectively detects and amplifies ribosome-bound mRNAs (Fig. 1A and fig. S1A). The RIBOmap tri-probe set includes: (i) a splint DNA probe that hybridizes to ribosomal RNAs (rRNAs) and serves as the splint to circularize the adjacent padlock probe; (ii) a padlock probe that targets specific mRNA species of interest and encodes a gene-unique barcode; (iii) a primer probe that targets the mRNA site adjacent to the one targeted by the padlock probe and serves as the primer for rolling circle amplification resulting in a DNA nanoball (amplicon). DNA amplicon signals are only produced when all three probes are present in proximity (Fig. 1, B and C). The gene-unique barcodes in the DNA amplicons are then decoded through in situ sequencing with error-reduction by dynamic annealing and ligation (Fig. 1A) (20). We also tested an alternative RIBOmap workflow that uses primary antibodies of ribosomal proteins (RPS3, RPL4) and splint-conjugated protein A/G to target ribosomes, which demonstrated a much lower signal-to-noise ratio (SNR) than that of the rRNA targeting strategy (fig. S1, B to G). Thus, we proceeded with the rRNA targeting strategy in the following studies.

To confirm that RIBOmap specifically targets ribosome-bound mRNAs, we designed four experiments. First, RIBOmap and previously reported STARmap RNA imaging (20) were used to detect protein-encoding ACTB mRNA and two negative controls of noncoding RNAs (Fig. 2A and table S1). As expected, RIBOmap only detected cytoplasmic ACTB mRNAs whereas STARmap detected both ACTB mRNA and the noncoding RNAs (Fig. 2, B and C and fig. S2). Second, RIBOmap specificity was validated using Harringtonine, a translation inhibitor that traps translation-initiating ribosomes at the start codon (21) (Fig. 2D). After Harringtonine treatment, there was a significant decrease in RIBOmap signal intensity using probes targeting 115 or 405 nucelotides (nt) downstream of the start codon, whereas there was minimal decrease in the signal generated by the probe targeting the −16 nt from the start codon (Fig. 2, E to G). Third, we used synthetic in vitro transcribed (IVT) mRNAs and lipid-mediated transfection (Fig. 2H). Our data demonstrated that STARmap could detect both small puncta corresponding to free cytosolic mRNAs and larger intracellular granules corresponding to lipid transfection vesicles containing many copies of nontranslating mRNAs, whereas RIBOmap signals were depleted in lipid transfection vesicles (Fig. 2, I and J), indicating that RIBOmap does not detect nontranslating mRNAs. Finally, we validated RIBOmap by comparing it with RiboLace, an established ribosome profiling technology (22). RIBOmap successfully detected differentially translated mRNAs in starved versus control human MCF7 cells, which is consistent with the reported RiboLace and proteomics measurements (22) (fig. S3). Together, these results demonstrate the specificity of RIBOmap in detecting ribosome-bound mRNAs.

Multiplexed RIBOmap in cultured cells

After benchmarking the SNR and specificity of RIBOmap, we performed a highly multiplexed 981-gene RIBOmap experiment in HeLa cells (Fig. 3A). The 981 genes represent a curated list composed of cell-cycle gene markers, genes with diverse subcellular RNA patterns, and genes of varying RNA stabilities (23–26) (table S2). We designed a multimodal RIBOmap imaging experiment that further incorporated the information on cell-cycle stages and subcellular organelles: (i) the cell-cycle phase of each cell was captured by the fluorescent ubiquitination-based cell cycle indicators (FUCCI) (27, 28); (ii) ribosome-bound mRNAs (981 genes) were in situ sequenced; (iii) finally, nuclei, endoplasmic reticulum, and cell shapes were stained and imaged (Fig. 3B). We also conducted a paired STARmap experiment for comparison (Fig. 3A). In total, RIBOmap and STARmap sequenced 1813 and 1757 cells, respectively (fig. S4A). The median distance between each read and its nearest neighbor in 3D was 0.69 mm for RIBOmap and 0.61 mm for STARmap (fig. S4, B and C), indicating that both techniques have high spatial resolution. To assess the reliability of RIBOmap in multiplexed experiments, we calculated the correlation between the RIBOmap results for 981 genes with published HeLa proteome and ribosome profiling datasets (26, 29, 30). The results showed that the correlation between RIBOmap and the proteome dataset is comparable to that between the ribosome profiling dataset and the proteome dataset (fig. S4D). Additionally, the correlation between RIBOmap and the two ribosome profiling datasets is comparable to the correlation between the two ribosome profiling datasets (fig. S4E).

Next, we evaluated whether RIBOmap can decipher cell cycle–dependent mRNA translation. The results showed that the G1, G1/S, and G2/M cell cycle phases identified from both RIBOmap and STARmap datasets agreed with the expected cell cycle–dependent patterns of FUCCI protein fluorescence (Fig. 3C and fig. S5), demonstrating the accuracy of RIBOmap in delineating cell states. We then identified the differentially expressed genes (DEGs) across cell cycle phases using RIBOmap and STARmap data (fig. S6, A and B, and table S3). Most of the RIBOmap DEGs overlapped with known cell cycle–dependent genes (31) (fig. S6C). We also identified DEGs detected in RIBOmap but not in STARmap (fig. S6D). One example is NOL6 (nucleolar protein 6), whose translation was significantly decreased from the G1 phase to the G2/M phase as detected by both RIBOmap and ribosome profiling (32), but whose overall RNA level showed little change across cell cycle phases as detected by STARmap and RNA-seq (fig. S6, E to G).

Next, to leverage the single-cell resolution in our dataset, we performed gene-expression covariation analysis and identified five coregulated translation modules (RTMs) with substantial intramodule correlation, each with enrichment of distinct functional pathways (RTMs 1 to 5, Fig. 3D, fig. S7, A to E, and table S4). RTM 3 and RTM 5 are enriched for G2/M and G1/S cell cycle marker genes, respectively (Fig. 3D, right), and are negatively correlated (fig. S7F). Moreover, RTM 2 contains genes encoding protein translation machineries (fig. S7D); these genes show a positive correlation with RTM 5 (G1/S marker genes) and a negative correlation with RTM 3 (G2/M marker genes) (fig. S7G). This observation suggests that the protein translation machinery is up-regulated in the G1/S phase to support the demand for protein products as cells enlarge and subcellular organelles are replicated.

Subcellular localized translation has implications for signal transduction and proteome organization (6). Using RIBOmap results, we identified five colocalized translation modules (LTMs) with highly correlated subcellular spatial organizations and distinct functional enrichment (LTMs 1 to 5, Fig. 3, E to G, fig. S8, A to C, and table S4). Among them, LTMs 2, 4, and 5 are enriched with a membrane protein and secretion pathway (Fig. 3, F and G and fig. S8 A and B) and physically colocalize with ER staining (Fig. 3, H and I and fig. S8, D to F), suggesting they are ER-translated genes. Indeed, 62.2 and 94.6% of LTM4 genes overlapped with the proximity ribosome profiling dataset (8) and APEX-RIP dataset (33), respectively (Fig. 3J, fig. S8D, and table S4), suggesting the accuracy of RIBOmap for subcellular spatial analysis. By contrast, LTMs 1 and 3 are enriched for genes encoding large protein complexes of the mitotic spindle and translation machinery, respectively (Fig. 3, F and G, and fig. S8, A and C). This observation suggests that the subunits of large protein complexes may be synthesized in spatial proximity for efficient assembly. Finally, we explored whether coregulated genes tend to be colocalized for translation and found that the five RTMs also show strong subcellular colocalization (fig. S8G). These results may imply that functionally related gene groups can be coregulated through subcellular colocalized translation, possibly through shared regulatory RNA elements or protein-protein interactions between nascent proteins.

RIBOmap in intact mouse brain tissue

We then applied RIBOmap to mouse brain slices to reveal single-cell translatome profiles in intact tissues. We mapped a targeted gene list of 5413 genes (table S5) curated from previously published single-cell sequencing studies of mouse cell atlas (34–40). We imaged two biological replicates of the mouse hemisphere coronal sections (60,481 cells for Rep1 and 58,692 cells for Rep2) containing multiple brain regions (Fig. 4, A and B). We benchmarked tissue RIBOmap data by comparing the spatial translation pattern of well-known cell-type marker genes and neurotransmitter genes with corresponding in situ hybridization (ISH) images from the Allen Brain database (41) (Fig. 4C and fig. S9), where the comparison showed consistent spatial patterns. The data quality of RIBOmap is comparable to previous STARmap PLUS data (42) (fig. S10A). Notably, we observed a high correlation (Pearson r = 0.95) between the two RIBOmap replicates (fig. S10B), validating the reproducibility of RIBOmap.

Encouraged by the benchmarking results, we then adopted a hierarchical clustering strategy (20, 36) to identify cell types (fig. S10, C to E), resulting in 11 major cell types and 38 subtypes (Fig. 4, D and E, fig. S10, C to E, and fig. S11). The cross-reference analyses revealed good correspondence between RIBOmap cell clusters and published regional scRNA-seq results of cell types (35) (fig. S10D). Based on these cell typing results, we generated spatial cell maps from the imaged hemibrain region (Fig. 4, F and G and fig. S12). This spatial cell-type map is consistent with previous reports (36), demonstrating the potential of RIBOmap single-cell translatomics to comprehensively identify diverse brain cell types and regions.

Cell type–specific and brain region–specific translational regulation in mouse brain

To compare the RIBOmap results with spatial transcriptomics, we conducted pairwise STARmap on a brain slice adjacent to RIBOmap sample followed by data integration for joint analyses (Fig. 5A). We observed consistent cell typing results between the two methods with respect to gene expression, cell-type composition, and spatial distribution of cell types (Fig. 5A and fig. S13). We also performed immunostaining of NeuN and GFAP proteins during RIBOmap and STARmap sample preparation and observed clear enrichment of NeuN protein signal in neurons and GFAP protein signal in astrocytes, respectively, in both samples (fig. S14), further supporting the accuracy of the cell typing results.

Leveraging the single-cell and spatial resolution of paired RIBOmap and STARmap data, we analyzed the heterogeneity of translational regulation across cell types and brain regions. We determined the reads correlation of 5413 genes between the two datasets for individual cell types or tissue regions (Fig. 5, B and C). Lower correlation scores indicate a greater degree of translational regulation, resulting in disparities between the translatome and the transcriptome. We found that non-neuronal cell types, especially oligodendrocytes, have lower correlation scores between translatome and transcriptome than neuronal cell types (Fig. 5B). Correspondingly, fiber tracts, the brain region where oligodendrocytes are enriched, had the lowest correlation (Pearson r = 0.58) between the translatome and transcriptome among the eight brain regions (Fig. 5C). This suggests a previously undercharacterized level of translational regulation in glial cells, particularly in oligodendrocytes.

To further pinpoint translationally regulated genes across various cell types, we performed gene clustering using major cell-type–resolved RIBOmap and STARmap profiles and identified 15 gene modules with distinct gene functions and expression patterns (fig. S15, A and B, and table S6). Given that the oligodendrocyte lineage showed the lowest correlation between translatome and transcriptome among major cell types (Fig. 5C), we focused on gene modules that were highly expressed in oligodendrocyte lineage cells (240 genes; fig. S15A and table S6). Pseudotime trajectory analysis revealed the differentiation path of the oligodendrocyte lineage in our datasets (40), from oligodendrocyte precursor cells (OPC) to oligodendrocyte subtype 1 (OLG1), then to oligodendrocyte subtype 2 (OLG2) (Fig. 5, D and E). This is supported by the increased expression gradients of genes involved in myelination (Plp1 and Mbp) along oligodendrocyte maturation (Fig. 5F) (43), and by the spatial enrichment of the more mature OLG2 in fiber tracts for myelination-based axon ensheathment (Fig. 5, G and H). To correlate translational regulation with oligodendrocyte subtypes and maturation, we further clustered oligodendrocyte-related gene modules using their subtype-resolved translatome and transcriptome profiles in the oligodendrocyte lineage (Fig. 5I, and table S6). We detected three gene modules: Module1 and Module 2 have consistent patterns between RIBOmap and STARmap with the highest signals in OPC and OLG2 for most genes, respectively, whereas Module 3 has the highest RIBOmap signal in OLG2 but the highest STARmap signal in OLG1 for most genes (Fig. 5I). This observation suggested strong translational regulation for genes in Module 3. We further calculated the relative translation efficiency (RTE) for each gene using their ratio of RIBOmap and STARmap signals and found that genes in Module 3 had higher RTE in OLG2 than OLG1 and OPC (Fig. 5I). GO analysis revealed that Module 3 genes are involved in myelin sheath (i.e., Cldn11 and Tspan2) (Fig. 5, J and K). Spatial analysis of Module 3 RTE patterns and representative genes across brain regions also showed the highest RTE value in the fiber tracts, where OLG2 is enriched (Fig. 5, L to O, and fig. S16A). We experimentally verified one example gene Cldn11 and found that Claudin 11 (the protein product of Cldn11) immunostaining correlated better with Cldn11 RIBOmap signal than STARmap signal, validating the accuracy of the spatial translatomic pattern we observed (fig. S16, B to E). Overall, through the integrative analysis of RIBOmap and STARmap results in the mouse brain, we identified gene modules that are translationally regulated during the maturation of oligodendrocyte lineage cells across different brain regions. The enhanced translation efficiency of Module 3 genes in OLG2 suggests substantial translation remodeling. This may functionally serve to augment the protein production for axon ensheathment and may be mechanistically relevant to the recent report on altered tRNA modification during oligodendrocyte maturation (44).

To systematically uncover genes with differential spatial patterns between translatome and transcriptome across brain regions, we performed gene clustering using region-resolved RIBOmap and STARmap profiles and identified gene modules with distinct spatial patterns between RIBOmap and STARmap across different brain regions (e.g., Module 7, fig. S17, A to C, and table S6). For example, G protein subunit gamma 2 (Gng2) in Module 7 has a low RIBOmap signal in the thalamus, which is consistent with the protein signal (extracted from the HPA database) and less correlated with the STARmap signal (fig. S17, D to I). By contrast, the RNA ISH signals of Gng2 from the Allen Brain database (41) correlated better with the STARmap signal than RIBOmap (fig. S17, J to M). These results were consistent with previous studies in which the translatome is better correlated with the proteome than the transcriptome (7, 11, 12).

Subcellular localized mRNA translation in mouse brain

In addition to brain-wide anatomical analysis, we investigated potential translational control at the subcellular level. In the brain tissue, subcellular localized translation in the cell body (soma) and periphery branches (processes) serves as a critical mechanism to assemble and adjust the network of neuronal and glial cells (45–50) in response to physiological signals during neurodevelopment and memory. To dissect localized translation in the somata and processes of neuronal and glial cells, we divided the RIBOmap reads into somata reads (i.e., inside the cell body area identified by ClusterMap) (51) and processes reads (i.e., the rest of the reads) (Fig. 6, A to C). Next, we defined the top 10% of genes with the highest and lowest processes-to-somata ratios as processes- and somata-enriched translation genes, respectively (Fig. 6D and table S7). GO analysis showed that processes-enriched translation genes are associated with translation machinery and postsynaptic density (Fig. 6E and fig. S18A). By contrast, somata-enriched translation genes are associated with the plasma membrane and extracellular matrix (Fig. 6F and fig. S18B), corresponding to localized translation at the ER in the somata. Notably, we observed abundant processes-enriched translation signals in hippocampal neuropils (Fig. 6G and fig. S18C). By contrast, the somata-enriched translation genes showed sparse RIBOmap signals in hippocampal neuropils (Fig. 6H and fig. S18D), which may be associated with ER translation. Beyond neuronal genes, we also observed localized translation in glial cells (Fig. 6, I and J, and fig. S18, E and F). Overall, we demonstrated that RIBOmap can be used to study subcellular-localized translation in processes of both neuronal and glial cells of the mouse brain tissue.

In summary, RIBOmap is a spatial translatomics method with single-cell and molecular resolution, which can be used to study mRNA regulation and protein synthesis in intact cellular and tissue networks. In contrast to existing approaches, RIBOmap bypasses complicated polysome isolation steps and genetic manipulation, thus holding promise for spatial and single-cell resolved studies in post hoc human tissue and disease samples. The detection efficiency and spatial resolution of RIBOmap can be further increased by combining with super-resolution imaging, expansion microscopy, and sparse labeling (52) to further illustrate translational events in fine subcellular structures (e.g., dissecting dendritic and axonic mRNA translation in neurons). It is noteworthy that the accuracy of RTE analysis depends on the precise joint cell typing of RIBOmap and STARmap data. Therefore, further computational development to improve the cross-modality integration of various spatial omics modalities will significantly enhance the analysis and interpretation of RIBOmap. Whereas RIBOmap in this manuscript focuses on mapping RNA translation, such proximity-based tri-probe design can be readily adapted to study RNA-RNA interactions, RNA-protein interactions, and RNA modifications. In the future, we envision that RIBOmap can be combined with other imaging-based measurements to enable spatial multiomics mapping of epigenome, transcriptome, and translatome in the same samples for an integrative understanding of biological systems.

Supplementary Material

Data and materials availability:

The RIBOmap preprocessed dataset and gene expression dataset are available on Zenodo (53, 54) and Single Cell Portal (SCP) (https://singlecell.broadinstitute.org/single_cell/study/SCP1835). All code and analysis are available on Zenodo (55).