MainThe diversity of proteins produced by cells is typically defined by the repertoire of individual genes encoded in the genome. This diversity can be amplified through pre-mRNA cis-splicing, where varied pairings of exons within a given transcript can be differentially fused. It is unclear whether mechanisms beyond cis-splicing exist to diversify the protein-coding capacity of mammalian cells. In unicellular and invertebrate organisms, the process of trans-splicing, whereby exons from distinct pre-mRNAs are fused to create hybrid proteins, has been reported. In trypanosomes and nematodes, a common pre-mRNA is trans-spliced to various other mRNAs to promote transcript stability and translation3; by contrast, in Drosophila, this process generates functionally diverse transcripts and proteins4. In healthy mammals, few examples of trans-splicing exist, and functional chimeric mRNAs (chRNAs) are typically associated with oncogenic transformation of cells where genomic translocations fuse disparate genes at the DNA level5,6. Expression of some chRNAs known to be produced by DNA translocation have also been described in non-malignant tissue and are proposed to form through trans-splicing7,8,9. However, it remains to be elucidated whether, in healthy mammalian cells, widespread fusion of mRNA from distinct genes can produce chRNAs that encode functional proteins. However, one can predict that such fusion events would expand the number of physiological protein-encoding mRNAs far beyond what is currently appreciated in our genomes.Despite advances and widespread adoption of RNA sequencing (RNA-seq) methodologies, endogenously expressed chRNAs in mammals have largely evaded discovery because of technical limitations. cDNA synthesis-based approaches are used extensively and rely on viral reverse transcriptase enzymes to convert RNA to cDNA. However, viral reverse transcriptase enzymes can generate artificial fusion transcripts through template switching10. Furthermore, as mammalian chRNAs are not annotated in reference transcriptomes, candidate chRNAs multimap to different parts of the genome during alignment, resulting in their routine removal during standard RNA-seq analyses. Short-read sequencing also limits chRNA detection, as read fragments are assigned as chimeric only if they span the junction at which one gene meets the other; otherwise, these reads are assigned to the parent genes of the chRNA11. Moreover, additional challenges arise from difficulties in accurately resolving multimapping reads that display short junction sequences12 and/or map to multiple similar loci13. Notwithstanding these caveats, conventional RNA-seq datasets have consistently identified putative chRNAs in healthy human tissue14,15. Whether these chRNAs arise artificially during library preparation, or whether they are physiologically relevant, remains an open question.Inflammation drives chRNA expressionTo identify chRNAs while avoiding artifacts from cDNA synthesis, we performed Oxford Nanopore PromethION direct RNA-seq analysis of polyadenylated RNA from steady-state, tissue-reparative and inflammatory mouse bone-marrow-derived macrophages (BMDMs; Fig. 1a). Ten biological replicates yielded 52.9 million passing reads (Extended Data Fig. 1a and Supplementary Table 1), of which over 90% mapped to the mouse genome (Extended Data Fig. 1b). Read identity (Extended Data Fig. 1c), average Phred scores (Extended Data Fig. 1d) and read lengths (Extended Data Fig. 1e) were consistent with high-quality published datasets, and marker-gene expression confirmed macrophage polarization states (Extended Data Fig. 1f).Fig. 1: Inflammation drives chRNA expression.a, Schematic of chRNA identification and differential expression analysis in mouse BMDMs. The diagram was created using BioRender; Jackson, R. https://BioRender.com/3le8vsp (2026). b, The chromosomal distribution of 30,390 exon–exon chRNAs detected for downstream analysis. c, Classes of chRNA species identified using long-read RNA-seq. d, Scaled normalized short-read fusion fragments per million (FFPM)-like values for chRNAs co-detected in long-read and short-read RNA-seq data in BMDMs stimulated as indicated for 24 h. Selected chRNAs are highlighted. e, Classes of chRNA species co-detected in long-read and short-read RNA-seq. f, Scaled normalized counts for chRNAs using NanoString nCounter analysis in BMDMs stimulated as indicated for 24 h. Selected chRNAs are highlighted. g, Classes of chRNA species co-detected using NanoString nCounter and long-read RNA-seq. h, PCR analysis of Gsdmd-Tmem106a (G-T) cDNA expression in BMDMs. The gel image is representative of three independent experiments. i, Chromatogram of the Gsdmd-Tmem106a junction in cDNA from C57BL/6J BMDMs. For i–k, the dotted line denotes the division between Gsdmd- and Tmem106a-derived sequences. j,k, Chromatograms of the Gsdmd-Tmem106a junction in cDNA from BMDMs from BALB/cJ (j) and WSB/EiJ (k) mice. l–n, BMDMs were treated as indicated and the gene expression of Gsdmd-Tmem106a (l), Gsdmd (m) and Tmem106a (n) was analysed using RT–qPCR. Gene expression normalized to Polr2a fold change (FC) over steady state is depicted. Data are mean ± s.e.m. n = 5 biological replicates. o–q, RNA extracted from macrophages after infection or mock infection for NanoString nCounter analysis. The mean of normalized counts is shown. The line represents the mean; each datapoint represents a biologically independent animal. o, Mice were treated intranasally with influenza A virus or PBS (7 days) and CD45+F4/80+CD64+ cells were isolated from the lungs. p, Mice were infected intracisternally with E. coli or PBS (18 h) and CD11b+ cells were isolated from the brain. q, Mice were treated intraperitoneally with LPS or PBS (16 h) and F4/80+ cells were isolated from peritoneal lavage. P values were determined using one-way analysis of variance (ANOVA) (l–n) and unpaired Student’s two-tailed t-tests (o–q).Source dataThe computational tools LongGF, JAFFAL and Genion16,17,18 were benchmarked to detect BCR-ABL1 in K562 cells19 (Supplementary Table 2) and used to identify chRNAs containing annotated splice donor and acceptor sites from two distinct genes. This identified 30,390 candidates across activation states (Fig. 1b, Extended Data Fig. 1g and Supplementary Tables 3 and 4). These chRNAs had a median length of 1,681 nucleotides, compared with 2,190 nucleotides for their parent transcripts (Extended Data Fig. 1h). Parent genes were distributed throughout the genome (Fig. 1b), with around 88% of chRNA species arising from interchromosomal pairings (Fig. 1c) and about 4.5% and 7.5% from proximal (<106 bases) and distal (≥106 bases) intrachromosomal loci, respectively (Fig. 1c and Extended Data Fig. 1i), suggesting contributions from both readthrough transcription via cis-splicing and trans-splicing from distal loci.Limited sequencing depth and modest agreement among the long-read detection tools (Extended Data Fig. 1g and Supplementary Table 1) precluded robust quantification, prompting orthogonal validation. We generated high-depth Illumina RNA-seq data (Supplementary Table 5) and analysed these data using six fusion-detection algorithms20,21,22,23,24, producing an implausibly large candidate set (Supplementary Table 6) consistent with false positives generated through reverse-transcription and PCR template switching10. Nevertheless, comparison with the direct RNA catalogue identified over 250 chRNAs with matching junctions (Supplementary Table 7), including candidates associated with differential macrophage polarization (Fig. 1d). Although direct RNA-seq analysis predominantly detected interchromosomal chRNAs, co-detected chRNA species were largely intrachromosomal (Fig. 1e and Extended Data Fig. 1j), with proximal species detected more frequently and supported by more reads than distal or interchromosomal species (Extended Data Fig. 1k–m), consistent with the prevalence of local and readthrough transcription25,26.As short reads rarely span sufficient sequence on both sides of a chRNA junction, this approach can also generate false negatives for specific isoforms11. We therefore designed tandem NanoString nCounter probes that require simultaneous hybridization across each RNA junction (Fig. 1a). Over 500 probe sets were designed to target chRNAs detected by all three long-read tools (Supplementary Table 8), alongside scrambled controls with the expectation of false negatives. More than 100 high-quality probes detected constitutive or polarization-regulated chRNAs (Fig. 1f and Supplementary Table 7). Some chRNA species were validated by both Illumina and NanoString, whereas others were unique to either platform (Fig. 1d,f and Supplementary Table 7); notably, NanoString validated a greater proportion of interchromosomal species (Fig. 1e,g).PCR and Sanger sequencing independently validated chRNA junctions (Extended Data Fig. 2a–o and Supplementary Table 9), including a Gsdmd-Tmem106a amplicon spanning Gsdmd exons 1–2 fused to Tmem106a exons 6–9 (Fig. 1h–i). Gsdmd-Tmem106a was among the most inflammation-induced NanoString candidates (Fig. 1f), and both parent genes have established roles in lipopolysaccharide (LPS)-induced inflammation2,27. The same exon–exon junction was detected in C57BL/6J, BALB/cJ and wild-derived mice (Fig. 1j–k). Analysis using quantitative PCR with reverse transcription (RT–qPCR) further confirmed LPS-inducible expression, peaking at 6 h in parallel with the parent genes (Fig. 1l–n). The junction and LPS responsiveness of Cd274-Lacc1 were similarly confirmed (Extended Data Fig. 2p–s). In vivo, Gsdmd-Tmem106a was detected and induced in lung macrophages after influenza A infection, brain macrophages/microglia during Escherichia coli meningitis and peritoneal macrophages after LPS administration (Fig. 1o–q).Applying this pipeline to steady-state and inflammatory human monocyte-derived macrophages (Extended Data Fig. 3a–g and Supplementary Table 10), we identified over 900 chRNAs (Extended Data Fig. 3h–i and Supplementary Table 11), including inter- and intrachromosomal species (Extended Data Fig. 3j) with properties resembling those of mouse chRNAs (Extended Data Fig. 3k–l). qPCR further confirmed their detection and differential regulation by LPS (Extended Data Fig. 3m–o).Despite lower sequencing depth in the human dataset (Extended Data Fig. 3b and Supplementary Table 10), cross-species analysis identified 33 chRNAs sharing parent genes between mice and humans (Extended Data Fig. 4a and Supplementary Table 12). Although we did not identify a GSDMD-TMEM106A chRNA in these cells, approximately 8 displayed at least 50% mRNA sequence homology and 28 encoded predicted proteins with at least 50% conservation (Extended Data Fig. 4a,b and Supplementary Table 12), accompanied by closely matched domain architectures (Extended Data Fig. 4c). These included HDAC8-CITED1 and TBC1D22B-RNF8, which retained functional domains from both parent genes (Extended Data Fig. 4d–e) and were validated by Sanger sequencing in mouse and human macrophages (Extended Data Fig. 4f–g).Thus, integrating direct RNA-seq with short-read RNA-seq, NanoString, PCR and expression analyses provides a generalizable framework for identifying, validating and prioritizing endogenous chRNAs for functional studies in immunity and inflammation.Inflammation controls chRNA gene contactTo distinguish chromosomal translocation from RNA-level fusion, we performed long-range genomic PCR analysis of DNA from untreated and LPS-stimulated BMDMs using primers positioned in the junction exons, including Gsdmd exon 2 and Tmem106a exon 6. Neither Gsdmd-Tmem106a nor Cd274-Lacc1 produced the expected kilobase-scale genomic amplicon under either condition (Extended Data Fig. 5a,b), arguing against detectable chromosomal translocations and supporting RNA-level fusion through trans-splicing.As our chRNA catalogue was defined using annotated splice sites, we tested whether canonical splicing machinery contributes to chRNA formation. Inhibition of the SF3B spliceosome complex with pladienolide B28,29 or knockdown of U1 small nuclear RNA abolished LPS-induced Gsdmd-Tmem106a expression (Fig. 2a,b). Inhibition of RNA polymerase II with actinomycin D30 similarly eliminated Gsdmd-Tmem106a and Cd274-Lacc1, as well as their parent transcripts (Fig. 2c–e and Extended Data Fig. 5c–e), supporting the co-transcriptional joining of newly synthesized RNAs.Fig. 2: Inflammation controls chRNA gene contact.a, BMDMs were treated with LPS for 6 h in the presence of DMSO or pladienolide B (PlaB) and analysed using RT–qPCR. b, BMDMs were transfected with non-targeting (siCtrl) or Rnu1-targeting (siRnu1) siRNA, treated with LPS for 6 h and analysed using RT–qPCR. n = 4 mice per group. c–e, BMDMs were left non-treated (n = 4 mice) or treated with LPS and DMSO or actinomycin D for 3 h (n = 5 (Gsdmd-Tmem106a) (G-T) and n = 6 (Gsdmd and Tmem106a) mice per group). The relative expression of Gsdmd (c), Tmem106a (d) and Gsdmd-Tmem106a (e) was analysed using RT–qPCR. f, Aggregate peak analysis (APA) of in situ Hi-C contacts between paired exons of interchromosomal chRNAs or random exons. g, RT–qPCR analysis of BMDMs that were transfected with siCtrl or Ctcf-targeting siRNA (siCtcf) and treated with LPS for 6 h. n = 6 mice per group. h, LPS-induced changes in the contact frequency at chRNA interaction sites in siCtrl- and siCtcf-transfected BMDMs. For the box plots, the central dot represents the median, the upper and lower hinges represent the 25th and 75th percentiles, and the upper and lower whiskers extend to 1.5× the interquartile range (IQR) of the upper and lower hinges, respectively. i, Density plot of log2[fold change (FC) ratio] interactions with FC ratios > 1 in LPS-stimulated siCtrl BMDMs. Ratios were calculated by dividing the LPS-stimulated FC by the corresponding unstimulated FC for each siRNA condition with log2[FC ratio] values plotted. j, RT–qPCR analysis of LPS-stimulated BMDMs that were transfected with siCtrl or siCtcf. n = 6 (Gsdmd and Gsdmd-Tmem106a) and n = 4 (Tmem106a) mice per group. k, The primer sets used to amplify alternative Gsdmd-Tmem106a exon combinations. The diagram was created using BioRender; Venezia, O. https://BioRender.com/xvepcox (2026). l, Gel electrophoresis of LPS-treated BMDM cDNA using primers shown in k. The gel image is representative of two independent experiments. RT–qPCR expression was normalized to Polr2a and plotted as fold change (FC) over the indicated control. Each datapoint represents a biologically independent animal. For all bar graphs, data are mean ± s.e.m. (a–e, g and j). Hi-C data were combined from three independent biological replicates (f, h and i). P values were calculated using paired two-tailed t-tests (h), two-sided Kolmogorov–Smirnov tests (i), one-way ANOVA (c–e), two-way ANOVA (a, b and j) and unpaired Student’s two-tailed t-tests (g).Source dataWe next examined whether chRNA formation is facilitated by spatial proximity between parent genes. Genome-wide in situ Hi-C analysis revealed that interchromosomal parent–gene pairs identified by direct RNA-seq (Extended Data Fig. 1g) were significantly enriched for DNA–DNA interactions relative to randomized exon pairs (Fig. 2f). As CTCF regulates genome organization31, we depleted Ctcf using small interfering RNA (siRNA; Fig. 2g) and performed Hi-C after 6 h of LPS stimulation. LPS enhanced contacts between chRNA parent genes, whereas Ctcf depletion abolished this increase, demonstrating that these interactions are dynamically regulated and CTCF dependent (Fig. 2h,i).Functionally, Ctcf depletion reduced LPS-induced Gsdmd-Tmem106a without affecting Gsdmd or Tmem106a expression (Fig. 2j), and similarly reduced Cd274-Lacc1 without altering either parent gene (Extended Data Fig. 5f). Finally, junction-specific primers (Fig. 2k) detected only the Gsdmd exon 2 and Tmem106a exon 6 junction isoform (Fig. 2l), demonstrating that fusion is highly selective rather than a random consequence of proximity. Together, these findings show that inflammation promotes CTCF-dependent interactions between chRNA parent genes, enabling specific trans-splicing and chRNA formation in macrophages.