Automated in situ microfluidic Random-seq for robust single-nucleus and spatial total RNA profiling of diverse FFPE specimens Authors: Haide Chen, Yu-Sheng Chen, Shunji Zhang, Jiaye Chen, Hao Yang, Tongqiang Fan, Jianfei Zheng, Yuan Liao, Qian Liang, Yongcheng Wang Published: 2026 ABSTRACT AbstractFormalin-fixed paraffin-embedded (FFPE) tissues are indispensable for clinical- pathological assessment, yet widespread nucleic acid degradation severely hinders high-throughput molecular profiling. Current RNA sequencing and in situ hybridization approaches for FFPE materials suffer from inconsistent performance across tissue types and preservation conditions, restricting standardized, automated transcriptome analysis. Here, we show an in situ microfluidic Random-seq (imRandom-seq) platform that enables unified bulk, single-nucleus, and spatial total RNA profiling of FFPE specimens. Using specially designed random primers for efficient transcript capture and total transcriptome analysis, single-nucleus imRandom-seq outperforms traditional snRNA-seq and probe-based 10X Flex, with enhanced gene detection, reduced nuclear loss, and biologically reasonable cell-type annotation. Validated in difficult samples with high enzymatic activity and fragmented RNA, this in situ microfluidics-driven workflow features low manual operation, broad tissue compatibility, and robust data quality, providing a reliable and scalable tool for FFPE transcriptomic research.Subject terms: RNA sequencing, Reverse transcription polymerase chain reaction, Microfluidics, Transcriptomics Introduction Formalin-fixed paraffin-embedded (FFPE) tissue blocks, the gold standard for human tissue preservation in clinical diagnostics, stand as a “treasure trove” of clinical resources1. Yet despite their immense potential for clinical investigation, FFPE samples have long posed formidable challenges for molecular analysis. The processes of formalin fixation and paraffin embedding induce chemical modifications, fragmentation, and cross-linking of nucleic acids—with RNA being particularly vulnerable to degradation—significantly impeding comprehensive capture of the gene expression landscapes within these specimens. The advent and advancement of RNA-seq and in situ hybridization (ISH) technologies have emerged as pivotal solutions for profiling RNA in FFPE samples. More notably, a suite of single-nucleus RNA sequencing (snRNA-seq), spatial RNA sequencing (spRNA-seq), and FISH (Fluorescence In Situ Hybridization) techniques has been developed specifically for FFPE specimens2–10. These innovations enable researchers to delve deeper into the molecular intricacies of tissue samples that were previously recalcitrant to detailed analysis, thereby ushering in an era of in-depth exploration for FFPE-based research. RNA-seq (including bulk RNA-seq, snRNA-seq, and spRNA-seq) provides a whole-transcriptome perspective for FFPE samples. Due to the high degree of transcriptome degradation in FFPE samples, RNA-seq and FISH primarily use target probes to capture fragmented RNA. Limited by known genes, the use of target probes was applied only for a few species, like human and mouse. It can only perform quantitative analysis of transcripts, with no ability to conduct qualitative analysis (such as splicing, mutations, etc.). Currently, RNA capture based on random primers is employed in FFPE RNA detection3,8. Random primers bind to RNA molecules irrespective of the poly(A) tail, thereby enabling the capture of non-mRNA transcripts and degraded RNA lacking this tail, while also facilitating comprehensive total transcriptome analysis. Additionally, random primers can hybridize with multiple different sites on a transcript, facilitating the capture and analysis of the full length of the transcript. Despite these advancements, the cumbersome, poorly automated procedures hinder the widespread application4–6,11. The numerous manual steps increase the risk of variability and inconsistency in the results, which is detrimental to the accuracy of the data obtained. The extraction process of RNA or nuclei can introduce selectivity for RNA length and nucleus size, thereby introducing biases that may lead to a distorted understanding of the transcript profile within a sample and potentially obscure crucial biological insights. This unpredictability poses a major obstacle to the clinical application of these techniques. Meanwhile, the sample variability caused by the tissue type, tissue size (puncture, surgical resection), preparation condition (fixation, dehydration, embedding), and storage (time, temperature) of FFPE samples has also increased the difficulty of automated processing. To overcome these challenges, we developed an in situ microfluidic Random-seq (imRandom-seq) with a device that enables the automatic detection of bulk, single-nucleus, and spatial total RNA from FFPE samples. We utilize microfluidic chips to precisely manipulate and analyze FFPE samples, especially micro-volume samples—a capability beyond the reach of previous manual techniques. We design special random primers for imRandom-seq to efficiently capture transcripts, which enables total transcriptome analysis. When benchmarked against conventional snRNA-seq and probe-based 10X Flex, our method achieves markedly higher gene detection sensitivity, reduced nucleus loss, and biologically more reasonable cell type composition. We also apply imRandom-seq across a diverse range of tissues, including digestive system samples with high levels of digestive enzymes and samples with low RNA quality, to fully demonstrate the wide applicability and efficiency of the platform. Furthermore, we show that spatial imRandom-seq applies widely across various tissues, including the brain, oral cavity, and lung. In summary, we develop imRandom-seq as a refined and automated approach to process FFPE samples, leveraging the precision and efficiency of microfluidics technology. This protocol streamlines the process, minimizes manual intervention, and significantly enhances sample applicability and data quality. Results Automated FFPE sample processing with microfluidics To enhance the analysis quality, throughput, and streamline the operations, we developed an in situ microfluidic Random-seq (imRandom-seq) that could automatically execute treatments of FFPE samples for RNA-seq (Fig. 1a, b). The automated FFPE RNA-seq processing follows a sequential and efficient in situ process3,8. Once the sample is loaded, the pump priming system sequentially introduces dewaxing and rehydration reagents. Subsequently, the pipetting system takes over and adds reagents for permeabilization, single-stranded DNA blocking, in situ reverse transcription, dA tailing, washing, and cDNA labeling. This in situ approach ensures that each reaction step occurs precisely, enhancing the accuracy and efficiency of the overall FFPE RNA-seq processing. Two adaptive modules were designed to adapt to the two forms of FFPE samples (unmounted and mounted). Overview of in situ microfluidic Random-seq (imRandom-seq). a Automated sample processing of bulk, single nucleus, and spatial transcriptomics imRandom-seq. b The photograph depicting the automated device. c Real-time temperature monitoring of reagents in the PDMS chip during thermal cycling. Source Data are provided as a Source Data file. d Design of the pump priming system and the waste collection system. e The module of imRandom-seq for unmounted FFPE samples. f The module of imRandom-seq for mounted FFPE samples. See also Supplementary Figs. 1–3. This platform encompasses a straightforward man-machine interface for ease of use, a chip module equipped with a temperature controller and a chip cover, and a programmable pipetting system with a pump priming system (Fig. 1c, d; Supplementary Movies 1, 2). It also includes refrigerated reagent reservoirs, a mobile device for pipetting and pump priming, and an efficient waste collection system. The enzyme-related reagents are stored on a refrigeration module to preserve their activity. During the reactions, the chip covers hermetically seal the sample slot to prevent evaporation, ensuring that the reactions proceed under controlled conditions. Waste management is streamlined through a waste tube that channels all waste into a designated waste bottle, maintaining a clean and organized workspace. The unmounted module is suitable for both bulk RNA-seq and single-nucleus RNA-seq applications. The microfluidic design significantly simplifies the workflow and aligns with modern high-throughput sequencing demands. The microfluidic chip is a pivotal component of this platform, meticulously designed for high-throughput automated sample processing (Fig. 1e and Supplementary Fig. 1a). Configured with eight channels, the chip enables simultaneous processing of up to eight samples. Each channel includes a designated sample slot for loading FFPE scrolls and a waste slot that connects to a waste tube for efficient disposal. The chip module provides a stable environment during chip incubation, safeguarding the samples from external disturbances and maintaining optimal temperatures. This stability is crucial for maintaining the precision and reliability of the results. The chip module achieves a slight oscillation with a frequency of 720 revolutions per minute and an amplitude of 3 mm, which enhances the mixing of the samples and reagents. For snRNA-seq, the tissue scrolls are carefully transferred from the sample slot to a centrifuge tube for nuclei isolation using an optimized lysis solution (Supplementary Fig. 1b). The resulting single-nucleus suspension is then prepared for droplet barcoding and subsequent sequencing. For FFPE samples that were mounted on slides, we also designed an adaptive module for operation, which mainly meets the detection needs of spatial transcriptomics (Fig. 1f and Supplementary Fig. 2). imRandom-seq can simultaneously accommodate two standard tissue sections. Tissue sections are loaded into the modules, and the target areas of the samples are aligned with the reaction frame. Reagent addition and removal are executed via the programmable pipetting system. After the RNA capturing, an encoded spatial chip is used for spatial labeling of RNA8. The device can also be adapted to other spatial omics workflows through programmatic editing. To further assess the imRandom-seq platform, we first performed bulk sequencing for a stomach tumor FFPE sample using different channels and observed a high level of reproducibility (R = 0.96) (Supplementary Fig. 3a). We then compared the performance of manual and instrument-based workflows using the same human kidney tumor FFPE sample. The cDNA qPCR CT value analysis showed no significant difference between the two workflows (Supplementary Fig. 3b). Consistently, gene expression levels derived from the manual and instrument-based workflows were highly correlated (R = 0.96) (Supplementary Fig. 3c), further supporting the robustness of the platform. Moreover, using a human breast cancer FFPE sample, we performed a head-to-head comparison between bulk imRandom-seq and standard RNA-seq, which requires a substantially larger amount of extracted RNA for library construction rather than in situ reaction. Both methods detected comparable gene categories and exhibited similar gene body coverage without evident 5′ or 3′ bias (Supplementary Fig. 3d,e). Gene expression profiles generated by the two approaches showed a moderate to high correlation (R = 0.74) (Supplementary Fig. 3f). Compared to previous comparisons, the slightly lower correlation may be attributable to differences in library preparation strategies, as imRandom-seq generates strand-specific libraries, whereas standard RNA-seq we applied to FFPE samples is non-stranded, which becomes even more pronounced in the context of FFPE tissues. Thus, frozen mouse lung and kidney tissues were processed by standard RNA-seq and compared with the corresponding FFPE samples processed by imRandom-seq (Supplementary Fig. 3g). Both lung and kidney tissues showed increased correlation (R = 0.82 for lung and R = 0.83 for kidney), indicating improved consistency and the reliability of imRandom-seq in capturing biologically relevant signals from FFPE tissues. Performances of bulk and single-nucleus transcriptome profiles on various FFPE tissues using imRandom-seq To assess the versatility of imRandom-seq, we first applied bulk imRandom-seq for clinical FFPE samples from three distinct tissues, including kidney, uterus, and intestine. In addition to detecting more than 26,000 genes with 14,000 protein-coding genes across these tissues, imRandom-seq excelled at validating a wide variety of non-coding RNAs, such as lncRNA, snRNA, snoRNA, and more (Fig. 2a), suggesting that the whole transcriptome profile of FFPE samples could be well identified by imRandom-seq. Comparison of imRandom-seq with snRNA-seq and 10X Flex platforms. a Barplot displaying the number of different gene biotypes detected in FFPE kidney, uterus, and intestine tissues using imRandom-seq. b Saturation analysis of snRNA-seq (red) and imRandom-seq (green). UMAP plots of the integrated datasets of lung tumor tissues from snRNA-seq and imRandom-seq (c), and respective dataset (d; left, snRNA-seq; right, imRandom-seq). Colored by identified cell types. UMAP plots of the integrated datasets of mouse brain tissues from snRNA-seq and imRandom-seq (e), and respective dataset (f; left, snRNA-seq; right, imRandom-seq). Colored by identified cell types. g Barplot displaying the proportion of various cell types annotated in the mouse brain tissues from snRNA-seq (upper) and imRandom-seq (lower). h UMAP plot of the integrated datasets of lung tumor tissues from imRandom-seq and 10X Flex. Colored by identified cell types. i Kaplan-Meier analysis showing that a high expression of C2CD4D-AS1 contributes to poor OS (overall survival) for LUAD patients (TCGA LUAD cohort; n = 504, low: 321, high: 183). P value is calculated by the two-sided log-rank test comparing survival distributions between the corresponding two groups. j Expression levels of non-coding oncogene identified by TCGA data among various cells in lung tumor tissues from imRandom-seq. Color key from gray to blue indicates relative expression levels from low to high. See also Supplementary Figs. 4–8. Source Data are provided as a Source Data file. Since bulk imRandom-seq exhibited no bias in transcriptomic profile, we then considered applying it to single-cell RNA-seq to generate a single-cell resolution transcriptional atlas for FFPE samples, providing a powerful technique for exploring clinical research using FFPE samples. However, the low quality and fragmented nature of RNA from FFPE samples might lead to substantial signal loss, resulting in inaccurate transcriptomic features, particularly in the highly sparse single-cell RNA-seq data. Here, to achieve high-fidelity signals from individual nuclei, imRandom-seq performs nuclear RNA labeling before nuclei isolation, which is beneficial for capturing low-abundance cellular RNAs in traditional single-nucleus RNA-seq. Thus, we prepared one FFPE lung tissue sample with relatively low RNA quality (DV200 less than 30%, under the recommended value for the 10X Flex platform; Supplementary Data 1) and then performed single-nucleus imRandom-seq. Compared to traditional snRNA-seq methods, which made it difficult to obtain bioinformation from this sample, data from imRandom-seq were able to identify various cell types (Supplementary Fig. 4a, b), including alveolar type 1 (AT1), alveolar type 2 (AT2), immune cells, and so on, even at low sequencing depth. Needle biopsy samples are another common type of clinical FFPE sample that holds valuable information. However, given their small size, the high RNA loss rate caused by multiple centrifugation steps poses a significant challenge for traditional snRNA-seq methods. A significant advancement in the imRandom-seq protocol is the in situ reaction within the tissue sections, an advancement that is especially beneficial for samples of fine-needle biopsies. Here, using imRandom-seq, we were able to successfully identify various cell types in the needle biopsy of lung tissue (Supplementary Fig. 4c, d), suggesting the broad application prospects of imRandom-seq. To compare single-nucleus imRandom-seq with traditional snRNA-seq methods on FFPE samples in parallel, we applied both methods to the same human lung cancer tissue with a similar sequencing depth, and analyzed the single-cell transcriptome landscapes using the same pipelines. As anticipated, imRandom-seq identified a greater number of UMIs and genes within over 80% mapping ratio (Fig. 2b and Supplementary Fig. 5a, b), and exhibited higher cellular gene diversity compared to snRNA-seq for the same depth per cell, with the relatively low proportion of rRNAs (Supplementary Fig. 5c). Then, we performed cell annotations for the merged data from snRNA-seq and imRandom-seq within batch correction using Harmony12. The main cell types in lung tissues, including epithelial cells, immune cells, and stromal cells, were well identified by marker genes, such as KRT5 (basal cells), JCHAIN (plasma cells), CD247 (T cells), COL1A1 (fibroblasts), and FLT1 (endothelial cells)(Fig. 2c and Supplementary Fig. 5d, e). Moreover, data from both methods displayed similar capabilities in identifying different cells and showed high consistency in the relative proportions (Fig. 2d and Supplementary Fig. 5f). Since the centrifugation wash by snRNA-seq might result in a bias towards cells within larger sizes, such as neuron and plasma cells13, the decreased proportion of plasma cells suggested that data derived from imRandom-seq might provide an unbiased way to capturing cells within different sizes. We also compared gene expression profiles between snRNA-seq and imRandom-seq at both pseudobulk and individual cell type levels, revealing high consistency in both comparisons (Supplementary Fig. 5g, h). To further assess the bias in cell sizes between traditional snRNA-seq and single-nucleus imRandom-seq, we applied both methods to mouse brain tissues. Despite both methods detecting identical cell types with highly consistent gene expression at pseudobulk and individual cell type levels (Fig. 2e and Supplementary Fig. 6), snRNA-seq exhibited a significant bias towards neuronal cells (over 90%) and resulted in the loss of a remarkable number of non-neuronal cells, including endothelial cells, smooth muscle cells, and glial cells (Fig. 2f, g). In contrast, data by imRandom-seq, which employed controlled microchannels, was effective in preventing the washing away or damage of nuclei, thus ensuring a more authentic representation of cell types (Fig. 2f, g). Owing to the use of random primers, both snRNA-seq and imRandom-seq displayed unbiased distributions along genebody (Supplementary Fig. 7a). Therefore, we further examined their capability to capture alternative splicing events (Supplementary Fig. 7b). As expected, both methods exhibited a similar capacity to detect various alternative splicing events (Supplementary Fig. 7c), including skipped exon (SE), retained intron (RI), mutually exclusive exon (MXE), alternative 5′ splice site (A5SS), and alternative 3′ splice site (A3SS). Most of these events were supported by 10–1000 UMIs (Supplementary Fig. 7d, e). Collectively, compared to the traditional method, imRandom-seq worked as a powerful tool in unbiasedly presenting the cellular landscape with higher gene diversity, which enhanced both sample applicability and the quality of the data obtained. imRandom-seq expands oncogenic non-coding marker genes in distinct cells Next, we compared the single-nucleus imRandom-seq with the 10X Genomics Flex platform, which utilized a series of probes for theoretically capturing specific transcripts from FFPE samples11. The public 10X Flex data of a lung cancer sample was downloaded from the 10X Genomics official website. As expected, 10X Flex could only detect those genes with designed probes, which were remarkably fewer than the genes detected by imRandom-seq using random primers, especially for non-polyadenylated (non-polyA) genes (Supplementary Fig. 8a, b). To further assess whether the high diversity of genes leads to systematic bias in cell annotation, we merged lung cancer data from imRandom-seq and 10X Flex, and then performed debatch correction. Both imRandom-seq and 10X Flex demonstrated similar abilities in detecting various cell types, including epithelial cells, fibroblasts, macrophages, T cells, and B cells (Fig. 2h), and the expressions of corresponding marker genes validated by both methods were consistently similar (Supplementary Fig. 8c), suggesting similar capability in annotating most cell types for both methods on FFPE samples. We also observed that the 10X Flex platform identified more plasma cells but fewer endothelial cells compared to imRandom-seq (Supplementary Fig. 8d), which might be due to different FFPE sample characteristics or technical bias. Moreover, lncRNAs played important regulatory roles in tumorigenesis and tumor microenvironment14, but most comprehensive single-cell transcriptional landscapes are limited by polyA- or probe-dependent libraries. Here, we demonstrated the advantages of imRandom-seq in validating lncRNAs across different cell types (Supplementary Fig. 8e–g), which could benefit the biological and pathological understanding of lncRNAs’ roles in tumorigenesis. Additionally, we downloaded LUAD (Lung adenocarcinoma) transcriptomic profiles by bulk RNA-seq from TCGA, and detected about 394 genes that were up-regulated in tumors compared to normal samples (Supplementary Fig. 8h). Among these up-regulated genes, we found that C2CD4D-AS1 seemed to be an oncogene whose high expression level resulted in shorter survival time (Fig. 2i). Intriguingly, previous study reported that C2CD4D-AS1 contributed to the malignant phenotype of lung adenocarcinoma cell lines15, and the comprehensive single-cell transcriptional landscapes by imRandom-seq provided another potential mechanism for C2CD4D-AS1 with its high expression in T cells (Fig. 2j). In summary, compared to the 10X Flex platform, imRandom-seq demonstrated comparable ability in unbiasedly capturing different cell types. Owing to the use of random primers instead of designed probes, imRandom-seq provided advantages in detecting lncRNAs and showed high versatility across a large number of species, while achieving a remarkable 8-fold reduction in the cost of library construction per sample (Supplementary Data 2). imRandom-seq reveals tissue-specific protein-coding and non-coding markers for epithelial cells Considering that RNA in FFPE samples is frequently compromised in quality and prone to fragmentation owing to the preparation and preservation processes16, the previous traditional snRNA-seq method might lead to relatively more RNA degradation due to nuclei isolation before RNA labeling. This was particularly in gastrointestinal tissues such as the esophagus, stomach, and intestine. To enhance quality of the cDNA library, imRandom-seq was optimized to perform RNA labeling and reverse transcription before nuclei isolation, and further refined by minimizing the number of centrifugation steps, given that the centrifugal washing of nuclei might facilitate RNA and cDNA loss (Supplementary Fig. 1). We then applied imRandom-seq to gastrointestinal tissues, including the esophagus, stomach, and intestine, as well as non-gastrointestinal tissues, including the embryo, skin, and uterus. The initial annotation of cell types from integrated data were well detected, with clear identification of epithelial cells, stromal cells, endothelial cells, and immune cells (Fig. 3a and Supplementary Fig. 9a). The classical markers displayed a high degree of specificity for their respective cell types, for instance, IGHA2 for plasma cells, COL1A1 for fibroblasts, MYH11 for smooth muscle cells, VWF for endothelial cells (Fig. 3b and Supplementary Fig. 9b). Furthermore, despite the variability in cell type ratios across the six tissues examined, imRandom-seq reliably identified all cell types in both gastrointestinal and non-gastrointestinal tissues (Fig. 3c and Supplementary Fig. 9c), demonstrating its versatility and broad applicability. Cell heterogeneity revealed in FFPE samples by imRandom-seq. a UMAP plots of the integrated datasets of gastrointestinal tissues, including the esophagus, stomach, and intestine, as well as non-gastrointestinal tissues, including the embryo, skin, and uterus. Colored by identified cell types. b Violin plot displaying expression levels of representative markers for 10 identified cell types in multiple FFPE tissues. c UMAP plots of the integrated datasets of gastrointestinal tissues, including the esophagus, stomach, and intestine, as well as non-gastrointestinal tissues, including the embryo, skin, and uterus. Colored by different FFPE tissues. d Violin plot displaying expression levels of protein-coding (left) and non-coding markers (right) for epithelial cells among 6 tissues. e Pie charts displaying the proportions for gene biotypes of tissue-specific markers related to epithelial cells (upper) and cell-specific markers (lower). f Co-stained PVT1 RNA smFISH (red) and EPCAM protein IF (green) in FFPE intestine (upper) and uterus tissues (lower), experiment performed once (n = 1). DAPI stain (blue) labels the cell nuclei. Scale bar, 10 µm. See also Supplementary Figs. 9 and 10. Source Data are provided as a Source Data file. Next, we investigated the cellular subtypes presenting within each tissue, and found that certain tissue-specific cell types, particularly those of epithelial origin, were clearly distinguished (Supplementary Fig. 9d). Given the diverse subtypes of epithelial cells across various tissues, we attempted to find out their own epithelial markers for multiple tissues (Fig. 3d and Supplementary Fig. 10a, b). Interestingly, lncRNAs showed a heightened sensitivity in tracing the tissue of origin for epithelial cells compared to their ability to differentiate between cell types (Fig. 3e and Supplementary Fig. 10b). This could be attributed to the fact that the differences among various cell types are largely driven by a spectrum of functional proteins, whereas the distinctions within the same cell types across different tissues are likely due to variations in regulatory mechanisms. Furthermore, we conducted FISH to ascertain the expression levels of the gastrointestinal lncRNA marker PVT1 in both the intestine and uterus. Our findings revealed a strong alignment between these phenotypes and the broader transcriptomic profiles obtained through imRandom-seq (Fig. 3f), indicating that lncRNAs might play a significant role in tracking the spread of malignant cells. Identification of malignant cells and tumor-specific non-coding markers Given the extensive inventory and long-term stability, FFPE samples provide significant advantages in tumor research, particularly for retrospective studies. We then applied imRandom-seq to various FFPE tumor samples, including esophageal cancer (EC), kidney cancer (KC), primary lung cancer (LC_primary), metastatic lung cancer (LC_metastasis), ovarian cancer (OC), lymphoma (LY), and melanoma (MA), to demonstrate its wide applicability. Within debatch processing, cells from different tumor samples were well-mixed (Supplementary Fig. 11a), allowing us to clearly distinguish cell types based on their markers (Fig. 4a–c and Supplementary Fig. 11b, c). Additionally, subtypes of cells were identified for each tumor sample, which exhibited the characteristics of their respective tissues (Supplementary Fig. 11d; esophageal cancer shown in Supplementary Fig. 9d as Esophagus). Intriguingly, we observed a few epithelial-labeled cells in close proximity to B cells, predominantly from lymphoma samples (Fig. 4a and Supplementary Fig. 11a), and these cells were further characterized as epithelial-like B cells in lymphoma samples (Supplementary Fig. 11d). Given that the disordered B cells can contribute to lymphoma17, these epithelial-like B cells might represent the malignant cells in lymphoma. imRandom-seq validated oncogeneic lncRNAs in clinical FFPE tumor samples. a UMAP plots of the integrated datasets of FFPE tumor tissues, including esophageal cancer (EC), kidney cancer (KC), primary lung cancer (LC_primary), metastatic lung cancer (LC_metastasis), ovarian cancer (OC), lymphoma (LY), and melanoma (MA). Colored by identified cell types. b Violin plot displaying expression levels of representative markers for 8 identified cell types in multiple FFPE tumor tissues. c Barplot displaying the proportion of various cell types annotated in multiple FFPE tumor tissues. d Inferred CNV profiles of epithelial cells and immune cells. Chromosomal amplification (red) and deletion (blue) are inferred in each chromosomal position (columns) across the single cells (rows). Top, immune cells. Bottom, malignant epithelial cells among 6 FFPE tumor tissues. e Bubble plot displaying expression levels of up-regulated genes in malignant and non-malignant epithelial cells among 6 different FFPE tumor tissues, respectively. For each tumor tissue, the specific oncogenes are labeled with different colors, and the common oncogenes with black. f Kaplan-Meier analysis showing the high expression of epithelial-related oncogenes identified in FFPE tumor tissues contributes to poor OS (overall survival) for corresponding patients in TCGA. EC: n = 173 (low: 82; high: 91); KC: n = 529 (low: 401; high: 128); LC_primary: n = 504 (low: 398; high: 106); OC: n = 399 (low: 220; high: 179); MA: n = 103 (low: 48; high: 55). P values are calculated by the two-sided log-rank test comparing survival distributions between the corresponding two groups. See also Supplementary Figs. 11–13. Source Data are provided as a Source Data file. To further identify malignant cells in other tumor tissues, we conducted CNV analysis on epithelial cells from each tumor sample. Through hierarchical clustering based on CNV profiles across the genome, epithelial cells were effectively categorized into malignant and non-malignant groups for each sample, and the malignant cells presented distinct, patient-specific clusters characterized by relatively high CNV scores (Fig. 4d and Supplementary Fig. 12a), consistent with mechanisms of tumor-related copy number variation. Additionally, we analyzed differentially expressed genes in each tumor tissue, observing both up- and down-regulated genes in malignant cells (Supplementary Fig. 12b). These findings could potentially serve as biomarkers or therapeutic targets for cancer treatment. In each tumor sample, up-regulated genes in malignant cells were enriched in various tumorigenesis pathways, such as cellular response to hypoxia in kidney cancer, DNA repair in primary lung cancer, the EGFR (epidermal growth factor receptor) signaling pathway in metastatic lung cancer, and intracellular signal transduction in melanoma, among others (Supplementary Fig. 12c). Specifically, we identified the lung-associated pathway of surfactant homeostasis in metastatic lung cancer tissues (Supplementary Fig. 12c), indicating that the distinguished malignant cells can reflect the characteristics of the original tissue. We then explored tissue-specific oncogenes and found well-studied protein-coding markers, such as YAP1 in esophageal cancer18, PLOD2 in kidney cancer19, RPS6KA2 in lung cancer20, PAX8 in ovarian cancer21, and DLC1 in melanoma22. We also found some non-coding markers, like EPIC1 in lung cancer23,24 (Fig. 4e). Interestingly, compared to tissue-specific oncogenes, the common oncogenes that were up-regulated in more than half of tissues tended to be non-coding genes rather than protein-coding genes, such as PVT125, and ZFPM2-AS126 (Fig. 4e). By integrating bulk RNA-seq data from TCGA database, we conducted survival analyses on the up-regulated lncRNAs identified by imRandom-seq. The significantly decreased overall survival times indicated their critical roles in tumorigenesis in the corresponding tissues (Fig. 4f). This highlights the potential of imRandom-seq to distinguish lncRNAs at single-cell resolution, which could promote identification of potential biomarkers or therapeutic targets for cancer treatment. Some lncRNAs act as competing endogenous RNAs (ceRNAs), which protect target mRNAs from degradation by sequestering common microRNAs27. Utilizing the lncACTdb28, we predicted lncRNA-microRNA-mRNA regulatory axes based on differential expressed genes in malignant cells for each tumor, and constructed the regulatory network for visualization (Supplementary Fig. 13). Intriguingly, the MALAT1-dependent ceRNA could rescue lots of critical oncogenes, such as EGFR, TGFA, and VEGFA, in various cancers, demonstrating its potential regulatory role during tumorigenesis as ceRNA. Performances in spatial transcriptome across tissues using imRandom-seq Combined with previous studies8, we also established spatial imRandom-seq for spatial transcriptomic analysis, providing high-quality data across various tissues. We initially tested its performance on spatial transcriptomics using FFPE mouse brain tissue, which has distinct region categories (Supplementary Fig. 14a). Despite the relatively low sequencing depth, each spot yielded over 5,000 UMIs and 1,200 genes, with no obvious spatial bias (Fig. 5a and Supplementary Fig. 14b). Moreover, given the random priming strategy, both protein-coding and non-coding genes could be well detected (Supplementary Fig. 14c), which was advantageous for studying lncRNA mechanisms. We then clustered the spatial spots based on their transcriptome profiles, identifying 12 distinct niches within the mouse brain tissue (Fig. 5b, c and Supplementary Fig. 14d). Additionally, we used CARD to deconvolute the different brain regions, leveraging a validated single-cell RNA-seq dataset29. Consistent with brain tissue staining, diverse brain regions were accurately identified by spatial transcriptomics (Fig. 5d), including the entorhinal cortex (ENT), inferior temporal cortex (IT), cerebral cortex (CTX), medial amygdaloid nucleus (MEA), hippocampus (CA2, CA3, DG), striatum (STR), and others. Spatial transcriptomics on various FFPE tissues are also capable of imRandom-seq. a Performance of imRandom-seq on FFPE mouse brain section. Top: spatial UMI count map. Bottom: spatial gene count map. b UMAP plots displaying unsupervised clustering of FFPE mouse brain section. c Spatial visualization for unsupervised clustering on FFPE mouse brain section at 50 µm resolution. d Cell-type deconvolution results showing various mouse brain regions in the imRandom-seq spatial transcriptomic profile. The reference mouse brain scRNA-seq dataset is retrieved from the Mouse Brain Atlas (MBA). Color scale indicates the min-max normalized proportions of a given cell type across all spots on the spatial map. e Performance of imRandom-seq on FFPE human oral squamous cell carcinoma (OSCC) sample. Top: spatial UMI count map. Bottom: spatial gene count map. f UMAP plots displaying unsupervised clustering of FFPE OSCC section. g Spatial visualization of unsupervised clustering on FFPE OSCC section at 15 µm resolution. h Spatial expression mapping of OSCC oncogenes MKI67, TP53, and SESN1. i Spatial feature score maps of genes in cell cycle, epithelial-mesenchymal transition (EMT), and quiescence pathways. j Violin plots showing the number of UMIs (left) and genes (right) detected in mouse normal lung tissue by imRandom-seq. The center line indicates the median, the box bounds represent the interquartile range (IQR), and the whiskers extend to the most extreme data points within 1.5× IQR. k Spatial visualization of unsupervised clustering on mouse lung tissue at 50 µm resolution. l UMAP plots displaying cell types from a paired imRandom-seq dataset. m Cell-type deconvolution results showing various mouse lung cell types in the imRandom-seq spatial transcriptomic profile. The paired single-nucleus dataset was used as a reference. Color scale indicates the min-max normalized proportions of a given cell type across all spots on the spatial map. See also Supplementary Figs. 14–16. Source Data are provided as a Source Data file. To further test the application in clinical FFPE tissues, we applied imRandom-seq on a human oral squamous cell carcinoma (OSCC) sample (Supplementary Fig. 15a). As sequencing depth increased, we validated nearly 10,000 UMIs and 5000 genes per spot, which also showed high sensitivity for non-coding RNAs (Fig. 5e and Supplementary Fig. 15b, c). The spots were then clustered into 14 niches (Fig. 5f, g and Supplementary Fig. 15d). Additionally, the expression patterns of marker genes across different cell types exhibited distinct distributions in the FFPE samples (Supplementary Fig. 15e), indicating a complex tumor microenvironment (TME). We further estimated the signatures of typical oncogenesis pathways, and the intratumoral heterogeneity was well identified by the spatial transcriptome profile (Fig. 5i). Notably, niche 11 was highly enriched in the cell cycle pathway, with co-expression of MKI67 and PTPRC (Fig. 5h, i and Supplementary Fig. 15e), suggesting high immune infiltration around niche 11. Combining paired single-cell RNA-seq with spatial transcriptomics enables precise annotation of cell types. Here, we performed both spatial and single-nucleus imRandom-seq on the same FFPE mouse normal lung sample (Supplementary Fig. 16a). As expected, each spot in the spatial transcriptomics generated over 4000 UMIs and 2000 genes (Fig. 5j and Supplementary Fig. 16b). Unlike FFPE mouse brain and human tumor samples, which exhibit distinct brain regions or tumor heterogeneity, normal lung tissues displayed uniform transcriptome profiles (Fig. 5k and Supplementary Fig. 16c). Concurrently, we conducted single-nucleus imRandom-seq on adjacent sections from the same FFPE lung sample, identifying a variety of cell types (Fig. 5l and Supplementary Fig. 16d, e). The spatial transcriptome profile by imRandom-seq was further deconvoluted using the single-nucleus transcriptome profile, confirming, as anticipated, that these cells were evenly distributed throughout the lung tissue (Fig. 5m). In summary, imRandom-seq has proven highly effective for constructing both single-nucleus and spatial transcriptomics libraries, enabling precise cell type annotation and detailed spatial profiling across various tissues. Its ability to generate high-quality data, even with low sequencing depth, highlights its robustness and versatility, making it a valuable tool for studying complex tissue architectures and disease mechanisms. Discussion We demonstrated in situ microfluidic Random-seq (imRandom-seq), a technology that integrates microfluidic automation with optimized random primer-based chemical reactions to enable automated, high-performance bulk, single-nucleus, and spatial RNA-seq for FFPE samples. By leveraging microfluidic platforms, imRandom-seq streamlines sample processing—from tissue permeabilization to cDNA synthesis—with minimal manual intervention, while its chemically optimized reagents (e.g., random primers and buffer systems) enhance compatibility with degraded RNA, a hallmark of FFPE tissues. This hardware-software integration not only ensures robust transcriptome capture across diverse sample types but also lays the foundation for its clinical applicability. FFPE RNA analysis is traditionally dominated by probe-based methods (e.g., 10X Flex), which excel in targeted gene panel analysis with strong species specificity. In contrast, random primers—a key chemical feature of imRandom-seq—offer distinct advantages: they bypass probe-dependent limitations, enabling comprehensive capture of diverse RNA species (mRNAs, lncRNAs, and other non-coding RNAs) across varied species and sample conditions4,8. This breadth is critical for exploring understudied non-coding transcripts, which are often overlooked by probe-based approaches but hold significant biological and clinical relevance. Complementing this chemical design, imRandom-seq’s microfluidic hardware enables in situ reactions within intact tissues or nuclei, reducing RNA loss and secondary degradation that frequently occur during manual isolation steps (e.g., nucleus extraction). This is particularly impactful for challenging samples with fragile RNA integrity, such as gastrointestinal tract tissues—previously recalcitrant to high-quality transcriptome profiling. Together, the combination of random primers and microfluidic automation allows imRandom-seq to generate high-fidelity transcriptomes even from micro-volume or highly degraded FFPE samples, outperforming conventional methods in both coverage and stability. The clinical value of imRandom-seq stems from its ability to unlock FFPE samples for transcriptomic analysis. imRandom-seq applies to common clinical tissues (e.g., breast, lung) and historically challenging ones (e.g., gastrointestinal tract), overcoming RNA integrity barriers that limited previous technologies. In various FFPE tissues and different types of FFPE samples, including those with low RNA quality or very small size, imRandom-seq demonstrated good performance on comprehensive transcriptome profiles across all three RNA sequencing applications, bulk, single-nucleus, and spatial RNA-seq. Compared to bulk RNA sequencing, which provides a blended cellular profile, both single-cell and spatial RNA-seq have facilitated life science and clinical research through their high-resolution capabilities30. Although some probe-based approaches have been applied to FFPE samples for sequencing and imaging methods, the limited subsets of RNAs still suppress comprehensive and systematic research. Within automatic pretreatment, our single-nucleus imRandom-seq provided comprehensive transcriptomic information at single-cell resolution and identified various oncogenic lncRNAs across different FFPE tumor tissues. Intriguingly, our data revealed that common oncogenic drivers across 6 tumor types were more frequently lncRNAs than protein-coding genes, suggesting non-coding RNAs may play conserved roles in tumorigenesis via epigenetic regulation (e.g., transcriptional activation, post-transcriptional processing)31–33. Another application is spatial RNA-seq for FFPE samples, which enables both histopathological assessment and molecular profiling to guide therapeutic strategies. Spatial imRandom-seq enhances this workflow by providing high-resolution transcriptomic data in parallel with standard histology. For instance, spatial RNA-seq from imRandom-seq can map gene expression patterns to morphological features (e.g., tumor nests, stromal regions), enabling correlations between molecular alterations and pathological phenotypes (e.g., grade, invasion). This integration improves diagnostic precision—for example, distinguishing between benign and malignant lesions based on both histology and transcriptomic signatures—and supports retrospective studies, where FFPE archives offer unparalleled access to long-term clinical outcomes. These findings highlight imRandom-seq’s potential to identify therapeutic targets or diagnostic signatures, particularly in retrospective studies leveraging archived FFPE cohorts. In conclusion, imRandom-seq represents a significant advancement in FFPE transcriptomics, merging microfluidic automation and random primer-based chemistry to overcome key limitations of existing technologies. Its ability to generate bulk, single-nucleus, and spatial transcriptomes from diverse FFPE samples—with comprehensive coverage of coding and non-coding RNAs—positions it as a powerful tool for clinical research and diagnostics. By bridging the gap between archived FFPE tissues and high-resolution molecular analysis, imRandom-seq unlocks opportunities for biomarker discovery, therapeutic development, and precision pathology, ultimately advancing the translation of genomic research into clinical practice. Methods Experimental model The FFPE samples were provided by the First Affiliated Hospital, Zhejiang University School of Medicine. No data regarding the sex and/or gender and age of participants were collected. All human tissue samples were destroyed after testing. Male and Female wildtype C57BL6/J mice (6–8 weeks of age) were ordered from Shanghai SLAC Laboratory Animal. Mice were single-housed under standard laboratory conditions, including a 12 h light/dark cycle, temperatures of 18–23 °C with 40–60% humidity, with free access to mouse diet and water. All experiments conformed to the relevant regulatory standards at Zhejiang University Laboratory Animal Center. FFPE samples of mouse tissues were prepared by Core Facilities, Zhejiang University School of Medicine. Samples used in this study were listed in Supplementary Data 1 within the sequencing depth. Automatic device As shown in Fig. 1, the device has a man-machine interface (set the program, and monitor the temperature and program progress), a chip module (load chip, with a temperature controller vortex mixer and a chip cover), a programmable pipetting system (for low temperature reagent) and pump priming system (for room temperature reagent), refrigerated reagent reservoirs (for low temperature reagent), a mobile device for pipetting and pump priming, and an waste collection system (liquid suction rate is 40 μL/s). For single-nucleus imRandom-seq (Supplementary Fig. 1): Turn on the machine and pre-cool the reagent storage rack to ensure optimal conditions for reagent stability and reaction efficiency. Carefully place the microfluidic chip onto the instrument platform and secure it with the provided clamps. Attach the waste outlet of the chip to the waste tube to facilitate the removal of used reagents and fluids during the process. Place the samples into the designated sample slots on the chip, ensuring they are properly seated and secure. Prepare and place all required reagents (listed in the Experimental procedures of bulk/single-nucleus imRandom-seq), including permeabilizing buffer, blocking reagent, RT reagent, dA tailing reagent, 0.2X SSCT, and PBST, onto the pre-cooled storage rack to maintain their activity and effectiveness. Fill the reagent bottles with dewaxing solution, RNase-free water, and ethanol. Start the instrument and select the pre-set program to run the process automatically from Paraffin dissolution to dA tailing and wash (Supplementary Data 3). After the procedure is completed, collect the sample from the sample slot and perform snRNA-seq following the Experimental procedures of single-nucleus imRandom-seq. For spatial transcriptomics imRandom-seq (Supplementary Fig. 2): Place the spatial transcriptomics modules on the reaction platform. Two modules can be placed, and each module can hold one tissue section. Prepare and place all required reagents (listed in the Experimental procedures of spatial transcriptomics imRandom-seq8), including permeabilizing buffer, blocking reagent, RT reagent, dA tailing reagent, 0.2X SSCT, and PBST, onto the pre-cooled storage rack to maintain their activity and effectiveness. Fill the reagent bottles with dewaxing solution, RNase-free water, and ethanol. Start the instrument and select the pre-set program to run the process automatically from Paraffin dissolution to dA tailing and wash (Supplementary Data 3). After the procedure is completed, collect the tissue section from the module and perform spRNA-seq following the Experimental procedures of spatial transcriptomics imRandom-seq. Design and fabrication of the microfluidic chip The microfluidic chip shown in Fig. 2d was designed using the computer-aided design software AutoCAD (2021, Autodesk, USA). The microfluidic chip has eight channels and enables the simultaneous processing of up to eight samples. Each channel includes a designated sample slot for loading FFPE scrolls and a waste slot that connects to a waste tube for efficient disposal. The established protocols were employed34 for fabricating the polydimethylsiloxane (PDMS) microfluidic chip. Molds for a microfluidic device were made using a photolithographic approach, consisting of centrifugal coating and modeling the SU-8. Silicon molds were employed for casting PDMS (Sylgard-184) to fabricate microfluidic chips. Experimental procedures of bulk/single-nucleus imRandom-seq (Manual operation process without microfluidic chip) RNA extraction from FFPE samples with the RNApure FFPE Kit (CWbiotech), and Qsep100 was used to detect the DV200 of RNA. After quality control, FFPE tissue scrolls (10–20 µm in thickness) were placed into a 1.5 mL centrifuge tube and suspended in 1 mL Dewax Solution. The tube was incubated at 37 °C for 7 min in the thermomixer with gentle shaking. Dewax Solution was removed, and the process was repeated twice (for a total of 3 times). Tissue was rehydrated at room temperature with 1 mL of 100%, 100%, 90%, 70%, 50%, 30%, and 0% ethanol (in RNase-free water) with 2 min incubations each. The rehydrated tissue was resuspended in 100 µL of permeabilizing buffer (1 mg/ml Pepsin in 0.1 N HCl) and incubated for 30 min at 37 °C in a thermomixer. After permeabilizing, the tissue was resuspended in PBST (0.05% Tween-20 in PBS) and gently shaken for 30 s (repeat this washing step for a total of 4 times). The tissue was transferred to the PCR tube with blocking reagent (Enzyme B 3 µL, dNTP (10 mM each) 15 µL, Buffer B1 19.5 µL, 10 µM Block Primer 9 µL, RNase Inhibitor 3.5 µL). The PCR tube was placed into the thermal cycler (program: 8 °C 8 min, 15 °C 8 min, 20 °C 8 min, 25 °C 8 min, 30 °C 8 min, 37 °C 50 min, 4 °C ∞; heated lid 50 °C). After blocking, 150 µL of 0.2X SSCT (0.05% Tween-20 in 0.2X SSC) was added to the tube and incubated with a new program (50 °C 5 min, 4 °C ∞; heated lid 60 °C). After the incubation, the tissue was transferred into a 1.5 mL centrifuge tube. The tissue was washed using 0.2X SSCT (a total of 2 times) and PBST (a total of 4 times) with a gentle shake for 30 s. The tissue was transferred to the PCR tube with RT reagent (Enzyme R 3 µL, Buffer RN 35 µL, RT Primer mix (RT Primer-N and RT Primer-T, 10 µM each) 9 µL, RNase Inhibitor 3 µL). The PCR tube was placed into the thermal cycler (program: 8 cycles (8 °C 1 min, 15 °C 3 min, 20 °C 3 min, 25 °C 3 min, 30 °C 3 min, 42 °C 3 min), 42 °C, 1 h; heated lid 50 °C). After the incubation, the tissue was transferred into a 1.5 mL centrifuge tube. The tissue was washed using PBST (a total of 4 times) with a gentle shake for 30 s. On the final supernatant removal, the remaining liquid in the tube was adjusted to a final volume of 54.6 µL, and dA tailing reagent (TdT enzyme 0.7 µL, TdT reaction buffer 7 µL, Buffer CoCl2 7 µL, 100 mM dATP 0.7 µL) was added. The centrifuge tube was placed into the thermal cycler and incubated for 30 min at 37 °C. After the incubation, the tissue was washed using PBST (a total of 3 times) with a gentle shake for 30 s. For bulk imRandom-seq: the tissue was transferred to the PCR tube with 4X MM mix 4 µL, 10 µM Bulk Primer 4 µL, PBST 8 µL. Then, the PCR tube was incubated at 37 °C for 1 h, 50 °C 30 min, 60 °C 30 min, and 75 °C 20 min. After the incubation, the cDNA was purified by Ampure XP beads, eluting with 38.5 µL of nuclease-free water. Then, cDNA amplification using DeepVent DNA polymerase 0.5 µL, dNTP (10 mM each) 1 µL, 10X Thermopol buffer 5 µL, 10 µM PCR-F/R 5 µL with program (95 °C, 2 min, 20 cycle (95 °C, 15 s; 60 °C, 15 s; 72 °C, 2 min), 72 °C, 4 min, 4 °C ∞). And library construction was performed using Universal DNA Library Prep Kit, according to the manufacturer’s instructions. Libraries were compatible with both Illumina and BGI sequencing platforms, with the recommended sequencing type being PE150. For single-nucleus imRandom-seq: the tissue was ground with a grinder pestle for 5 min. After grinding, 1 mL lysis buffer (PBS 970 µL, 1 mg/mL Proteinase K 10 µL, 10% Tween-20 10 µL, 10% Triton x-100 10 µL) was added. The centrifuge tube was placed into the thermal cycler and incubated for 5-20 min at 37 °C. The lysate was filtered through a 10–40 µm cell strainer into a new 1.5 mL centrifuge tube. The nuclei extract was centrifuged at 700 × g for 5 min at 4 °C. After removing the supernatant, the pellet was resuspended and washed with PBST (repeat this washing step for a total of 4 times). The nuclei were stained with DAPI and counted under a fluorescent microscope. Nuclei, 4X MM mix, and Barcoded Beads were encapsulated into droplets using the microfluidic platform. Then, 120 µL emulsions were incubated at 37 °C for 1 h, 50 °C 30 min, 60 °C 30 min, and 75 °C 20 min. After the barcoding reaction, droplets were broken by mixing with 30 µL 50 mM EDTA and 150 µL 20% (vol/vol, in HFE-7500) PFO. The aqueous phase was taken out and purified by 1X Ampure XP beads, eluting with 38.5 µL of nuclease-free water. Then, cDNA amplification using DeepVent DNA polymerase 0.5 µL, dNTP (10 mM each) 1 µL, 10X Thermopol buffer 5 µL, 10 µM PCR-F/R 5 µL with program (95 °C, 2 min, 20 cycle (95 °C, 15 s; 60 °C, 15 s; 72 °C, 2 min), 72 °C, 4 min, 4 °C ∞). And library construction was performed using Universal DNA Library Prep Kit, according to the manufacturer’s instructions. Libraries were compatible with both Illumina and BGI sequencing platforms, with the recommended sequencing type being PE150. Experimental procedures of spatial transcriptomics imRandom-seq The procedures were performed as previously described with some optimizations8. FFPE sections were cut from the paraffin block and kept at 60 °C for 2 h. The sections were then washed three times with 100 µL Dewax Solution at 37 °C for 20 min. The sections were rehydrated at room temperature with 100 µL of 100%, 100%, 90%, 70%, 50%, 30%, and 0% ethanol (in RNase-free water) with 2 min incubations each. The sections were washed three times with 100 µL pre-cold wash buffer (1X PBS with 1U/μL RNase inhibitor) at room temperature for 5 min. The FFPE sections were then incubated in 100 µL permeabilizing buffer at 37 °C for 15 min and then washed three times with 2X SSC buffer (2X SSC buffer with 1U/μL RNase inhibitor). The FFPE sections were incubated with the RT reagent with multiple annealing ramping from 8 °C to 42 °C and overnight at 42 °C. After incubation, the sections were washed with 2X SSCT (2X SSC with 0.05% Tween-20) three times to wash away the residual random primers and primer dimers. The FFPE sections were incubated with the dA tailing reagent at 37 °C for 30 min, and then were washed with 2X SSCT three times. The transfer mix was prepared: 34.5 μL ddH2O, 5 μL RNase H reaction buffer, 5 μL 1 mg/mL Proteinase K, 3 μL RNase H, and 2.5 μL 10% TritonX 100. Following dA tailing, the cDNA in the FFPE sections was transferred into a gene expression slide with the transfer mix at 37 °C for 30 min. The extension mix was prepared: 25 μL 4X MM Mix, 75 μL ddH2O. The gene expression slide was incubated with the extension mix at 60 °C for 60 min. Following DNA extension, the single cDNA containing the spatial barcode was eluted using 50 μL 0.08 M KOH and 3 μL 1 M Tris-HCl pH 8.0. Amplify the mix using pre-amplification mix for 10 cycles and cleanup by Ampure XP beads. The program was: 98 °C 3 min;10 cycles of 98 °C 15 s, 63 °C 20 s, 72 °C 30 s; 72 °C 1 min; 4 °C ∞. Then the library construction was performed using the Universal DNA Library Prep Kit, according to the manufacturer’s instructions. Libraries were compatible with both Illumina and BGI sequencing platforms, with the recommended sequencing type being PE150. Paraffin—SweAMI -FISH +IF protocol The specific fluorescently labeled PVT1 FISH probes (5′-3′: TCCAGCTTTAGGTCACGTAAGGACA, GACTGGCTCTTAATTCTCCAATCTCA, CAGATGAACCAGGTGAACACAGAGC) were designed and synthesized by Servicebio (Wuhan, China). Tissue paraffin blocks were sliced by slicer with a thickness of 4μm, fished by spreading machine, and baked in 62 °C oven for 2 h. Sequentially put the slices into 2 changes of BioDewax and Clear solution, 15 min each, then dehydrate in 2 changes of pure ethanol for 5 min each, then put into 85% alcohol and 75% alcohol, 5 min each, and soak them in DEPC water. Put the slices into the citric acid (pH 6.0) repair box and place it in a water bath at 90 °C for 48 min. After natural cooling, the stroke circle was assembled, and washed with pure water, PBS washes for 3 times, 5 min each time. Add pre-hybridization solution and incubate at 37 °C for 1 h. Pour out that pre-hybridization solution, add the hybridization solution containing probes, and hybridize overnight in a constant temperature box (40 °C). Wash off the hybridization solution, 2X SSC, 37 °C for 10 min;1X SSC, washed at 37 °C for 2 × 5 min; 0.5X SSC was washed at room temperature for 10 min. Gently spin-dry the slice, drop the preheated branch probe hybridization solution (60 μL), and horizontally put it in a wet box for hybridization at 40 °C for 45 min. In this process, 50 ml 2X SSC was added to the bottom of the wet box to prevent drying. Pour off the hybridization solution, and rinse the slices with 2X SSC, 1X SSC, 0.5X SSC and 0.1X SSC preheated at 40 °C in turn for 5 min. Adding the hybridization solution containing the signal probe, and incubating at 42 °C for 3 h at a dilution ratio of 1:200. After that, they were washed with SSC buffer in turn: 2X SSC, 37 °C for 10 min, 1X SSC, 37 °C for 2 × 5 min, and 0.5X SSC at 37 °C for 10 min. Pour out PBS, add 3%BSA, and seal for 30 min. Remove the blocking solution, drop the prepared primary antibody (Anti-EpCAM Mouse mAb, dilution 1: 2000; GB12274, Servicebio), and lay it flat in a wet box for incubation at 4 °C overnight. The slides were put 5 min PBS (PH7.4) and washed on a decoloring shaker for 3 times, each time for 5 min. Drop the corresponding fluorescein-labeled secondary antibody and incubate at room temperature for 50 min. DAPI solution was dripped into the slice and incubated at room temperature for 8 min in the dark, then washed with running water, cover slip was placed with anti-fade mounting medium. UV glows blue by UV excitation wavelength 330–380 nm and emission wavelength 420 nm; Alexa Fluor 488 glows green by excitation wavelength 465–495 nm and emission wavelength 515-555 nm; CY3 glows red by excitation wavelength 510–560 nm and emission wavelength 590 nm. Standard RNA-seq for FFPE and FF (fresh frozen) samples Total RNA was isolated and purified from the FFPE sample using the Qiagen RNeasy FFPE Kit (Cat. No. 73504), fresh frozen sample using TriZol (Ambion, Cat # 15596026CN) according to the manufacturer’s protocol. The quantity and purity of the extracted RNA were assessed using a NanoDrop ND-1000 spectrophotometer (NanoDrop, Wilmington, DE, USA). RNA integrity was evaluated with the Agilent Bioanalyzer 2100 (Agilent, CA, USA) and further verified by agarose gel electrophoresis. Samples meeting the criteria of DV200 > 30% or RIN > 4.0 were deemed suitable for downstream applications. Library preparation was performed using the SMARTer Stranded Total RNA-Seq Kit v2—Pico Input Mammalian (Takara Bio USA, Inc.), starting with 10–50 ng of total RNA. The RNA was mixed with SMART Pico Oligos Mix and denatured at 72 °C for 3 min. Reverse transcription was then carried out at 42 °C for 90 min using the Template Switching Oligo Mix (TSO mix) as the reverse transcription primer. The resulting cDNA was amplified for 5 cycles using the SeqAmp DNA Polymerase system to incorporate sequencing adapters and sample index sequences. The amplified DNA library was purified using SPRI beads, and the pellet was resuspended in ZapR-Master-Mix to generate the initial crude library. To deplete ribosomal cDNA, the library was denatured in the presence of R-Probes v2, followed by incubation with ZapR v2 at 37 °C for 60 min. The rRNA-depleted product was then amplified using universal paired-end primers. The number of PCR cycles was adjusted based on input amount and RNA quality, typically ranging from 12 to 16 cycles. The final library was purified again with SPRI beads and eluted in nuclease-free water. Libraries were stable for up to 2 months when stored at –20 °C. Sequencing was performed on an Illumina NovaSeq 6000 (LC Bio Technology Co., Ltd., Hangzhou, China) using standard protocols with a paired-end 150-bp (PE150) configuration. Preprocessing and clustering analysis of traditional snRNA-seq & bulk/single-nucleus imRandom-seq data The preprocessing and clustering analysis of raw sequencing data was automated by VITAseer (v1.2) with mode R201221 for traditional snRNA-seq data and mode R201231 for imRandom-seq data, respectively. Raw sequencing data were cleansed and subsequently mapped to the reference genome to generate the gene expression matrix. For bulk imRandom-seq, the gene expression profile was estimated by read counts along the gene body. For snRNA-seq and single-nucleus imRandom-seq, the downstream analysis was performed, including quality control, clustering, and differential expression testing. Integrations were performed across the paired snRNA-seq and single-nucleus imRandom-seq data by Suerat and Integration_SCP function in SCP (0.5.6) with parameters (batch = “orig.ident”, integration_method = “Suerat”). Scaling and principal component analysis (PCA) were then applied to the integrated dataset, followed by the shared nearest-neighbor (SNN) graph construction using FindNeighbors with 30 PCs. Clusters were identified using the FindClusters function with a resolution of 0.6, and visualized with the Uniform Manifold Approximation and Projection (UMAP) algorithm as implemented by the CellDimPlot function in SCP. Marker genes were identified using FindAllMarkers function in Seurat (4.3.0) and filtered by the criteria (only.pos = TRUE, min.pct = 0.1, logfc.threshold = 0.25). The cell-type annotation for each cluster was determined manually using published lists of marker genes. SCP and ggplot2 (3.4.4) in R (4.3.2) were used to generate the raw plots. Genebody coverage The gene body coverage from 5′ end to 3′ end is calculated by using the “genebody_coverage.py” script from RSeQC (v4.0.0) software. Comparison between 10X Flex and single-nucleus imRandom-seq lung cancer data The filtered expression matrix of the Lung Cancer sample generated by the 10X Flex platform was downloaded from the 10X Genomics website (https://www.10xgenomics.com/datasets/40k-mixture-of-cells-dissociated-from-4-fixed-tumor-tissues-multiplexed-samples-4-probe-barcodes-1-standard). The pre-processing steps of the fastq files of the Lung Cancer dataset generated by imRandom-seq was the same as the steps mentioned in the previous section. The resulting expression matrices for both technologies were processed with Seurat V5 (v5.0.3) package in R (v4.3.3). We first filtered out genes that were expressed in fewer than 3 cells. The rRNAs were further removed based on the rRNA annotation in the GTF file. The “percent.mt” values were calculated with the “PercentageFeatureSet()” function, and cells with “percent.mt” values that are higher than a cutoff (mean(percent.mt) + 2*sd(percent.mt)) were excluded from the downstream analysis. Then the 10X Flex and imRandom-seq datasets were integrated together, within scaling and normalizing by “ScaleData()” and “NormalizeData()” functions. The top 3000 highly variable genes were found with the “FindVariableFeatures()” function. To perform the dimension reduction, both datasets were re-normalized with the “SCTransform()” function by regressing out the “percent.mt”. To further debatch the technical bias between 10X Flex and imRandom-seq, the Integration_SCP function in SCP (0.5.6) was performed with parameters (batch = “orig.ident”, integration_method = “Suerat”). Scaling and PCA were then applied to the integrated dataset, followed by the SNN graph construction using FindNeighbors with 30 PCs. Clusters were identified using the FindClusters function with a resolution of 0.6, and visualized with the UMAP algorithm as implemented by the CellDimPlot function in SCP. To investigate the overlapping of captured genes between two technologies, we limited the gene’s biotype to only protein-coding. The Venn plot was done by “ggvenn()” function in ggvenn (v0.1.10) package in R. To compare the ability of capturing the total transcriptome, we displayed the number of genes captured with seven different biotypes in 10X Flex data and imRandom-seq data with a barplot generated with ggplot2 package in R. Cell type annotation Manual annotation for each cell cluster was carried out using the top 100 markers according to published research papers, referencing annotation information by the ACT database (Annotation of Cell Types, http://xteam.xbio.top/ACT/)35. (For most datasets, markers of each cluster were identified by the “FindAllMarkers()” function with parameters (only.pos = TRUE, min.pct = 0.25, logfc.threshold = 0.5). For the tumor dataset of kidney cancer (KC) and ovarian cancer (OC), markers were identified with default parameters (only.pos = TRUE, min.pct = 0.1, logfc.threshold = 0.25). The list of top 100 canonical markers is provided in Source Data. Identification of markers and differentially expressed genes The markers of different cell types or markers of tissue-specific epithelial cells were identified by the “FindAllMarkers()” function with parameters (only.pos = TRUE, min.pct = 0.25, logfc.threshold = 0.5), and the top 100 genes ordered by “avg_log2FC” were then picked up for heatmap. The differentially expressed genes were analyzed by the “FindMarkers()” function using the Wilcoxon test, and genes with adjusted p-values less than 1e-3 and fold change more than 2 were defined as differentially expressed genes. Gene ontology analysis Gene ontology (GO) analysis of differentially expressed genes was performed using DAVID. GO terms with p < 0.05 were determined to be statistically significant. Alternative splicing analysis rMATS-turbo (v4.1.223) pipeline was employed to identify all 5 types of alternative splicing events with parameters “-t single –allow-clipping –variable-read-length --libType fr-firststrand”. The candidate alternative splicing events were retained for downstream analysis if there are >2 reads supporting both inclusion and exclusion junctions, and the FDR < 0.05. Calculating CNV inference and estimating CNV score The InferCNV package was applied to infer the CNVs36, and the expression matrix was first calculated by inferCNV within default parameters. For each tumor, the epithelial cells from the corresponding sample and annotated immune cells from all samples, including plasma cells, B cells, T cells, and myeloid cells, were utilized to estimate their CNV score. Referring previous study37, the matrix obtained by inferCNV was then imputed by incorporating the expression level of the cell itself and its nine adjacent cells, as determined by the neighborhood graph of clustering results. The neighborhood graph of epithelial cells was determined by an individual clustering map for each tumor sample (Supplementary Fig. 4d), and immune cells grouped into different cell types were further determined by a global clustering map, respectively (Fig. 5a). Next, every ordered 50 genes along chromosome loci were merged, defined as one bin. For each cell from epithelial cells and immune cells, the intensities of bins were adjusted by the average intensity from immune cells, and the adjusted intensity of each bin was defined as the CNV level, which was limited to -3 to 3. Moreover, the CNV score for each cell was determined by calculating the sum of the squares of CNV levels. Malignant cell identification For each tumor, epithelial cells were employed for hierarchical clustering according to CNV levels along the genome, and grouped into eight subclones as the default setting. Then, subclones characterized by relatively low CNV scores were identified as nonmalignant cells for the respective tissue, while the remaining cells were classified as malignant. Statistical analysis of differentially expressed genes in LUAD datasets The transcriptomic profiles were obtained from The Cancer Genome Atlas (TCGA) database, including 524 LUAD and 58 related normal samples. The edgeR package was applied for downstream analysis with default settings38. Genes with adjusted p-values less than 1e-3 and fold change more than 2 were defined as differentially expressed genes. Survival analysis Data, including transcriptomic profiles and corresponding clinical information, were obtained from the Cancer Genome Atlas (TCGA) database. The 173 ESCA, 529 KIRC, 504 LUAD, 399 OV, and 103 SKCM patients from the TCGA database were applied for survival analysis of non-coding oncogenes detected by imRandom-seq datasets of esophageal cancer (EC), kidney cancer (KC), primary lung cancer (LC_primary), melanoma (MA), and ovarian cancer (OC), respectively. As few patients carrying metastatic LUAD were found in the TCGA datasets, the survival analysis was not performed for non-coding oncogenes in metastatic lung cancer (LC_metastasis). For each oncogene, the maximally selected rank statistics method was used to determine the optimal expression cutoff for stratifying patients into high- and low-expression groups, with candidate cutoffs restricted to the 20th–80th percentile range of expression values. Only TCGA primary tumor samples (sample code “-01A-”) were selected for downstream analysis. Subsequently, Kaplan-Meier curves and two-sided log-rank tests were used to compare survival distributions. Univariate Cox regression provided hazard ratios (HR) and 95% confidence intervals (CI). The patient numbers in low/high groups, HRs, CIs, P values, and other relevant information are provided in Supplementary Data 4. Construction of the ceRNA regulatory network Interaction data among ceRNAs, microRNAs, and mRNAs were extracted from the LncACTdb (http://bio-bigdata.hrbmu.edu.cn/LncACTdb/index.html). Acting as potential ceRNA, lncRNAs showing significant upregulation in malignant cells of each tumor type were initially identified, and their target mRNAs were subsequently retrieved from the same database, retaining only those that were also differentially expressed in malignant cells. The regulatory network was visualized using Cytoscape software (https://cytoscape.org/). Preprocessing and alignment of the spatial transcriptomics data To pre-process the spatial transcriptomics imRandom-seq data, the barcode (BC) and UMI sequences for each read were extracted from the fastq file for R1 based on their standard lengths (16 bp BC + 12 bp UMI). For read 2, the polyA sequence was trimmed for each read, and we limited the read length (after polyA removal) to 90 bp at most to match the read length of typical 10X Visium results. The processed fastq files were then aligned to the mouse reference genome (mm39) and human genome (hg38) downloaded from GENCODE with STARsolo (v2.7.10b). Spatial transcriptome data analysis and cell deconvolution For data generated by spatial transcriptomics imRandom-seq, the spots were first selected manually based on the eosin images, and then the spots with fewer than 200 genes detected were further removed. For all the data, the normalization (SCTransform), dimension reduction (runPCA, runUMAP), clustering (FindCluster), and differential expression analysis (FindAllMarkers) were performed with built-in functions in Seurat (v4.3.0.1) package in R (v4.2.1) with all default parameters. The cell deconvolution of the data from spatial transcriptomics imRandom-seq data was estimated by CARD package (v1.1) in R with all default parameters29. The reference single-cell RNA-seq data for the mouse brain was the Adolescent Mouse Brain dataset downloaded from the Mouse Brain Atlas (https://mousebrain.org/). The reference single-cell RNA-seq data for mouse lung was derived from paired single-nucleus and spatial transcriptomics imRandom-seq data using the same FFPE mouse lung tissue. Statistics and reproducibility No statistical method was used to predetermine sample size. No data were excluded from the analyses. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. Ethical statement The collection of human samples in this study was approved by the Research Ethics Committee of the First Affiliated Hospital, Zhejiang University School of Medicine (IRB number: IIT20220893A). Written informed consent was obtained from all participants or their legal guardians prior to sample collection and experimentation. This study is compliant with the Guidance of the Ministry of Science and Technology (MOST) for the Review and Approval of Human Genetic Resources (approval numbers: 2026BAT00329, 2026BAT00328). Mouse experiments in this study were approved by the Animal Ethics Committee of Zhejiang University (IRB number: ZJU20240395); experiments conformed to the regulatory standards at Zhejiang University Laboratory Animal Center. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Information Supplementary Movie 1 Supplementary Movie 2 Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgement The project was supported by the National Key R&D Program of China (2025YFF0512800, Y.W.), the State Key Laboratory of Space Medicine, China Astronaut Research and Training Center (SKL2024K04), the Pioneer R&D Programs of Zhejiang Province (2024C03005, Y.W.), the Key R&D Program of Zhejiang (2024SSYS0022, Y.W.). Author contributions H.C. and Y.S.C. conceived and designed the research; Y.W. provided materials and supervised the program; H.C., S.Z., H.Y. and Y.L. performed imRandom-seq experiments; Y.S.C., J.C. and T.F. performed data analysis; J.Z., Y.L. and Q.L. designed and manufactured the instrument and chip. H.C., Y.S.C. and Y.W. wrote the paper. All authors have given approval to the final version of the manuscript. Peer review Peer review information Nature Communications thanks Rui Chen, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Data availability The RNA sequencing data are available at the Genome Sequence Archive (GSA) (Accession HRA007241, HRA012650, and CRA028651). Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request. The expression matrix of the Lung Cancer sample generated by 10X Flex platform was downloaded from the 10X Genomics website (https://www.10xgenomics.com/datasets/40k-mixture-of-cells-dissociated-from-4-fixed-tumor-tissues-multiplexed-samples-4-probe-barcodes-1-standard). TCGA datasets are downloaded from GDC Data Portal using TCGAbiolinks R package. Source data are provided with this paper. Code availability The scripts used for downstream analyses and to produce all the figures in the manuscript are deposited in Zenodo (10.5281/zenodo.20278269)39 and mirrored on Github (https://github.com/WangycLab/imRandom-seq). Competing interests H.C. and H.Y. have submitted a patent application to the Chinese patent office for part of the technologies in this study (application number: CN202410244039.1). Y.W. and Q.L. are cofounders, H.C., Y.S.C., H.Y., T.F. and J.Z. are employees of M20 Genomics. The remaining authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Haide Chen, Yu-Sheng Chen, Shunji Zhang, Jiaye Chen. Contributor Information Yuan Liao, Email: 0619834@zju.edu.cn. Qian Liang, Email: qian.liang@m20genomics.com. Yongcheng Wang, Email: yongcheng@zju.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-74203-9. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Information Supplementary Movie 1 Supplementary Movie 2 Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The RNA sequencing data are available at the Genome Sequence Archive (GSA) (Accession HRA007241, HRA012650, and CRA028651). Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request. The expression matrix of the Lung Cancer sample generated by 10X Flex platform was downloaded from the 10X Genomics website (https://www.10xgenomics.com/datasets/40k-mixture-of-cells-dissociated-from-4-fixed-tumor-tissues-multiplexed-samples-4-probe-barcodes-1-standard). TCGA datasets are downloaded from GDC Data Portal using TCGAbiolinks R package. Source data are provided with this paper. The scripts used for downstream analyses and to produce all the figures in the manuscript are deposited in Zenodo (10.5281/zenodo.20278269)39 and mirrored on Github (https://github.com/WangycLab/imRandom-seq).