Mutant TP53 hijacks RNA-splicing factor RBM28 to suppress double-stranded RNA triggered antitumor immunity

Abstract

TP53 mutation may not only compromise its multifaceted tumor-suppressive functions but also confer oncogenic properties. Here, we demonstrate that DNA-binding domain mutations of TP53 unexpectedly confer a transcriptional regulatory function, directly driving RNA-splicing factor RBM28 overexpression. Overexpressed RBM28 excessively splices transposon elements, inhibiting double-stranded RNA (dsRNA) formation, thereby suppressing dsRNA-triggered type I interferon (IFN) signaling and subsequent anti-tumor immunity. We demonstrate in mouse tumorigenesis models and human multi-stage esophageal cancer development that mutant p53 (mutp53)-driven aberrant RBM28/dsRNA/IFN axis plays a crucial role in cancer initiation, progression and resistance to immune checkpoint blockade (ICB) therapy through innate immune suppression. Pan-cancer analysis indicates that this mechanism underlies ICB resistance in most cancers. Pharmacological restoration of normal p53 conformation or targeted RNA-splicing inhibition enhances anti-tumor immunity and ICB efficacy. Collectively, our study has unveiled an important function of mutp53 in establishing immunosuppressive tumor microenvironment, which provides an actionable framework for intervention and therapy in TP53-mutated cancers.

Mutant p53 (mutp53) drives tumor immune evasion partly through restricting innate immune sensing triggered by transposable elements (TEs). Here, the authors find that mutations in the DNA-binding domain enable mutp53 to transcriptionally activate the RNA-splicing factor RBM28, leading to excessive splicing of TE transcripts and suppression of antitumor immunity in esophageal cancer.

Introduction

The tumor suppressor p53 plays a critical role in many aspects of cellular functions and, thus, its mutations are prevalent in various types of human cancer^1,2. It is well known that mutant p53 (mutp53) often loses its normal tumor-suppressor functions and acquires other oncogenic properties resulting in the disruption of cell cycle arrest, DNA repair and apoptosis in response to cellular stress, which are performed by wild-type p53 (wtp53) function. Therefore, cells with the inactivating TP53 mutations show high genomic instability, characterized by high tumor mutational burden and widespread copy number alterations^3,4. Mutp53 can inhibit the wtp53 activity by dominant-negative effect and interact with other proteins to promote the malignant phenotypes of cancer cells such as proliferation, invasion and metastasis^5,6. Also, mutp53 plays a critical role in tumor immune evasion by suppressing anti-tumor immune responses⁷ and altering the tumor microenvironment (TME)⁸. For instance, mutp53 has been reported to interact with transcription factors like NF-kB and STAT, driving the expression of genes that promote immune suppression⁹. Mutp53 can also promote the secretion of certain interleukins to reprogram tumor-associated macrophages in the TME into pro-tumor macrophages by altering their lipid metabolism¹⁰. These previous findings indicate that mutp53 may affect cancer initiation and progression in multiple aspects. However, the non-canonic oncogenic properties of mutational events in TP53 remain largely unexplored.

Cellular double-stranded RNA (dsRNA) plays a crucial role in anti-tumor immunity. dsRNAs are generated from genomic transposable elements (TEs), a diverse group including long interspersed elements (LINEs), short interspersed elements (SINEs), and endogenous retroviruses (ERVs)¹¹. These TEs trigger innate immune sensing not only through the dsRNA they produce—which is recognized by intracellular sensors such as RIG-I-like receptors (RIG-I and MDA5) to initiate type I interferon (IFN) production¹²—but also through other nucleic acid components. For instance, cytoplasmic LINE-1 DNA can be sensed by cGAS, while RNA-DNA hybrids may also provoke an immune response¹¹. Additionally, other cancer-associated stresses, such as DNA damage and the formation of micronuclei, are potent activators of the cGAS/STING signaling pathway, further contributing to IFN-driven anti-tumor immunity^7,13. This immune response not only helps in combating viral infections but also enhances the ability of the immune system to detect and eliminate cancer cells¹¹. The presence of dsRNA can trigger immunogenic cell death in cancer cells, making them more visible to the immune system. Furthermore, dsRNA can activate dendritic cells, promoting a robust anti-tumor T-cell response¹⁴. Thus, endogenous dsRNA serves as a critical molecular signal that bridges innate and adaptive immunity, contributing significantly to the defense against cancer. It has been documented that RNA sensing induced by chromosome missegregation can augment anti-tumor immunity¹⁵. However, tumor cells can maintain dsRNA homeostasis through diverse mechanisms, including promoting dsRNA degradation, preventing dsRNA recognition and suppressing dsRNA formation, to evade immune activation. For instance, mutations in KRAS, a gene often mutated in certain human cancers, can accelerate dsRNA degradation, thereby impairing dsRNA sensing and IFN response that consequently drives immunosuppression in colorectal cancer¹⁶. For p53, it has been shown that in the context of pancreatic cancer, mutp53 can suppress Alu-derived dsRNA formation by controlling L1 proteins to bind Alus¹⁷, and in ovarian cancer, p53 loss of function reduces the level of H3K9me3 at the promoter of TEs, which elevates TE transcription, but decreases the levels of H3K27Ac at the promoter of interferon-stimulated genes (ISGs), ultimately suppressing the expression of ISGs¹⁸. Since dsRNA-triggered innate immunity plays a key role against carcinogenesis and antitumor immunotherapy, elucidating the mechanisms for suppressing this antitumor immunity is of fundamental significance for intervention and treatment of cancers.

Esophageal squamous-cell carcinoma (ESCC) develops through a well-defined multistage tumorigenesis process, including precancerous lesions such as low-grade intraepithelial neoplasia (LGIN) and high-grade intraepithelial neoplasia (HGIN)^4,19–21. Several whole-genome sequencing studies have shown a high occurrence of TP53 mutations in ESCC^4,22–24. Recently, we have further demonstrated that TP53 multimutations or biallelic loss emerged as early as in LGIN stage and its rates significantly increased in HGIN and ESCC, accompanied by widespread copy number alterations⁴. Furthermore, ESCC often presents markedly immunosuppressive TME during its initiation and progression^19,21,25,26 and resistance to immune checkpoint blockade (ICB) therapy^27,28. On the other hand, proteomics studies have reported an aggressive and short survival time ESCC subtype marked by upregulated spliceosomal and ribosomal proteins and elevated spliceosome pathway activity^29–31, suggesting that this may involve the aberrant RNA splicing. These previous findings indicate that ESCC may be a good model for such research.

In this work, we have explored the relationship between aberrant dsRNA splicing events and anti-tumor immunity in mouse and human ESCC. We have found that the dsRNA levels are significantly reduced during the multi-stage ESCC tumorigenesis, which is correlated with cancer initiation, progression and immunosuppressive TME. The dsRNA level decline in ESCC is caused by mutp53, which gains a non-canonic transcriptional regulatory function, directly driving RNA-splicing factor RBM28 expression. Overexpressed RBM28 promotes excessive splicing of TEs, inhibiting cellular dsRNA accumulation, suppressing anti-tumor immunity via attenuating the IFN signaling. We have also found that ESCCs with RBM28 overexpression were resistant to ICB therapy and had shorter survival time, which are also validated in multiple other types of human cancer. Finally, we have shown that pharmacological restoration of normal p53 conformation by arsenic trioxide or targeted inhibiting RNA splicing activity by TG003 enhance antitumor immunity and improve the ICB therapy efficacy. Collectively, our study has not only discovered a non-canonic mutp53 function but unveiled a critical role of mutp53-driven aberrant RBM28/dsRNA/IFN axis in creating immunosuppressive TME. These findings may have important implications in the intervention and treatment of cancer with TP53 mutations.

Results

Aberrant RNA splicing plays a vital role in the development of ESCC

To profile the proteomic alterations specifically within esophageal epithelial cells during 4-nitroquinoline 1-oxide (4-NQO)-induced esophageal tumorigenesis, we established a genetically engineered ED-L2-Cre, APEX2^F/F mouse model, in which the ED-L2 promoter activates expression of Cre in the esophageal epithelia, leading to specific expression of the proximity labeling enzyme APEX2³² in epithelial cells (Fig. 1A). The 4-NQO model recapitulates the multi-stage pathogenesis of human ESCC³³ and APEX2 technology allows for the precise labeling of proteins in living cells, providing unparalleled spatial resolution over conventional proteomic approaches. We treated the APEX2-expressed mouse esophageal tissue sections with biotin-phenol and performed immunofluorescence staining of epithelial marker KRT6A, APEX2, and the biotinylated proteins. We found that EGFR labeled by APEX2 was colocalized with epithelial cell marker KRT6A, indicating that APEX2 was successfully and specifically expressed in epithelial cells. Upon the addition of biotin-phenol labeling, APEX2⁺Cre⁺ tissues showed broad biotinylating patterns (Fig. S1A) and the biotinylated proteins were well colocalized with KRT6A (Fig. 1B), demonstrating that this method can effectively label proteins in mouse esophageal epithelium. We then examined the differences among the biotinylated proteins by the streptavidin magnetic bead method in normal mouse esophageal tissues (NOR) and precancerous or cancerous esophageal tissues obtained at various times from 4-NQO-treated mice (Fig. S1B). Immunofluorescence staining of biotinylated tissue sections (Fig. 1C) and Western blot analysis of tissue lysate input, flow-through and bead eluate (Fig. S1C and S1D) all validated the success of our enrichment method. We performed a Data Independent Acquisition quantitative proteomics analysis on the streptavidin bead pulldown samples after on-bead tryptic digestion and identified a total of 5263 proteins, among them 4949 in NOR (n = 4), 4880 in LGIN (n = 4), 4857 in HGIN (n = 4) and 4845 in ESCC (n = 5). Of the 5263 proteins, 4682 occurred in all esophageal tissue samples while the rest 581 were identified either in NOR, LGIN, HGIN or ESCC, respectively (Fig. S1E, S1F and Supplementary Data 1). Principal component analysis showed that NOR and LGIN tissues had a similar protein pattern, but HGIN and ESCC had distinct ones (Fig. S1G).

Fig. 1. Aberrant RNA splicing plays a vital role in the development of ESCC.

A Workflow for APEX2-based spatial proteomic analysis of esophageal epithelial cells in mouse ESCC formation induced by 4-NQO. Proteins in esophageal epithelial cells with APEX2 expression were specifically labeled with biotin, followed by streptavidin enrichment and mass spectrometry analysis (B) Fluorescence images showing APEX2 localization (EGFP as an indicator) and biotinylation activity (Neutravidin as an indicator). Scale bars, 100 μm. C Fluorescence images showing APEX2 biotinylation activity in normal esophageal (NOR), low-grade intraepithelial neoplasia (LGIN), high-grade intraepithelial neoplasia (HGIN) and esophageal squamous-cell carcinoma (ESCC) tissue samples. Scale bars, 100 μm. D Heatmap displaying the normalized levels of proteins and the top 5 associated gene ontology pathways across the 4 stages of mouse ESCC development. N, NOR; L, LGIN; H, HGIN; and E, ESCC. E Heatmap displaying normalized protein levels and the top 5 associated gene ontology pathways across 4 stages of human ESCC development. The proteomic data was obtained from a published study²⁴. F The number of 7 types of aberrant alternative splicing events (AASEs, red bars) and the corresponding number of genes (blue bars) affected by these AASEs. A3, alternative 3’ splice sites; A5, alternative 5’ splice sites; AF, alternative first exons; AL, alternative last exons; MX, mutually exclusive exons; RI, retained introns; and SE, skipping exons. Data are from bulk RNA-seq of human ESCC samples²⁰. G RNA splicing activity scores in NOR, LGIN, HGIN and ESCC samples from two patients. The left spatial plots show RNA splicing activity scores and the right box plot is the quantitative statistics. Box shows median (central line) and 25th to 75th percentile (box edge) with 1.5x quantile range. Trend test P-values in G are from two-sided Jonckheere-Terpstra test. Source data are provided as a Source Data file.

We hence performed protein co-expression and gene ontology analysis on the 5263 proteins to identify the representative ones and their biological roles in each histopathological stage of ESCC development. We observed that during the evolution from precancerous lesions to ESCC, there were 1114 proteins with their levels continuously increased, which were significantly enriched in the RNA splicing-related pathways, but there were 884 proteins that showed their levels continuous decreased during ESCC development, which were significantly enriched in the immune-related pathways (Fig. 1D). Similar results were seen when a previously published human esophageal cancer proteomic dataset²⁴ was analyzed using the same method, showing a significant elevation of proteins in RNA splicing-related pathways during ESCC development (Fig. 1E and Fig. S1H). We also analyzed our human bulk RNA-sequencing data on ESCC published previously²⁰ and the results showed that there were 7329 aberrant alternative splicing events that affected 1352 genes, encompassing 7 types of alternative splicing events (Fig. 1F). Furthermore, we performed a spatial transcriptomic analysis of human multi-stage esophageal lesion samples¹⁹ and confirmed that the RNA-splicing activity scores in epithelial cells were significantly increased from NOR, LGIN to HGIN and ESCC (Fig. 1G). These results indicated a parallel correlation between widespread aberrant RNA alternative splicing events and significantly elevated RNA-splicing proteins in precancerous and cancerous esophageal epithelial cells, suggesting that RNA splicing may play a key role in the development of ESCC.

TP53 mutations drive aberrant expression of RNA splicing factor RBM28 in ESCC

We next wanted to seek why the aberrant RNA splicing events in the precancerous and cancerous esophageal epithelial tissues were significantly more than that in normal esophageal tissues. By integrative analysis of proteomic and genomic data from human esophageal cancer samples, we found that among the frequent proposed ESCC driver mutations, only mutTP53 samples had significantly increased RNA splicing scores (Fig. 2A and Fig. S2A). Gene ontology analysis revealed that in the mutTP53 ESCC samples, 4 of the top 10 upregulated pathways were related to RNA splicing, and some DNA damage repair pathways (Fig. 2B). In ESCC, most TP53 mutations are missense, which often lead to a gain of function in mutp53^34,35. By stratification analysis of TP53 mutation types, we found that the RNA-splicing activity scores were significantly higher in the samples with TP53 missense mutations compared with that in the samples with TP53 non-missense mutations or wild type (Fig. 2C), suggesting that missense mutations may confer an additional function upon the p53 protein, which could potentially contribute to enhanced RNA splicing. These results were also the case at the transcriptomic level. When using their own matched adjacent normal tissues as controls, the mutTP53 samples had significantly more alternative aberrant splicing events than the wtTP53 samples (Fig. S2B). Furthermore, to identify mutation-associated aberrant alternative splicing events (AASEs), when compared directly with wtTP53 tumor samples, significantly more alternative aberrant splicing events and affected genes occurred in samples with TP53-missense mutations than those with TP53 non-missense mutations (Fig. S2C and S2D). These results emphasize that mutant p53 may drive aberrant RNA splicing in ESCC development.

Fig. 2. TP53 mutations drive aberrant expression of RNA splicing factor RBM28 in ESCC.

A The RNA splicing activity score is significantly higher in ESCC samples with mutant TP53 (MUT, n = 54) than with wild-type TP53 (WT, n = 36). B Dot plot displaying the top 10 enriched GO pathways of overexpressed proteins in TP53 mutations samples. C Differential RNA splicing activity score in ESCC samples with TP53 missense mutations (M, n = 43), non-missense mutations (non-M, n = 11) and WT (n = 36). D Venn diagram showing the selection of key splicing factors. E RBM28 RNA levels detected by spatial transcriptomic analysis. Left, spatial plots showing normalized epithelial RBM28 RNA levels; right, quantitative statistics. The horizontal lines in the violin plots represent mean values. TP53 + , wildtype TP53; TP53-, TP53 loss of heterozygosity. F Western blot analysis showing differential RBM28 protein expression in HET-1A, KYSE30, KYSE150 and KYSE450 cells with TP53-knockout (KO), mEC25 cells with Trp53 knockdown (KD), and KYSE510 cells with wtTP53(WT) or various mutTP53 overexpression. G The effects of mutTP53 or wtTP53 overexpression on RBM28 RNA levels in KYSE510 cells. H The ChIP-seq tracks show the occupancy of mutp53 in the promoter of RBM28 compared with wtp53. I The differential RBM28 DNA enrichment in KYSE510 cells with stably overexpressed of wtTP53 or various mutTP53 determined by ChIP-qPCR. J Electrophoretic mobility-shift assays demonstrating binding of all mutp53 proteins to RBM28 promoter except R337C. Arrow points to the DNA-protein binding bands. K The differential effects of wtTP53 or mutTP53 on luciferase reporter gene expression driven by the RBM28 promoter in KYSE510 cells. Box in A and C shows median (central line) and 25th to 75th percentile (box edge) with 1.5x quantile range whiskers. Data in G, I and K are mean ± SEM of 3 biological replications. Trend test P-values in E are from two-sided Jonckheere-Terpstra test. Two-sided Fisher’s exact test for B, two-sided Wilcoxon rank-sum test for A, C, and E, two-sided Student’s t-test for G, I and K. Source data are provided as a Source Data file.

We examined what are the critical factors contributing to the aberrant RNA splicing in mutTP53 ESCC. By the overlap analysis of mouse, human and TP53-missense mutation specific proteomic data, we found five RNA splicing related proteins, RBM28, TCERG1, PPM1G, PRPF4 and SNRPB2, upregulated in both human and mouse ESCC (Fig. 2D and Supplementary Data 2). Further confirmation assays with quantitative PCR and immunoblotting revealed that RBM28, PPM1G and TCERG1 were altered at both mRNA and protein levels in the TP53-null ESCC cell line KYSE510 upon stable expression of exogenous wild-type or mutant TP53 (Fig. S2E and S2F). Additionally, we found that the expression levels of Rbm28 and Ppm1G in mouse esophageal tissue samples were increased along with the disease lesion progression (Fig. S2G). Since the expression level change of RBM28 in tumor versus normal tissues and in missense TP53 mutation versus non-missense TP53 mutation was greater than PPM1G (Fig. S2H) and was the only splicing factor that was significantly associated with poor patient prognosis (Fig. S2I), we thus focused on it for further analysis. Spatial transcriptomics analysis showed a gradual increase in RBM28 expression as ESCC progressed, with higher expression in samples with TP53 loss of heterozygosity (LOH) (Fig. 2E). In the TCGA ESCC samples, RBM28 was also highly expressed in TP53 LOH samples (Fig. S2J). Immunofluorescence analysis of human multi-stage tissues further confirmed that RBM28 expression increased during ESCC development (Fig. S2K). We then conducted biochemical assays and revealed that TP53 knockout or Trp53 knockdown enhanced the RBM28 protein level in HET-1A cells (wild-type TP53) but reduced RBM28 protein level in KYSE30 (p.R110L/p.E326K), KYSE150 (R248Q), KYSE450 (H179R/E339*) and mEC25 (V170M) cells, all of which harbor TP53 DNA-binding domain (DBD) mutations (Fig. 2F). We also generated KYSE510 cell lines (TP53-null) stably expressing wild-type TP53, various DBD mutants (R175H, Y220C, R248W, R273H, and R280K), or the non-DBD mutant, R337C. Western blot analysis showed that expression of wild-type TP53 substantially reduced the RBM28 protein level, whereas DBD mutants significantly elevated the RBM28 protein level, but the non-DBD mutant R337C decreased the RBM28 level (Fig. 2F). Notably, knockout or knockdown of mutTP53 suppressed RBM28 expression while re-expression of DBD mutants sustained RBM28 levels above those in TP53-null (empty vector control) cells. These results suggest that the mutTP53 mutated at the DBD might have transcriptional activity for RBM28 expression. Truly, RT-qPCR analysis showed that the RBM28 RNA level can be significantly upregulated by DBD mutated TP53, which could be suppressed by inducing wtTP53, but non-DBD mutated TP53 (R337C) had no such function (Fig. 2G).

We previously found that during ESCC initiation and progression, TP53 mutations or LOH caused genomic copy number alterations⁴. However, by analysis of the DNA sequencing data, we did not find significant copy number alteration in the RBM28 locus at chromosome 7q32.1 (Fig. S2L), supporting that the TP53 inactive mutations directly activate RBM28 expression at transcription level. We thus analyzed the Cistrome DB data³⁶ and the results showed that the mutp53 (R280K and R273H) but not wtp53 could bind the RBM28 promoter region, and the binding was eliminated when the mutTP53 was silenced by short hairpin RNA (Fig. 2H). Similar results were obtained when analyzing a previously reported ChIP-seq dataset³⁷, showing that mutp53 (R248Q, R273H and R249S) but not wtp53 could bind to the RBM28 promoter (Fig. S2M). Notably, despite having some variation in binding signals, we identified a conserved C-rich cis-element in the promoter of RBM28 (Fig. S2N) similar to that in the promoters of KMT2A, KAT6A and CTNNB1 can be bound by mutp53^37,38. The qPCR analysis of chromatin immunoprecipitation product produced by the p53 antibody demonstrated a significant enrichment of the RBM28 promoter in KYSE510 cells with mutTP53 mutated at the DBD, but not that with mutTP53 mutated at non-DBD or with wtTP53 (Fig. 2I). This binding pattern mirrored that of the known mutp53 target KMT2D (Fig. S2O) and was opposite to the canonical wtp53 target CDKN1A (Fig. S2P). This binding was functionally consequential, as evidenced by a concurrent increase in the active histone mark H3K27Ac at the promoter specifically in DBD mutp53 cells (Fig. S2Q). To investigate whether this interaction is direct, we performed an electrophoretic mobility shift assay and the result confirmed that purified DBD mutant p53 protein, but not wild-type p53, could bind directly to the RBM28 promoter probe. Moreover, when the C-rich motif on the probe was mutated, this binding was abolished (Fig. 2J). Reporter gene assays further confirmed that the interaction between the DBD mutated TP53 and the RBM28 promoter are of biological function, showing that the plasmid construct with the C-rich motif-deleted RBM28 promoter had almost no reporter gene activity but the plasmid construct with the full-length of the RBM28 promoter had significant reporter gene activity when the plasmid constructs were expressed in KYSE510 cells transfected with various types of TP53 (Fig. 2K).

RBM28 overexpression inhibits transposable element-derived dsRNA formation

Our previous studies have shown that TP53 mutations or LOH resulted in many CNAs in the ESCC genome⁴. Transposable elements (TEs) constitute nearly half of the human genome and are the major source of cellular dsRNAs³⁹. By analyzing the DNA sequencing data⁴, we found that the copy number of TEs increased along with the development of ESCC (Fig. 3A and Fig. S3A). However, the dsRNA response score did not parallelly increase during the disease development in human and mice (Fig. 3B and Fig. S3B). We reasoned that this apparent paradox could be explained if the transcription of amplified TEs was suppressed. Through analyzing the bulk RNA-sequencing data of mouse multi-stage diseased esophageal samples²¹, we found that the expression of most TEs was significantly decreased in LGIN, HGIN, and ESCC compared with NOR (Fig. S3C). These downregulated TEs were predominantly from the LINE and LTR families (Fig. S3D). Based on this finding, we selected five classic TEs to check their expression levels in these samples by using a qPCR method and the results were confirmative (Fig. S3E). It is known that multiple genes are involved in maintaining dsRNA homeostasis. For instance, the RISC complex DICER1, AGO2 and TARBP1 and RISC antagonist DDX60¹⁶, RNase ERN1¹⁴, RNA transporter SIDT2⁴⁰, and RNA helicases SKIV2L and DHX9^41,42 are responsible for degrading dsRNA, while ADAR and FTSJ3 prevent the activation of dsRNA sensors by editing dsRNA^43,44. However, we found that the expression of these molecules did not increase during esophageal carcinogenesis (Fig. S3F), nor was it significantly elevated in tumors with TP53 missense mutations compared with tumors where TP53 was not mutated (Fig. S3G). We therefore proposed that it was dsRNA formation rather than degradation or recognition that caused this paradoxical phenomenon. Since aberrant splicing is a source of dsRNA formation^45,46, it is warranted to hypothesize that RBM28 functions by splicing TEs to inhibit their dsRNA formation.

Fig. 3. RBM28 overexpression inhibits transposable element-derived dsRNA production.

Significant differences of transposon element (TE) copy number (A) human NOR (n = 14), LGIN (n = 4), HGIN (n = 8) and ESCC (n = 14) samples and dsRNA response scores (B) across human ESCC stages (NOR, n = 15; LGIN, n = 6; HGIN, n = 8 and ESCC, 14). C The effect of RBM28 knockout (KO) on the dsRNA level in KYSE30 cells. Left, multiplex immunofluorescence (mIF) staining of RBM28 and dsRNA; right, quantitative statistics of mean immunofluorescence intensity (MFI). Scale bars, 20 μm. Data are mean ± SEM of dsRNA intensity from cells in non-overlapping visual fields from 3 biological independent experiments. D The effect of Rbm28 KO on the dsRNA level in mouse allograft derived from mEC25. Left, mIF staining images; right, quantitative statistics. Scale bars, 100 μm. Data are mean ± SEM from 5 mice. E The effect of mutTP53 on the intracellular dsRNA level in KYSE510 cells by flow cytometry. MutTP53 with RBM28 versus Vector shows dsRNA suppression, reversed by RBM28 KO. F Volcano plot showing that RBM28 KO in KYSE30 cells upregulated most TEs at locus level. The pie chart illustrates the proportion of upregulated TEs among various chromatin regions (G) and TE types (H). TTS, Transcription termination site; TSS, Transcription start site. I The effect of wtTP53/mutTP53 overexpression in KYSE510 cells on the TE RNA expression by qPCR. *P < 0.05, ** P < 0.01, *** P < 0.001 and **** P < 0.0001. The effects of RBM28/Rbm28 KO on the TE RNA expression in KYSE30 cells (J) and mEC25 cells (K) by qPCR. CLIP-qPCR analysis shows increased TEs RNA levels captured by RBM28 antibody in RBM28 KO-KYSE30 cells (L) and Rbm28 KO-mEC25 cells (M). RIP-qPCR analysis KYSE30 (N) and mEC25 (O) cells with RBM28/Rbm28 KO had higher retention of dsRNA derived from TEs with dsRNA-specific antibody J2. Data in E and I − O are mean ± SEM of 3 biological replications. Two-sided Jonckheere-Terpstra test for A and B, two-sided Wilcoxon rank-sum test for C and D, two-sided Student’s t-test for E and I-O, and empirical Bayes moderated t-test with Benjamini-Hochberg correction for F. Source data are provided as a Source Data file.

We found that in RBM28-knockout cells (Fig. S4A), the upregulated genes were significantly enriched in the pathway of response to dsRNAs (Fig. S4B). Immunofluorescence staining and flow cytometry analysis showed that RBM28-knockout KYSE30 and mEC25 cells had significantly increased levels of dsRNAs (Fig. 3C and Fig. S4C − S4E). Comparable results were obtained in C57BL/6 J mouse subcutaneous allografts derived from Rbm28-knockout mEC25 cells, and in ESCC induced by carcinogen 4-NQO in esophageal-specific Rbm28-knockdown mice by using adeno-associated virus (AAV)-shRbm28 system (Fig. S4F, S4G and Fig. 3D). Consistently, we found that knockout of mutTP53 in KYSE30 cells or knockdown of mutTrp53 in mEC25 cells also increased the dsRNA levels (Fig. S4H and S4I). We then transferred the distinct types of mutTP53 mutated at the DBD (R175H, Y220C, R248W, R273H, and R280K) into KYSE510 cells to check their effect on dsRNA formation (Fig. S4J). Flow cytometry analysis showed that all these mutTP53 inhibited dsRNA production; however silencing RBM28 expression in these cells could restore the dsRNA production (Fig. 3E and S4K).

TEs with increased copy numbers can be transcribed in a bidirectional manner or as inverted repeats, forming sense-antisense hybrids or foldback hairpin dsRNA¹¹. Since elevated RBM28 may splice out the amplified TEs preventing them from forming additional dsRNA, we examined the effects of RBM28 knockout on the RNA levels of TEs in ESCC cells by RNA-sequencing. The results showed that the RBM28 knockout significantly increased the RNA levels of most TEs compared with RBM28 non-knockout, both at the family level and at individual genomic loci (Fig. S4L and 3F). Most of these upregulated TEs were located within introns (Fig. 3G), with the SINE and LINE elements being the majority (Fig. 3H). We also identified specific introns containing these TEs that were robustly regulated by RBM28 (Fig. S4M). We then selected 8 TEs in human KYSE510 cells and KYSE30 cells and 5 TEs in mouse mEC25 cells for qPCR analysis. The results showed that the transfection of mutTP53 mutated at the DBD to KYSE510 cells significantly suppressed the expression of these TEs (Fig. 3I). Conversely, knockout of mutTP53 in KYSE30 cells or knockdown of mutTrp53 in mEC25 cells elevated TE expression (Fig. S4N and S4O). Furthermore, RBM28 knockout in KYSE30 and mEC25 cells significantly elevated the expression of these TEs compared with RBM28 non-knockout (Figs. 3J, K). As an RNA splicing factor, RBM28 may bind to TEs in the pre-mRNA sequences when it performs splicing activity. We thus conducted RNA immunoprecipitation (RIP)-qPCR assays and confirmed that RBM28 is indeed bound to these repetitive sequences in both human and mouse cell lines (Fig. S4P, S4Q). We further conducted cross‑linking immunoprecipitation qPCR (CLIP‑qPCR) analysis and the results demonstrated that RBM28 interacts directly with these TEs (Figs. 3L, M). Importantly, by RIP-qPCR assays using the specific antibody against dsRNA J2, we found that RBM28 knockout in human and mouse ESCC cells significantly retained more dsRNAs derived from the TEs compared with controls (Figs. 3N, O). These results suggest that mutTP53-driven RBM28 overexpression inhibits endogenous dsRNA formation via excessively splicing out TEs.

RBM28 overexpression shapes an immunosuppressive microenvironment

We next systematically examined the effects of RBM28 on the progression of ESCC. Interestingly, we found that RBM28 knockout did not significantly alter the proliferation of mouse and human ESCC cell lines in vitro (Fig. 4A, Fig. S5A and Fig. S5B) and in vivo in immunodeficient NSG mice and Rag2^-/- mice (Fig. 4B, Fig.4C, and Fig. S5C–E); however, the Rbm28 knockout significantly inhibited mEC25 cell proliferation in immunocompetent C57BL/6 J mice (Fig. 4D and Fig. S5F), suggesting that the pro-tumor function of RBM28 may be immune dependent. In the esophageal-specific Rbm28-knockdown mouse ESCC model induced by 4-NQO, mice with Rbm28 knockdown in the esophagus had significantly reduced ESCC number and tumor size compared with mice without Rbm28 knockdown (Fig. 4E). Similarly, knockdown of Trp53 significantly inhibited the proliferation of mEC25 cells in immunocompetent C57BL/6 J mice (Fig. S5G). These results encouraged us to investigate the effect of aberrant RBM28 expression on the tumor immune microenvironment. We first performed deconvolution analysis on published bulk RNA-sequencing data²⁰ and proteomic data²⁴ of human ESCC. Curating individual immune cell gene signatures by single-sample gene set enrichment analysis and tumor immune dysfunction and exclusion algorithm, we assigned immune cell infiltration scores to these datasets. The heatmap showed that samples with a high RBM28 level had higher exclusion scores, reflected by a decrease in tumor-infiltrating lymphocytes such as CD8⁺ T cells, CD4⁺ T cells and other immune cells but an increase in myeloid-derived suppressor cells (MDSCs) (Fig. S5H), suggesting that RBM28 overexpression driven by mutTP53 may suppress immune cell infiltration. Second, we analyzed the immune cell infiltration in our published spatial transcriptomics data of multi-stage esophageal lesion samples¹⁹ by dividing the epithelial spots into RBM28 high or low status (Fig. 4F). We defined a cell neighborhood score by calculating the average cell-type score surrounding the corresponding spots. As shown in the heatmap, spots with mutTP53 exhibited an immunosuppressive phenotype and spots with high RBM28 level showed reduced infiltration of CD8⁺ T cells but increased infiltration of MDSCs (Figs. 4G, H and S5I). Third, we also analyzed the immune cell infiltration using our previously published single-cell RNA sequencing data of multi-stage esophageal lesion samples¹⁹ and ESCC samples²⁵ and found that compared with low RBM28 level samples, high RBM28 RNA level samples had significantly less CD8⁺ T cell infiltration, particularly the effector CD8⁺ T cells (Figs. 4I, J). Finally, we performed multiplex immunofluorescence (mIF) staining analysis of 4-NQO induced mouse ESCC tissues and the results showed that Rbm28 silenced ESCC had significantly higher infiltration of CD8⁺ T cells especially GZMB⁺ CD8⁺ T cells but lower infiltration of MDSCs in the TME (Figs. 4K, L). Similarly, flow cytometry and multiplex immunofluorescence staining analysis also revealed that there were significantly more CD8⁺ T cells (especially GZMB⁺ CD8⁺ T cells) but significantly fewer MDSCs in C57BL/6 J mouse allografts derived from Rbm28-knockout mEC25 cells than in allografts derived from mEC25 cells without Rbm28-knockout (Fig. 4M − 4 P, and S5J). Further FCM analysis indicated that the decrease in MDSCs was primarily due to a reduction in the PMN-MDSCs, with no significant change in the M-MDSCs (Fig. S5K). These results clearly indicate that during the development of ESCC, mutTP53-driven RBM28 overexpression promotes an immunosuppressive phenotype characterized by decreased CD8⁺ T cell infiltration and increased MDSC accumulation.

Fig. 4. RBM28 overexpression drives an immunosuppressive tumor microenvironment.

A The in vitro proliferation of Rbm28-knockout (KO) mEC25 cells. The growth of Rbm28-KO mEC25 cells in NSG (B, n = 4), Rag2^-/- (C, n = 5) and C57BL/6 J mice (D, n = 5). (E) The effect of Rbm28 knockdown by shRbm28-AAV system on 4-nitroquinoline 1-oxide (4NQO)-induced ESCC formation in C57BL/6 J mice (n = 8). Left, gross ESCC image; middle and right, quantitative statistics of tumor numbers and volume. F Spatial transcriptomic analysis of human ESCC and precancerous tissues and the tumor neighborhood score calculation formula. High/low groups were defined by RBM28 expression presence/absence; the neighborhood score of a cell type was the average score of 7 spots adjacent to the target tumor spot. Mut, mutant; wt, wildtype. G Heatmap displaying the tumor neighborhood scores of various immune cells in ESCC samples with wtTP53 or mutTP53 (upper) and with RBM28 high or low (lower). TIL, tumor-infiltrating lymphocytes; B, B cells; Macro, macrophages; M2, M2 macrophages; Neutro, neutrophils; MDSC, myeloid-derived suppressor cells. H Spatial transcriptomic showing that MDSCs were adjacent to high RBM28 ESCC cells but CD8⁺ T cells were far. Summary of immune cell analysis on single-cell RNA sequencing in 43 human multi-stage esophageal lesion samples¹⁹(I) and 60 human ESCC samples²⁵ (J) Group was defined by the median value. K, LThe effect of Rbm28 knockdown by shRbm28-AAV system on mouse ESCC microenvironment in C57BL/6 J mice. Left, immunofluorescence staining images; right, quantitative statistics from 8 tumors. The effect of Rbm28 KO in mEC25 on the numbers of CD8⁺ T cell (M and N) and MDSC (O and P) infiltration in C57BL/6 J mice detected by flow cytometry (M and O) and immunofluorescence (N and P). Left, flow cytometry or immunofluorescence staining images; right, quantitative statistics from 5 tumors. *P < 0.05, *P < 0.01 and ***P < 0.001. Data in A–E and K–P are mean ± SEM. Two-sided Wilcoxon rank-sum test for all P value in this figure. Source data are provided as a Source Data file.

RBM28 overexpression prevents the IFN pathway activation

It is known that dsRNAs play a vital role in regulating innate immunity, primarily through the activation of IFN signaling pathways. Since the aberrant RBM28 expression suppressed the production of cellular dsRNAs, we thus investigated whether the levels of dsRNA sensors (RIG-I and MDA5) and the upstream molecules involved in the activation of the IFN genes (MAVS, TBK1, IRF3 and their phosphorylated forms) are altered in human and mouse ESCC cells. Western blot assays showed that these protein levels significantly increased in human or mouse cells with mutTP53 knockout or mutTrp53 knockdown but significantly decreased in the cells overexpressing mutTP53 (Fig. 5A). We also observed that in RBM28 knockout cells, the levels of all these proteins increased (Fig. S6A), suggesting that mutTP53-induced aberrant expression of RBM28 may indeed affect the activation of IFN pathway. We then performed siRNA-mediated knockdown of the RNA sensors IFIH1 (MDA5) and DDX58 (RIG‑I) and found that RBM28 knockout significantly reduced the levels of proteins downstream of IFN pathway, confirming that RBM28 suppresses the IFN response specifically by limiting immunostimulatory dsRNA (Fig. 5B). We next wanted to seek the molecular mechanism underlying aberrant RBM28 expression related to immunosuppressive tumor environment indicated by increased MDSCs infiltration but decreased CD8⁺ T cell infiltration. By analyzing the RNA-seq data of RBM28-knockout/non-knockout-KYSE30 cells, we found many differentially expressed genes and among them were CXCL10, CCL5, ISG15 and IFNB that significantly upregulated in RBM28-knockout cells (Fig. 5C). Pathway enrichment analysis of these differentially expressed genes showed that the upregulated genes were enriched in immune-related pathways, particularly IFN-associated pathways, while the downregulated genes were significantly enriched in RNA splicing pathways (Fig. 5D and Fig. S6B). We then looked at the data from mouse and human ESCC samples. The scRNA-seq data²¹ and APEX2 proteomics data of mouse esophageal samples showed that the scores of IFN-related pathways gradually decreased with the development of ESCC, and in human ESCC samples²⁴, the lowest IFN-related pathway score was presented in those having TP53 missense mutations (Fig. S6C). Correlation analysis revealed that RBM28 was significantly and negatively correlated with type I IFN but not type II IFN in both human and mouse samples (Fig. 5E and Fig. S6D). We also looked at our previously published mouse multi-stage esophageal lesion sample scRNA-seq data²¹ and found 12 representative and classic ISGs, including CXCL10, CCL5 and ISG15, decreased with the development of ESCC (Fig. S6E). The qPCR results for selected ISGs were largely consistent with the scRNA-seq data, with 10 out of 12 genes showing concordant changes (Fig. 5F).

Fig. 5. Mutant TP53-driven RBM28 overexpression inhibits IFN pathway activation.

A The effect of TP53 KO, Trp53 knockdown (KD) and wtTP53 or various mutTP53 overexpression (OE) on the levels of the dsRNA sensors RIG-I and MDA5, and downstream molecules MAVS, TBK1, phosphorylated-TBK1 (pTBK1), IRF3 and phosphorylated-IRF3 (pIRF3) in KYSE30, mEC25 and KYSE510 cells. B The effects of IFIH1 and DDX58 knockdown in RBM28/Rbm28 knockout (KO) on the levels of dsRNA sensors, and downstream proteins in KYSE30 and mEC25 cells. GAPDH served as loading control. C Volcano plot showing differentially expressed genes in RBM28-KO KYSE30 cells. Labeled are the top 10 upregulated genes (red) and top 10 interferon-stimulated genes (ISGs, black). D Gene set enrichment analysis of upregulated genes in RBM28-KO KYSE30 cells showing that the top 10 enriched REACTOME pathways belonged to immune-related pathways. E Spearman analysis showing negative correlations of RBM28 protein levels and type I IFN levels. Left, the correlation in mice multi-stage esophageal lesion samples; right, the correlation in human ESCC samples with different TP53 mutation status. Shaded region indicates 95% confidence interval for the correlation. F Heatmap shows downregulated ISGs during mouse ESCC progression. The effect of RBM28 (G) or mutTP53 KO (H) on ISGs expressions in KYSE30 cells. I The effects of ectopically expressed wild-type or various mutTP53 on the expression of ISGs in TP53-null KYSE510 cells. JThe effect of RBM28 KO on the expression of ISGs in mutant-TP53-overexpressed KYSE510 cells by qPCR. K RBM28 KO-KYSE30 cells produced significantly higher IFNβ level in medium by ELISA. L The effect of RBM28 level on CD8⁺ T cell IFNβ response in 60 human ESCC samples²⁵. Left, UMAP plot; right, quantitative statistics. Group for high or low RBM28 expression was by the median value. The horizontal lines in the violin plots represent mean values. Data in G − K are mean ± SEM from 3 biological replicates. P-values were determined by the empirical Bayes moderated t‑statistic with Benjamini‑Hochberg method (C), two-sided Jonckheere-Terpstra test (F), two-sided Student’s t-test (G − K), and two-sided Wilcoxon rank-sum test (L). Source data are provided as a Source Data file.

Next, we examined the effects of mutTP53 and RBM28 on the expression of CXCL10, CCL5, ISG15 and IFNB in both human and mouse cell lines. Real-time qPCR analysis showed that RBM28 knockout in KYSE30, mEC25 and Het1A cells, or mutTP53/mutTrp53 knockout or knockdown in KYSE30 and mEC25 cells significantly increased the expression of ISGs (Figs. 5G, 5H and S6F-S6H). In contrast, knockout of wild‑type TP53 in Het1A cells (Fig. S6I) or ectopic expression of DBD‑mutant TP53 in KYSE510 cells significantly suppressed the expression of these ISGs (Fig. 5I). Importantly, RBM28 knockout in KYSE510 cells reversed mutTP53 (R280K)-mediated decrease in the transcription levels of ISGs (Fig. 5J and S6J). Consistently, in RBM28‑knockout cells, subsequent knockdown of the cytoplasmic RNA sensors (IFIH1/DDX58) abolished ISG expressions induced by RBM28 loss (Fig. S6K and S6L), confirming that RBM28 acts in the upstream of dsRNA sensing to regulate the IFN pathway. We measured IFNβ, CCL5, CXCL10 and ISG15 levels in KYSE30 cell culture medium by enzyme-linked immunosorbent assays and the results demonstrated that RBM28-knockout cells secreted significantly higher ISGs than control cells (Fig. 5K and S6M). We also compared the functional state of CD8⁺ T cells between RBM28-high and RBM28-low ESCC samples based on the scRNA-seq data in our previously published study²⁵. We examined 1000 CD8⁺ T cells in each sample and found that CD8⁺ T cells in RBM28-high samples had significantly lower responsiveness to IFNβ and lower levels of cytotoxic or pro-apoptotic molecules such as granzyme family members (GZMA, GZMB, GZMH and GZMK), TNF, TNFSF10, IFNG, FASLG encoding FAS ligand and PRF1 encoding Perforin than that in RBM28-low ESCC samples (Fig. 5L and Fig. S6N). These results indicate that RBM28 overexpression inhibits the IFN signaling in ESCC, resulting in an immunosuppressive TME.

RBM28 overexpression in tumors is correlated with resistance to immunotherapy in patients

We then examined the relationship between the RBM28 or dsRNA status in tumor tissues and the therapy efficacy of neoadjuvant PD-1 antibody and chemotherapeutic agent treatment in ESCC patients. We performed mIF staining of tissue sections and the results revealed that the relative RBM28 IF levels were significantly higher in non-responders (N = 50) than in responders (N = 50) while the dsRNA IF levels were the opposite. Moreover, the levels of RBM28 showed a significant and negative correlation with the dsRNA levels (Fig. 6A), and non-responders exhibited significantly higher RBM28 expression than responders (Fig. S7A). We further analyzed the publicly available one bulk RNA-seq dataset (N = 30) from clinical trials of single-agent PD1 antibody treatment and two scRNA-seq datasets (N = 7 and N = 12, respectively) from clinical trials of PD1 antibody plus chemotherapy for ESCC^28,47,48. The results showed that patients with low tumoral RBM28 RNA levels responded better to the therapy than those with high tumoral RBM28 RNA levels (Fig. 6B–D and Fig. S7B), and responders exhibited higher dsRNA scores in epithelial cells (Fig. S7C, D). These results indicate that the high RBM28 activity caused low dsRNA content may be a factor contributing to ESCC immunotherapy resistance. We also examined the correlation between RBM28 levels and overall survival time in two independent ESCC patient cohorts^25,49 (Figs. 6E, F) and their pooled sample (Fig. S7E). The results consistently showed that patients with high tumoral RBM28 RNA levels had significantly shorter overall survival time, probably reflecting the detrimental role of mutTP53-driven aberrant RBM28/dsRNA/IFN pathway in anti-tumor immunity determining the prognosis of cancer.

Fig. 6. RBM28 overexpression in tumors is correlated with resistance to immunotherapy in patients.

A The effect of RBM28 and dsRNA levels in ESCC on the efficacy of neoadjuvant ICB therapy plus chemotherapy in our own cohort (N = 100). Shown are hematoxylin and eosin (H&E) staining and multiplex immunofluorescence staining of RBM28 and dsRNA (upper) and quantitative statistics (bottom). Non-responders had significantly higher RBM28 but lower dsRNA levels in tumors than responders, which were significantly and negatively correlated. Scale bar, 100 μm. Shaded region indicates 95% confidence interval for the correlation. B − D The effect of RBM28 on ICB therapy efficacy was validated in 3 publicly available ESCC patient sets. Data in B was from bulk RNA-sequencing⁴⁷, and in C and D were from single-cell RNA-sequencing^28,48. E, F Kaplan-Meier survival estimate shows ESCC patients^25,49 with high RBM28 RNA level in tumor had significantly shorter survival probability than those with low RBM28 RNA level in tumor. P value from log-rank test. HR, hazard ratio of death for high RBM28 RNA level (95% confidence interval). G Comparison of the RBM28 RNA levels between TP53-mutated and TP53-non-mutated tumors across TCGA multiple cancer types. H RBM28 RNA levels were significantly and positively correlated with the infiltrated numbers of MDSCs (left) but significantly and negatively correlated with T cell exclusion scores (right) in 15 of 16 cancer types with the RBM28 expression significantly differed between wtTP53 and mutTP53 shown in G. I, J The effect of the RBM28 expression level in tumor on ICB therapeutic efficiency were validated in multiple other cancer types with the data being publicly available. Box shows median (central line) and 25th–75th percentile (box edge) with 1.5x quantile range represented by whiskers. The horizontal lines in the violin plots represent mean values. All P values in this figure were from two-sided Wilcoxon rank-sum test except specifically indicated. Correlation in A and H were by spearman analysis. Source data are provided as a Source Data file.

Since TP53 mutations are prevalent in human cancer, we thus speculated that mutTP53 promotes RBM28 expression and drives immune evasion via IFN pathway suppression may be a general phenomenon in TP53-mutated cancer types. By analysis of the Cancer Cell Line Encyclopedia, we revealed that cancer cell lines with TP53 mutations usually had higher RBM28 RNA levels compared with cell lines derived from the same type of cancer without TP53 mutation (Fig. S7F). Analysis of the Cancer Genome Atlas data across 23 cancer types revealed significantly higher RBM28 RNA levels in mutTP53 versus wtTP53 tumors in pooled analysis (Fig. S7G), with 16 individual cancer types showing significant elevation (Fig. 6G). Notably, the magnitude of RBM28 changes between mutTP53 tumors and wtTP53 tumors exceeded that of established targets, including the wild-type p53-regulated gene CDKN1A and the mutant p53-regulated genes KMT2A, KMT2D, and CTNNB1 (Fig. S7H). Moreover, we found that in these 16 cancer types, RBM28 RNA levels in tumors were significantly and positively correlated with the infiltrated numbers of MDSCs and the exclusion scores of CD8⁺ T cells, with both correlation coefficient (r) >0.3 in HNSC, LIHC, LUAD, LUSC, SARC, and SKCM (Fig. 6H), suggesting that high tumor RBM28 expression level also causes immunosuppressive TME in TP53-mutated cancer types beyond ESCC. Consistent with these transcriptomic findings, analysis of the Clinical Proteomic Tumor Analysis Consortium dataset also confirmed that RBM28 protein levels were significantly elevated in mutTP53 tumors and the magnitude of RBM28 changes also exceeded CDKN1A in a pooled analysis (Fig. S7I), with significant upregulation observed in four individual cancer types (Fig. S7J). We also observed that high RBM28 expression is associated with shorter overall survival in the majority of cancer types from TCGA (Fig. S7K). We further searched for the publicly available data on ICB therapy of other non-ESCC cancer types to analyze the relationship between RBM28 expression level in tumors and the ICB therapeutic efficacy, and found that RBM28 RNA levels (bulk RNA-seq or scRNA-seq) or protein levels, measured before ICB treatment, were significantly higher in ICB non-responders than in ICB responders across almost all cancer types including melanoma^50–53, non-small cell lung cancer^54–56, head and neck squamous cell carcinoma⁵⁷, gastric carcinoma^58,59, colorectal cancer⁶⁰, and breast cancer⁶¹ (Figs. 6I, J). Despite limited sample size in certain cohorts, the consistent pan-cancer association supports RBM28 as a biomarker for ICB resistance. These results strongly indicate that the aberrant RBM28-dsRNA-IFN pathway plays a significant role in cancer immunity, thereby affecting cancer progression and ICB efficacy in pan-cancers.

Restoring p53 normal function in cancer cells enhances anti-cancer immunity

Given the presence of extensive RNA splicing in ESCC, it is significant to examine whether inhibiting the aberrant RNA splicing activity may prevent ESCC development. Based on the finding that mutp53 can directly regulate RBM28 transcription, we further explored the potential therapeutic role of arsenic trioxide (ATO) in ESCC since it has been shown of broad applicability in reactivating druggable p53 mutants^62,63. Surprisingly, ATO treatment significantly suppressed the mRNA expression of RBM28 in TP53-mutated human and mouse ESCC cell lines (KYSE30 and mEC25) whereas it had no such effect on RBM28 in ESCC cell lines with wtTP53 (HET-1A) or TP53 deletion (KYSE510) in vitro (Fig. 7A and Fig. S8A). Also, ATO treatment elevated the levels of dsRNA sensors and downstream molecules that participate in the activation of the IFN pathways in TP53 mutated ESCC cells but not in HET-1A cells or KYSE510 cells (Fig. 7B and Fig. S8B). Furthermore, immunofluorescence staining and flow cytometry analysis demonstrated that ATO treatment significantly elevated cellular dsRNA contents in mEC25 cells and KYSE30 cells, which could be reversed by ectopically overexpressed RBM28 (Fig. 7C, Fig. S8C and S8D) while ATO had no such effect on dsRNA in HET-1A and KYSE510 cells (Fig. S8E). We also examined the effects of ATO on the expression of effector genes, i.e., IFNB and some ISGs in KYSE30 and mEC25 cells and found that ATO treatment markedly suppressed RBM28 RNA level but significantly elevated the RNA levels of IFNB, CCL5, CXCL10 and ISG15; however, the restoration of RBM28 expression significantly attenuated the effect of ATO (Fig. 7D and Fig. S8F). These in vitro results prompted us to conduct experiments in mice. We subcutaneously implanted mEC25 cells into mice and 5 days later, ATO was administered (5 mg/kg, i.p., daily) for 12 days (Fig. S8G). The results showed that ATO treatment significantly inhibited allograft growth, achieving an effect comparable to RBM28 knockout (Fig. 7E). Nevertheless, ATO treatment were less effective for the allografts derived from RBM28-knockout mEC25 cells than the allografts derived from RBM28-non-knockout cells (inhibitory rate of ATO for RBM28-knockout allografts 29.61% versus ATO for RBM28-non-knockout allografts 44.89%), indicating that in this experimental conditions, ATO exerts its antitumor effect through RBM28 reduced sensitivity to ATO in the absence of RBM28 (Fig. 7E). During the experiment course, ATO treatment did not significantly change mouse body weight gain compared with non-ATO treatment (Fig. S8H). mIF staining of the allografts showed that although either ATO treatment or RBM28 knockout all resulted in increased levels of dsRNAs, a higher number of GZMB⁺ CD8⁺ T cells and reduced infiltration of MDSCs, combination of RBM28 knockout and ATO treatment had more significantly pronounced effect than the sole treatment (Fig. 7F).

Fig. 7. Restoring p53 tumor suppressor function in cancer cells enhances anti-cancer immunity.

Arsenic trioxide (ATO) treatment significantly inhibited the RBM28 (Rbm28) expression and elevated CDKN1A (Cdkn1a) at RNA level (A), while suppressed RBM28 and elevated dsRNA sensors and downstream molecules at protein level (B) in KYSE30 cells and mEC25 cells. ATO significantly suppressed RBM28 level but elevated dsRNA mean immunofluorescence intensity (MFI) (C), and the expression of interferon-stimulated genes (D), which reversed by RBM28 overexpression in KYSE30 cells. Nuclei was stained with DAPI. Scale bars, 20 μm. Data in C are mean ± SEM from cells in non-overlapping visual fields from 3 biologically independent experiments. E ATO treatment significantly promoted the effect of RBM28 knockout (KO) on suppressing mEC25 allografts growth in C57BL/6 J mice. Left, tumors at the endpoint; right, allografts growth curves. F ATO treatment significantly promoted the effects of RBM28 KO on increasing dsRNA accumulation in ESCC cells and on increasing GZMB⁺CD8⁺ T cell numbers but decreasing MDSC numbers. Scale bar, 100 μm. Data in E and F are mean ± SEM from 8 allograft tumors. G TG003 treatment significantly inhibited 4-NQO-induced mouse primary tumor numbers and volumes. H TG003 treatment significantly reduced the incidence of ESCC and precancerous lesions. Left, hematoxylin and eosin images; right, proportions statistics. I TG003 treatment significantly increased dsRNA accumulation in ESCC cells, and increased GZMB⁺CD8⁺ T cell numbers but decreased MDSC numbers. Scale bar, 100 μm. Data in G-I are mean ± SEM from control (n = 6) or treatment group (n = 9). Combined PD-1 antibody (αPD-1) with ATO (J) or TG003 (K) therapy was most effective on inhibiting mouse mEC25 ESCC allograft growth. Left, allografts growth curve; right, quantitative statistics of tumor size at the endpoint. Data are mean ± SEM from 8 (J) or 10 (K) allograft tumors. Data in A and D are mean ± SEM from three biologically independent experiments. Panels in C, F, G and I show immunofluorescence images (left) and quantitative statistics (right). Two-sided Wilcoxon rank-sum test for all P values in this figure, except two-sided Student’s t-test for A and D. Source data are provided as a Source Data file.

We further performed another set of in vitro and in vivo experiments using TG003, a broad-spectrum of RNA splicing inhibitor that inhibits splicing factor phosphorylation⁶⁴. As reported, TG003 treatment reduced RBM28 phosphorylation (Fig. S8I) but significantly elevated dsRNA levels in both KYSE30 cells and mEC25 cells, which could be reversed by ectopic overexpression of RBM28 (Fig. S8J − S8K). TG003 treatment also elevated the levels of IFNB and the downstream ISG expression and this effect could be counteracted by ectopic expression of RBM28 (Fig. S8L). We then treated mice bearing ESCC induced by 4-NQO with TG003 (Fig. S9A) to look at whether this RNA splicing inhibitor works in suppressing primary tumors. TG003, given from day 148 to day 184 after 4-NQO attack, significantly reduced the tumor volume compared with vehicle control and more importantly, we found that despite administration at late stage of 4-NQO carcinogenesis, TG003 still significantly prevented 4-NQO-induced body weight loss and the occurrence rates of the carcinogen induced-precancerous lesions and ESCC (Figs. 7G, 7H, S9B and S9C). Parallelly, mIF staining analyses demonstrated that tumor tissues from TG003 treated mice had significantly more dsRNA content, more effector CD8⁺ T cell numbers, but fewer MDSCs than tumor tissues from control mice (Fig. 7I). Together, these results suggest that the immune suppression due to dsRNA and IFN reduction in ESCC caused by mutTP53-induced aberrant RBM28 expression might be overcome by the therapies targeting RNA splicing. Finally, we conducted experiments in mice bearing subcutaneous allograft derived from mEC25 cells to explore whether ATO and TG003 may enhance the therapeutic efficacy of PD1 antibody (Fig. S9D and S9E). We found that PD-1 antibody plus either ATO or TG003 treatment achieved better anti-tumor efficacy than PD-1 antibody, ATO or TG003 alone treatment (Figs. 7J, 7K, Fig. S9F and S9G). Furthermore, the combined therapy did not significantly alter mouse body weight gain (Fig. S9H and S9I), indicating it is safe in mice.

For cell lines of other cancer types, we found that ATO can inhibit RBM28 expression in MDA-MB-231 and NCI-H446 cells (Fig. S9J and S9K), both harboring mutTP53 that are vulnerability to ATO. Conversely, for PANC1 and SW620 cell lines, which had TP53 mutations either outside the DBD region or not targeted by ATO, ATO treatment did not significantly change the levels of RBM28 RNA (Fig. S9L). Nevertheless, we treated these cell lines (PANC1, SW620 and Cal27) with PRIMA-1MET (APR-246), another small molecule compound that can restore p53 tumor suppressor function⁶⁵, significantly inhibited their RBM28 RNA expression (Fig. S9M and S9N). Parallelly, flow cytometry analysis showed that both ATO and APR-246 treatment significantly elevated the dsRNA levels in these TP53 mutated cell lines (Fig.  S9O − S9S).

Discussion

It is well known that mutations in the tumor suppressor gene TP53 play a critical role in the initiation and progression of human cancer^3,4,66–68. However, the aberrant functions of mutp53 remain to be further unveiled although many pro-cancer mechanisms have been known. In the present study, using ESCC as a model where TP53 mutation rates are particularly high^{2,4,20,22–24}, we explored the potential non-canonic functions of mutp53 in cancer. By analyzing the multi-omics data of mouse and human esophageal tissues with various precancerous and cancerous lesions, we have revealed aberrant RNA splicing as an early event in ESCC tumorigenesis, which was caused by overexpression of splicing factors such as RBM28. Remarkably, we have found that mutp53 mutated at the DBD gains a transcriptional regulatory function driving RNA-splicing factor RBM28 overexpression. This non-canonic role of mutp53, via exploiting RBM28 and other possible splicing factors to splice TEs, inhibits the endogenous dsRNA production in cancer cells, suppresses type I IFN response pathway and anti-tumor immunity, thus promoting ESCC tumorigenesis and progression. We have also found that in ESCC and most other cancer types, the expression level of RBM28 is significantly higher in TP53-mutated tumors compared to their TP53-unmutated counterparts and RBM28-overexpressed tumors usually exhibit immunosuppressive TME and respond inefficiently to ICB therapy. Furthermore, we have demonstrated that some compounds that can restore p53 normal conformation can inhibit aberrant RNA splicing activity, increase endogenous dsRNA accumulation and enhance anti-tumor immunity, therefore preventing cancer initiation and progression and enhancing the drug vulnerability to ICB treatment. Thus, our study has unveiled an important non-canonic function of mutp53 in cancer, elucidating the underlying mechanism for TP53 mutations to cause anti-tumor immunosuppression (Fig. 8), which might provide a new avenue for the early intervention and treatment of cancers with TP53 mutation.

Fig. 8. A proposed action model for mutTP53 hijacking RBM28 to suppress double-stranded RNA triggered antitumor immunity.

In cancer cells, the DNA binding domain-mutated TP53 (mutTP53) loses its canonic tumor suppressor functions but gains non-canonic transcriptional regulatory functions, driving overexpression of RNA-splicing factors such as RBM28 to inhibit dsRNA formation from transposon elements (TEs). This dsRNA formation perturbation impairs dsRNA-triggered type I IFN signaling and its downstream gene expression, sequentially establishing immunosuppressive tumor microenvironment (TME) that promotes cancer initiation, progression and resistance to immune checkpoint blockade (ICB) therapy. In contrast, cancers with wild-type TP53 (wtTP53) have immunoreactive TME, which responds actively to ICB therapy and thus have better prognosis since the type I IFN-triggered anti-tumor immunity may normally function.

As a key tumor suppressor protein, TP53 is one of the most mutated genes in human cancers and the mutations in DBD are particularly prevalent. These mutations typically lead to the loss of p53 ability to bind to the promoters of target genes, thereby impeding the activation of classic tumor-suppressing pathways such as cell cycle arrest, DNA repair, or apoptosis. However, a growing body of recent studies suggests that some DBD-mutated TP53 not only lose tumor-suppressive functions but may also actively drive tumor progression through a gain-of-function mechanism. The finding in our study that DBD-mutated mutp53 acts as a transcription factor controlling RBM28 overexpression serves as a paradigmatic example of this phenomenon. Although wtp53 suppresses RBM28, we observed that cells expressing DBD-mutated mutp53 maintain RBM28 even higher than those in cells completely lacking wtp53 function and knockout mutTP53 reduce RBM28 levels. These results indicate that the elevation in RBM28 is not merely due to the loss of wtp53 repression, but stems from an active and gain-of-function mechanism. Mechanistically, mutp53 may acquire additional transcriptional regulatory properties through conformational changes or abnormal protein interactions. For instance, mutp53 can form complexes with chromatin remodeling complexes (such as SWI/SNF) or other transcription factors (e.g., NF-κB or members of the AP-1 family), hijacking the promoter regions of non-canonic target genes^9,69,70. However, in the present study, we have demonstrated that mutp53 can directly bind to the RBM28 promoter to activate RBM28 transcription. We have found a C-rich cis-element in the promoter of RBM28 and some other RNA splicing factor genes, which has been documented to be a mutp53-binding element in other non-canonical target genes in previous studies^37,38,71. Notably, different p53 mutants exhibit varying efficacies in binding RBM28 promoters and regulating RBM28 expression, yet they consistently activate its transcription through this gain-of-function mechanism rather than through mere loss of function. We also observed that wtp53 suppresses RBM28 expression. This repression may involve indirect regulatory mechanisms, rather than direct promoter binding, as suggested by the absence of RBM28 promoter occupancy in both public ChIP-seq datasets and our own ChIP-qPCR and EMSA experiments. Nevertheless, the precise mechanism by which wtp53 suppresses RBM28 remains an important question for future studies.

Notably, because the overexpression of RBM28 and other RNA splicing factors can suppress endogenous dsRNA formation and the type I IFN pathway activation, the results in the present study provide a direct molecular basis linking TP53 mutations to immunosuppression in ESCC. It is well-known that DBD mutations frequently occur in various cancers other than ESCC, suggesting that this mechanism may be common for tumors to escape from immunosurveillance. Furthermore, mutp53 is often more stable than wtp53, and its accumulation may continuously drive RBM28 overexpression, establishing an immunosuppressive TME. Indeed, by analysis of the TCGA data, we have found that significant increment of RBM28 level and suppression of anti-tumor immunity are present in most cancers with mutTP53 compared to cancers with wtTP53 and related to resistance to ICB therapies. These findings suggest that mutTP53-driven aberrant RBM28/dsRNA/IFN axis contributes to anti-tumor immunosuppression, which is a pan-cancer phenomenon and thus has broad implications.

RBM28, a nucleolar RNA-binding protein involved in spliceosomal ribonucleoprotein complexes⁷², plays a pivotal role in RNA processing and ribosome biogenesis⁷³. Its dysregulation in cancers may disrupt normal splicing events, thus diminishing the accumulation of endogenous dsRNA, a key trigger for innate immune activation via cytosolic RNA sensors like RIG-I and MDA5¹². Furthermore, such widespread splicing alterations could potentially generate novel antigenic peptides that enhance T cell-mediated anti-tumor immunity¹¹. dsRNA derives from dysregulated or bidirectional transcription of TEs although which TE classes contribute to dsRNA formation remain contentious. While IR-Alus (a SINE subclass) have been suggested as a major dsRNA source activating broad sensor pathways^17,74, compelling evidence indicates that LTRs and LINEs—but not SINEs—drive dsRNA formation and interferon response^43,75–77. In the present study, we have demonstrated that RBM28 binds all three TE classes (LTRs, LINEs, and SINEs) thereby inhibiting their transcription and subsequent dsRNA formation rather than through splicing-related neoantigen peptides. However, Alu elements, the dominant primate SINEs, are absent in mice. Given our cross-species design encompassing human and mouse systems, we therefore prioritized LTRs and LINEs for analysis. The suppression of dsRNA signaling could thus impair the production of type I IFNs (e.g., IFNβ), which are essential for recruiting immune cells and establishing an anti-tumor microenvironment. Indeed, we have found that mutp53-induced upregulation of RBM28 diminishes cellular dsRNA accumulation, reduces IFNβ production and the expression of downstream ISGs such as CXCL10 and CCL5, forming an immunosuppressive TME featured by increased MDSC infiltration but decreased CD8⁺ T cell infiltration. This mechanism has been observed with FTSJ3, an RNA methyltransferase, which promotes hepatocellular carcinoma immune evasion by inhibiting dsRNA-induced IFNβ signaling⁴⁴, and with KRAS mutation, which drives immunosuppression in colorectal cancer by accelerating dsRNA degradation and impairing IFN response¹⁶. In the present study, we have shed light on an important cancer immunosuppression mechanism that mutp53-caused RBM28 overexpression compromises dsRNA-dependent IFN response, thereby creating an immunosuppressive niche favorable for cancer initiation and progression, and even poor response to ICB therapies. These findings may further improve our understanding of the interaction between malignant epithelial cells and immune cells in cancer development and resistance to immunotherapy. It is worth noting that previous studies have reported that RBM28 can bind wtp53 at the DBD, suppressing p53’s transcriptional activity and consequently inactivating the p53 signaling pathway^78,79. We have found that in human and mouse ESCC cells where RBM28 level is high, mutp53 can still significantly activate RBM28 transcription. These results suggest that RBM28 may not be able to bind mutp53 mutated at the DBD although this assumption needs to be experimentally validated. Interestingly, a recent study has shown that loss of p53 function reduces H3K9me3 at the TE promoters and decreases H3K27Ac at the promoters of ISGs, resulting in elevated transcription of immunogenic repetitive elements and chronic viral mimicry activation and reduced ISG expression, which diminishes cellular immunogenicity¹⁸. The divergent mechanisms further underscore the role of p53 dysfunction in suppressing anti-tumor immunity via IFN pathway.

The discovery of small-molecule compounds such as ATO and APR-246 capable of restoring normal function to mutp53 and subsequently suppressing RBM28 expression^63,65 and TG003, a specific inhibitor of Cdc2-like kinases that regulates the alternative splicing of pre-mRNA⁶⁴, provides not only evidence supporting the role of mutp53/RBM28/dsRNA/IFN axis in suppressing anti-tumor immunity but also a promising therapeutic strategy for targeting tumors driven by p53 dysregulation. In the present study, we have revealed that ATO and APR-246 can suppress RBM28 overexpression, reverse cellular dsRNA depletion, reengage cytosolic sensors (e.g., RIG-I/MDA5), and reinstate IFN-driven anti-tumor immunity, suggesting that their effects are through rescuing wtp53 activity. These effects enable the drugs to suppress the growth of allografts derived from mutTP53 ESCC cells in mice. In the meantime, we have found that TG003 has the same effects as the wtp53 activity rescued by ATO and demonstrated its effect in suppressing the initiation and progression of mouse primary ESCC induced by carcinogen 4-NQO. These results clearly indicate that the mutp53-RBM28-dsRNA signaling axis is a therapeutic target for ESCC and even other cancer types. Since p53 has its canonic tumor suppressor functions, restoring normal function to mutp53 by the drugs such as ATO or APR 246 may not only function on the anti-tumor immunity but other p53 function related pathways. Notably, a growing body of evidence suggests that these compounds can also inhibit tumor growth through p53-independent mechanisms. For instance, APR-246 has been reported to trigger ferroptosis⁸⁰ and alter the cellular GSH/ROS balance⁸¹. Similarly, it has been reported that ATO, independent of its role in restoring p53 function, induces increased ROS production and DNA damage in ESCC cells, resulting in the degradation of Cyclin D1 and functionally suppresses the CDK4/6 pathway, thereby sensitizing oral squamous cell carcinoma and ESCC to checkpoint inhibitor treatment⁸². This dual mechanism, i.e., direct tumor cell killing via canonic p53 tumor suppressor functions and immune microenvironment remodeling via dsRNA signaling, positions such compounds as potential gamechanger in cancers with TP53 mutations and RBM28 overexpression. Indeed, we have also demonstrated that ATO and TG003 have synergic effect with PD1 antibody on treatment of ESCC in mice. These results open an avenue for conducting clinical trials to examine whether ATO, an FDA-approved drug, can elevate therapeutic impact of ICB in human solid tumors where TP53 mutation is common because the response rates of solid tumors to ICB are low.

In summary, we have unveiled an important mechanism for TP53-mutated ESCC and other multiple cancer types to evade anti-tumor immunosurveillance. Mutp53 plays a non-canonical transcriptional regulatory function which directly drives RBM28 overexpression, resulting in reduced endogenous dsRNA formation, impaired anticancer immunity, and thus allowing tumors to create immunosuppressive TME that promotes cancer initiation, progression and resistance to immunotherapy. These findings hold profound significance for understanding the actions of TP53 mutations in the initiation and progression of cancers and for advancing cancer intervention and treatment.

Methods

Ethics statement

Our research adheres to all applicable ethical guidelines. The study protocol was approved by the Institutional Review Boards of Cancer Hospital, Chinese Academy of Medical Sciences (NCC2022C-141 and NCC2024C-836). All patient specimens were obtained with informed consent. Animal experiments were performed in accordance with the protocols approved by the Institutional Animal Care and Use Committee of the Chinese Academy of Medical Sciences (NCC2021A271). The approved guidelines specified that mice were to be euthanized when the estimated tumor burden exceeded 10% of their normal body weight or when the maximum allowed tumor volume exceeded 2000 mm³. Throughout this study, no animal reached or exceeded these limits.

Establishing a mouse model carrying esophageal epithelium-specific APEX2

C57BL/6 J mice (male and female, 6–8 weeks) with APEX2^flox/flox were kindly gifted by Dr. Mo Li of Peking University Third Hospital. These mice were crossed with ED-L2-Cre mice (male and female, 6-8 weeks, Shanghai Model Organisms Center) to generate genetically engineered mice with APEX2 overexpression in the esophageal epithelium (APEX2^+/+: ED-L2-Cre; APEX2^F/F). The genotypes of mice were verified via PCR using the primers detailed in Supplementary Data 3. Mice were maintained in a local housing facility under the controlled conditions of 23 ±  1 °C, 50  ±  10% humidity, and 12 − 12 h light-dark cycle.

Induction of ESCC in mice with esophageal epithelium-specific APEX2

We induced mouse ESCC models as described previously²¹. Briefly, mice with esophageal epithelium-specific APEX2 were treated with 4-NQO (Sigma-Aldrich) at a concentration of 100 μg/mL drinking water for 16 weeks. Drinking water containing the carcinogen was replaced weekly with freshly prepared solution, and mice had ad libitum access throughout the treatment period. After 16 weeks, 4-NQO solution was replaced with sterile water until the experiment endpoint. The esophageal samples obtained at various times, i.e., week 0 (NOR), week 18 (LGIN), week 22 (HGIN), and week 26 (ESCC) were subjected to APEX2-mediated biotinylation and proteomics analysis.

APEX2-mediated biotinylation of mouse esophageal tissues

Mouse esophageal tissues expressing ED-L2-APEX2-EGFP from biologically distinct mice in the NOR (n = 4), LGIN (n = 4), HGIN (n = 4) and ESCC (n = 5) groups were incubated with 1 mM biotin-phenol in DMEM medium for 2 h at 37 °C. H₂O₂ at a final concentration of 2.5 mM was added to initiate the reaction. To distinguish specific biotinylation from background binding, control mouse esophageal tissues (n = 4) were processed identically but in the absence of biotin-phenol. Twenty minutes later, the reaction was quenched with an equal volume of quencher solution consisting of 10 mM sodium ascorbate, 10 mM sodium azide, and 5 mM Trolox in Dulbecco’s phosphate-buffered saline. After homogenization, tissues were lysed in lysis buffer containing phenylmethanesulfonyl fluoride, 1× protease inhibitor cocktail and a quencher solution for 20 min. Lysates were sonicated and centrifuged at 12,500 x g. Then ice-cold methanol was added to precipitate protein at −80 °C for 6 h. After centrifugation at 1500 x g for 30 min, proteins were redissolved in 1% SDS RIPA buffer. The protein sample (1.0 mg in 0.2% SDS RIPA buffer) was incubated for 2 h with 150 μl of streptavidin-coated magnetic beads slurry (Thermo-Fisher, No. 88817). The beads were then washed twice with 1 ml of RIPA lysis buffer, once with 1 ml of 1 M KCl, once with 1 ml of 0.1 M Na₂CO₃, once with 1 ml of 2 M urea in 10 mM Tris-HCl (pH 8.0), and twice with 1 ml RIPA lysis buffer. Protein-bound streptavidin beads were then resuspended in 200 μl of digestion buffer (50 mM HEPES, 1 μM CaCl₂ and 2% acetonitrile) and digested in 2 μg of trypsin/LysC mixture overnight at 37 °C. Magnetic beads were removed by magnetization and the digested supernatant was collected. A 10 μl of sample was saved for Pierce fluorometric peptide quantification assay. Peptides were frozen and dried in a vacuum concentrator before mass spectrometry analysis.

Mass spectrometry analysis and protein identification

A total of 21 samples were analysed by LC-MS/MS. DDA spectral-library construction and DIA protein quantification were performed by Novogene Co., Ltd. (Beijing, China) using an EASY-nLC 1200 nano-UHPLC system coupled to a Q Exactive HF-X Orbitrap mass spectrometer (Thermo Fisher Scientific). Peptides were separated using a self-packed trap column (4.5 cm × 75 μm, 3 μm) and a self-packed analytical column (15 cm  × 150 μm, 1.9 μm) at a flow rate of 600 nl min⁻¹. Mobile phase A was 0.1% formic acid in water, and mobile phase B was 0.1% formic acid in 80% acetonitrile. The gradient was 5–8% B over 1 min, 8–30% B over 75 min, 30–50% B over 5 min, and 50–95% B over 1 min, followed by washing and column equilibration, with a total run time of 100 min. For DDA, full MS spectra were acquired over m/z 350–1500 at a resolution of 120,000, with an AGC target of 3 × 10⁶ and a maximum injection time of 80 ms. The 40 most intense precursor ions were fragmented by higher-energy collisional dissociation at a normalized collision energy of 27%. MS2 spectra were acquired at a resolution of 15,000, with an AGC target of 5 × 10⁴, a maximum injection time of 45 ms, an intensity threshold of 1.1 × 10⁴, and a dynamic-exclusion time of 20 s. For DIA, 4 μg peptide from each sample was mixed with 0.8 μl iRT reagent, and half of the resulting sample volume was injected for analysis. The spray voltage was 2.1 kV, and the capillary temperature was 320 °C. Full MS spectra were acquired over m/z 350–1500 at a resolution of 60,000, with an AGC target of 5 × 10⁵ and a maximum injection time of 20 ms. DIA MS2 spectra were acquired using 50 variable-width isolation windows covering m/z 400–1004 at a resolution of 30,000, with an AGC target of 1 × 10⁶ and a normalized collision energy of 27%. DDA raw files were searched using the Pulsar search engine in Spectronaut (version 18, Biognosys) against the UniProt Mus musculus database (Mus_musculus_uniprot_2022_9_5.fasta; 86,436 sequences) to construct the spectral library. Trypsin was specified as the protease, allowing up to two missed cleavages. Carbamidomethylation was specified as a fixed modification, whereas methionine oxidation and protein N-terminal acetylation were specified as variable modifications. The precursor- and fragment-ion mass tolerances were 10 ppm and 0.02 Da, respectively. Peptide and protein identifications were filtered at a false discovery rate of 1%, and proteins were required to contain at least one unique peptide. DIA raw files were searched against the DDA spectral library in Spectronaut for chromatographic peak extraction, peptide identification, and fragment-ion-based protein quantification. Retention times were corrected using iRT calibration, and the precursor Q-value cutoff was 0.01.

A total of 7299 proteins were identified, and their quantitative abundances were log₂-transformed, with missing values imputed as one-tenth of the minimum observed abundance. Proteins detected in at least three replicates in any group were retained (n = 7,020). Differences between each experimental group and the negative control group were assessed using a two-sided Wilcoxon rank-sum test, followed by Benjamini–Hochberg false discovery rate (FDR) correction for multiple comparisons. Among these, 5263 proteins (Supplementary Data 1) were significantly enriched in at least one experimental group relative to the negative control group (fold change >1 and adjusted P < 0.05) and were included in subsequent analyses. We then performed principal component analysis (PCA) on a total of 5263 proteins of 17 samples to illustrate the global proteomic difference among the four stages of esophageal tumorigenic lesions. Dimension reduction was performed using scatterplot3d R package (version 0.3–44).

Consensus clustering analysis and enrichment analysis of proteomics data

The time-series protein expression matrices of 17 mouse samples and 789 human samples were analyzed to identify proteomic subtypes using the consensus clustering method. Consensus clustering was performed with the ClusterGVis R package (version 0.1.1; https://github.com/junjunlab/ClusterGVis). EnrichCluster function was used to do batch enrichment analysis for each cluster based on the GO term data. The top 5 pathway terms ordered by P value were extracted to display.

Identification of aberrant alternative splicing events in mouse ESCC

The generateEvents module of SUPPA2 (version 2.3)⁸³ was initially employed to identify alternative splicing events based on the master transcriptome. Subsequently, we utilized the psiPerEvent module of SUPPA2 to quantify event inclusion levels measured by percent spliced in (PSI) values for every identified alternative splicing event in each sample. Alternative splicing events with PSI > 0.1 in at least 20% ESCC samples were kept for further downstream analysis. To explore the aberrant alternative splicing events, Wilcoxon signed-rank test was performed to evaluate the difference in PSI score for each alternative splicing event between ESCC and adjacent nontumor samples. The alternative splicing events with adjusted P value < 0.05 and the absolute value of ΔPSI > 0.1 (the difference of the median PSI of each alternative splicing events between tumor and normal) were identified as aberrant alternative splicing events.

Gene set variation analysis

Gene set variation analysis (GSVA) was performed using the GSVA package⁸⁴ (version 1.34.0) to derive pathway activity scores for our multi-omics datasets, including spatial transcriptomics, single-cell RNA sequencing, proteomics, and bulk RNA sequencing. These scores reflect the relative activation level of each pathway per sample. Pathway gene sets were obtained from the Molecular Signatures Database (MSigDB, version 6.2).

Spatial transcriptomic data analysis

We obtained spatial transcriptomic expression of Patient 1 (NOR and LGIN samples with wtTP53 and ESCC sample with TP53 biallelic loss) and Patient 2 (NOR sample with wtTP53 and HGIN and ESCC samples with TP53 biallelic loss) from our previously published study¹⁹. The tumor neighborhood score for various cell types was calculated following the published method⁸⁵. The tumor neighborhood was defined as the seven spots surrounding a tumor spot including the tumor spot itself. For each tumor spot, the tumor neighborhood score for a specific cell type was determined by averaging the corresponding cell-type scores within the tumor neighborhood. The Euclidean distance between the spot with a score of CD8⁺ T cells or a score of MDSCs and the RBM28⁺ spot estimated based on their respective coordinates⁸⁵ was then used to analyze the Spearman correlation between the CD8⁺ T cell or MDSC infiltration scores and distances.

Cell lines, cell culture and drug treatment

Human ESCC cell lines KYSE30, KYSE450 and KYSE510 were generous gift from Dr. Y. Shimada of Hyogo College of Medicine, Japan. Mouse ESCC cell line mEC25 was gifted by Dr. Li Fu of Shenzhen University International Cancer Center. Immortalized esophageal epithelial cell line HET-1A and human pancreatic cancer cell line PANC-1 were purchased from American Type Culture Collection. MDA-MB-231, Cal-27 and SW620 cell lines were obtained from the Cell Bank of the Institute of Basic Medical Sciences, Chinese Academy of Medical Sciences (Beijing). Human small-cell lung cancer cell line NCI-H446 was purchased from the Cell Bank of Type Culture Collection of Chinese Academy of Sciences Shanghai Institute of Biochemistry and Cell Biology (Shanghai). KYSE30, KYSE450, KYSE510, SW620 and NCI-H446 cells were cultured in RPMI-1640 medium with 10% fetal bovine serum (FBS), while HET-1A, mEC25, Cal-27 and PANC1 were cultured in Dulbecco’s modified Eagle’s medium with 10% FBS in a humidified incubator with 5% CO₂, routinely authenticated by short tandem repeat analysis, and regularly treated with a mycoplasma-removing agent. The cultured cell lines were treated with ATO (10 μM, Hayida Yaoye CO., LTD.), APR-246 (40 μM, SML1789, Sigma-Aldrich) or TG003 (200 nM, HY-15338, MCE) for 48 h.

Establishing cell lines with altered TP53 or RBM28 expression

The CRISPR/Cas9 system was used to knock out TP53, RBM28/Rbm28 in human or mouse cell lines. Cas9 and single-guide RNAs targeting the genomic TP53, RBM28 or Rbm28 sequence were designed using the CRISPR design tool and cloned into the GV708 plasmid by Genechem (Shanghai). Cells were infected by the CRISPR/Cas9 virus in the complete medium for 48 h and selected by puromycin. For Trp53 knockdown in mEC25 cells, lentivirus carrying shRNA targeting Trp53 was utilized, with the Trp53 sequence cloned into the LVN10 vector (GenePharma, Suzhou). For establishing cell lines with ectopic RBM28 overexpression, RBM28 sequence was cloned into the GV492 vector. The lentivirus was packed by Genechem (Shanghai). mEC25 and KYSE30 cells were infected with lentivirus and selected by puromycin. The overexpression or silencing efficiency of TP53 and RBM28 was examined by Western blotting assay. All sequences used are provided in Supplementary Data 3.

RNA extraction and real-time quantitative PCR analysis

We extracted total RNA from cell or tissue samples using the RNA-Quick Purification Kit (ES Science) and conducted reverse transcription using PrimeScriptTM RT Master Mix (Takara). The mRNA level was detected in QuantStudio 5 Real-Time PCR system and Quantagene q900 qPCR system (Kubo Technology) with TB Green® Premix Ex TaqTM II (Takara) and the corresponding primers synthesized by TSINGKE Biological Technology (Beijing, China) (Supplementary Data 3). The expression level of the specified gene was determined using ΔΔCt relative to the expression of GAPDH or β-ACTIN as internal control.

Western blot assay

We extracted total protein from tissues or cells using RIPA lysis buffer (Beyotime) containing Protease and Phosphatase Inhibitor Cocktail (NCM Biotech). Total protein separated on SDS-PAGE (Meilunbio) was transferred to PVDF membrane (Millipore) and then incubated with primary antibody overnight followed by incubation with secondary antibody (Cell Signaling Technology). Antibody against MDA5 (5321), RIG-I (3743), TBK1 (3504), phospho-TBK1 (Ser172, 5483), IRF3 (11904), or phospho-IRF3 (Ser396, 29047) was from Cell Signaling Technology. Antibody against GAPDH (60004-1-Ig), MAVS (66911-1-Ig), RBM28 (16484-1-AP), SNRPB2 (13512-1-AP), PRPF4 (10728-1-AP) or TCERG1 (21858-1-AP) was from Proteintech. Antibody against p53 (sc-126) or PP2Cγ (sc-390625) was from Santa Cruz Biotechnology. Protein signals were detected with Amersham Imager 600 using a high sensitivity ECL chemiluminescence detection kit (Meilunbio). Phosphorylated RBM28 was detected using Phos-Assay SDS-PAGE (Vazyme), which was prepared by the addition of 25 μM Phos-assay and 50 μM MnCl₂ at the final concentration. Uncropped and unprocessed scans of the blots are provided in the Source Data file.

Chromatin immunoprecipitation and quantitative PCR

The chromatin immunoprecipitation (ChIP) assay was performed using the SimpleChIP Plus Sonication Chromatin IP Kit (Cell Signaling Technology). KYSE510 cell lines stably expressing various TP53 mutants were cross-linked with 37% formaldehyde for 12 min at room temperature. After quenching with glycine, cells were lysed and nuclei were isolated using cell lysis and nuclear lysis buffers. Chromatin was fragmented using a Covaris S220 Focused-ultrasonicator with the following parameters: Peak Power: 140w, Duty Factor: 5%, Cycles per Burst: 200, Duration: 540 s. The sheared chromatin was incubated overnight at 4 °C with 5 μg of p53 antibody (sc-126x, Santa Cruz), 10 μg H3K27ac (#8173, CST) or normal IgG as a negative control. Three independent biological replicates were performed for each experimental group. The eluted DNA was treated with RNase A and proteinase K, and purified DNA was analyzed by quantitative PCR using primers specific for the RBM28 promoter. CDKN1A and KMT2D promoters were used as positive and negative controls for wild-type and mutant p53 binding, respectively (Supplementary Data 3). The ChIP enrichment efficiency was calculated using the following formula: Enrichment percentage = 2% × 2^{(Ct[Input] - Ct[IP])}, where Ct value represents the quantification cycle of the PCR reaction. The fold enrichment was determined as the ratio of enrichment percentage in the antibody group to that in the IgG control group. Statistical significance was determined using Student’s t-test comparing three independent biological replicates.

Electrophoretic mobility-shift assays

Following the instructions of the manufacturer (Thermo Fisher Scientific), binding reactions containing purified wild-type or mutant p53 proteins (CUSABIO) and biotin-labeled oligonucleotides (with or without unlabeled competitor, Beijing TSINGKE Biological Technology) were performed to form protein-nucleic acid complexes. The reaction mixtures were electrophoresed on a non-denaturing polyacrylamide gel (Beyotime) in 0.5X Tris-Borate-EDTA (TBE) buffer at 110 V for 100 min at 4 °C. The complexes were then transferred to a positively charged nylon membrane (Cytiva) at 350 mA for 40 min. Finally, the DNA was UV cross-linked to the membrane at 120 mJ/cm² and visualized using the Smart-EMSA™ EMSA Kit (IF9501, Engibody) and LightShift Chemiluminescent EMSA Kit (20148, Thermo-Fisher). The sequences of the RBM28 probes are provided in Supplementary Data 3.

Luciferase reporter gene assay

The RBM28 promoter sequence produced by PCR was cloned into the GV238 vector (MCS-firefly_Luciferase). The reporter plasmid containing the RBM28 promoter with or without the TP53 binding motif was designed by Genechem in Shanghai (Fig. 2H). KYSE510 cells were seeded in 48-well plates and infected with the plasmids using jetPRIME® (Polyplus). When cells grew to 80% to 90% confluency, luciferase activity was determined according to the instruction (Promega).

Cell viability assay

The cell viability was evaluated using the CCK8 (CK04, Dojindo Lab) assays. Cells were seeded in 96-well plates at a density of 3000 cells/well. CCK8 solution was added to each well and cell viability was determined at 450 nm after incubation for 2 h.

Establishing mouse xenograft and allograft models and drug treatment

We established mouse ESCC xenograft or allograft models by transplanting Rbm28-knockout or non-knockout, or Trp53 knockdown or non-knockdown, mEC25 cells (8 × 10⁶ in 150 μl of growth factor-reduced Matrigel) into the hind leg of C57BL/6 J (male and female, 6–8 week, Beijing HFK Bioscience) and Rag2^{-/ -}mice (male and female, 6–8 weeks, Beijing Biocytogen Pharmaceuticals) and RBM28-knockout or non-knockout mEC25 cells or KYSE30 cells (5 × 10⁶) into the hind legs of NSG (NOD/SCID/IL-2Rγnull, male and female, 6–8 weeks, Beijing HFK Bioscience) mice. The tumor growth rate was monitored using manual caliper measurements every 2 days starting from day 5 for C57BL/6 J mice and day 15 for NSG mice. The mouse body weight was also measured every 2 days. Tumor volume was determined by length × width² × 0.52. For drug treatment, PD-1 antibody (BE0146, BioXCell) or IgG (BE0089, BioXCell; both 200 μg/tumor) was injected intratumorally every 2 days, TG003 (30 mg/kg in DMSO, every 2 days) and ATO (5 mg/kg in PBS, daily) was given intraperitoneally for a total of 5 doses.

Treatment of carcinogen-induced mouse primary ESCC

C57BL/6 J mice (male and female, 6–8 weeks) suffered from precancerous lesions and ESCC were intraperitoneally treated with TG003 (30 mg/kg; HY-15338, MCE) every 2 days starting from week 22 after 4NQO administration. The body weight of each mouse was examined every 4 days during the experiment. For esophageal-specific knockdown of Rbm28, AAV-shRbm28 or its control (Genechem, Shanghai) (Supplementary Data 3) was injected into the lower part of the esophagus at multiple spots by surgery under anesthesia. On week 26, the esophagus of each mouse was removed and the tumor number counted. The tumor volume was measured and expressed as length × width² × 0.52. The body weight of each animal was examined every 3 days during the experiment.

Immune cell infiltration analysis

We used single-sample gene set enrichment analysis to quantify the relative abundance of specific immune cell types in the tumor microenvironment of human ESCC tissue samples. The Tumor Immune Dysfunction and Exclusion algorithm was used to calculate the abundance of MDSCs and model the distinct tumor immune evasion mechanisms including dysfunction of tumor-infiltrating cytotoxic T lymphocytes (CTLs) and the exclusion of CTLs by immunosuppressive factors⁸⁶. The single-cell RNA-sequencing data of NOR, LGIN, HGIN and ESCC samples were obtained from our previously published studies^19,25. Samples were divided into RBM28 high or RBM28 low group using the median RNA level as a cutoff and the proportion of immune cells and CD8⁺ T cell subtypes in each sample were calculated.

Multiplexed immunofluorescence staining assay

We performed multiplexed immunofluorescence staining on formalin-fixed and paraffin-embedded tissue section with the 7-color multi-labeling kit (PANOVUE). The tissue section were incubated with primary antibodies against CD8 (ab217344) and panCK (ab9377) from Abcam, GZMB (46890) from Cell Signaling Technology, CD11b (14-0112-82) and Ly-6G/Ly-6C (14-5931-82) from eBioscience, RBM28 (16484-1-AP) and KRT6A (10590-1-AP) from Proteintech and dsRNA (10010500) from SCICONS. The sections were counterstained with DAPI for nuclear visualization and coverslipped with anti-fade mounting medium. Multispectral images were scanned using NanoZoomer S60 (Hamamatsu) and the integrated fluorescence density of each sample was analyzed using Image J software.

Flow cytometry analysis

Fresh tissues were minced and incubated in RPMI 1640 medium containing collagenase I (17018029, Gibco), collagenase IV (17104019, Gibco), and DNase I (10104159001, Roche) at 37 °C for 1h, followed by filtering through a cell strainer to prepare a single cell suspension. The cell suspension was treated with RBC lysis buffer (555899, BD). Cell surface markers were stained with fluorophore-conjugated antibodies in the dark in 2% FBS-PBS for 1 h at 4 °C. Flow cytometric analysis was performed on flow cytometer (FACSMelody, BD) and data was analyzed using FlowJo software (Treestar). Antibodies used were Zombie Viability Kit (423105), PerCP anti-mouse CD45 (103129), APC anti-mouse CD8a (100712), PE anti-mouse CD11b (101207), PE/Cyanine7 anti-mouse Ly-6G/Ly-6C (Gr-1, 108416), PE/Cyanine7 anti-mouse CD11b (101215), APC anti-mouse Ly-6G (127613), PE anti-mouse Ly-6C (128007) from Biolegend.

For intracellular dsRNA staining, Foxp3/Transcription Factor Staining Buffer Set Kit (00-5523-00, Ebioscience) was used for fixation and permeabilization in the dark for 1 h. Cell suspension was incubated with antibody J2 against dsRNA (10010500, SCICONS) for 1 h, followed by incubation with Dylight 649, Goat Anti-Mouse IgG as secondary antibody (A23610, Abkine) in the dark for 1 h. FlowJo software was used to analyze the fluorescence intensity of flow cytometry data. The relative dsRNA MFI was calculated as the MFI value normalized to that of the control group.

Immunofluorescence assays

KYSE30 and mEC25 cells were grown to 50–70% confluency on the coverslips (Aoqing Biotechnology). After washing, cells on coverslips were fixed with 4% paraformaldehyde and permeabilized with 0.2% Triton X-100. The coverslips were blocked with 3% bovine serum albumin and then incubated with antibody against RBM28 (16484-1-AP, Proteintech) and dsRNA (10010500, SCICONS). Goat anti-mouse secondary antibody with Alexa Fluor 594 (A-11005) and goat anti-rabbit secondary antibody with DyLight 633 (35563) from Thermo-Fisher were used as the secondary antibody. Confocal microscopy (AKOYA) was then used to capture the images after coverslips were sealed with mounting medium containing DAPI (Beyotime). Three representative fields from each sample were selected and the fluorescence intensity of each cell in the select field was analyzed using Image J software. Relative RBM28/dsRNA MFI was calculated as the MFI value normalized to that of the control group.

Enzyme-linked immunosorbent assay

KYSE30 were first cultured in medium containing 10% FBS for 12 h and then cultured in serum-free medium for 36 h. The concentration of IFNβ, CCL5, CXCL10 and ISG15 in culture supernatant was determined by enzyme-linked immunosorbent assay using the ELISA kits from R&D Systems.

RNA sequencing and data analysis

Total RNA from cells were isolated using the TRIzol reagent (Invitrogen) and subjected to sequencing by Novogene Co., Ltd. (Beijing). RNA-sequencing data were mapped to the GRCh38 human genome by HISAT2 (version 2.0.5). The gene expression matrix of raw read counts after annotation by FeatureCounts (version v1.5.0-p3) was processed using limma (version 3.58.1). Gene set enrichment of differentially expressed genes was analyzed based on REACTOME pathways.

RNA-sequencing analysis of transposon element expression

The expression of transposon elements in RBM28 knockout or, RBM28 non-knockout KYSE30 cells and in mouse multi-stage lesion esophageal tissue samples²¹ were analyzed with rRNA-depleted RNA-sequencing. Residual rRNA reads were removed using SortMeRNA. Post-filtered reads were aligned to the human reference genome (hg38) for KYSE30 cells or mouse reference genome (GRCm38) for mouse tissue samples using the STAR algorithm (v2.7.10b). STAR index files were generated using the genomeGenerate function with the parameter “--sjdbOverhang 100”.

TE quantification at the family level was performed using TEtranscripts (v2.2.3)⁸⁷ with the parameters “--sortByPos --stranded reverse --mode multi --minread 1 -i 10”. Read counts were assigned using the annotated GRCh38_Ensembl_rmsk_TE.gtf (for human) and GRCm38_Ensembl_rmsk_TE.gtf (for mouse) files downloaded from https://hammelllab.labsites.cshl.edu/software/. To further characterize the genomic localization of TEs, we also performed locus-level quantification using the command: ‘featureCounts -s 2 -t exon -g transcript_id -p -f -a hg38_GENCODE_rmsk_TE.gtf -o “outfile” “bam file” –primary -Q 20 –ignoreDup‘. We then filtered TEs present in intergenic or intronic regions and directly extracted count values from RNA-seq BAM files using SAMtools for differential analysis with DESeq2.

RNA immunoprecipitation assay

We performed RNA immunoprecipitation assay using the Magna RIP™ Kit (Millipore). Cell lysate was sonicated and subjected to immunoprecipitation at 4 °C overnight with antibody against RBM28 or dsRNA, with a homologous IgG serving as a control. The RNA from the input and immunoprecipitation sample was subsequently treated with proteinase K, extracted by phenol/chloroform, and precipitated by ethanol. The purified RNA was converted to cDNA and detected by RT-qPCR assay and the level was determined by normalizing the Ct value obtained from the immunoprecipitated sample against the Ct value from the input fraction. For reactions with no detectable amplification, the Ct value was set to the predefined maximum Ct value of the assay.

Crosslinking and Immunoprecipitation followed by quantitative PCR (CLIP-qPCR)

Crosslinking and immunoprecipitation assays were conducted using a commercial CLIP Kit (IEMed, K319). KYSE30 cells and mEC25 cells were crosslinked by exposure to 254 nm UV light at a dose of 0.15 J/cm² for 1 min. The cells were subsequently sonicated on ice using a cycle of 5 s ON and 5 s OFF at 500 W for 30 min. Lysates were incubated with anti-RBM28 antibody (16484-1-AP, Proteintech) and homologous IgG on a vertical mixer with shaking at room temperature for 1 h. RNA was extracted using polysome elution buffer and Proteinase K treatment. The supernatant was collected using a magnetic stand and further purified with a Nucleic Acid BF C column. Finally, qPCR was performed to assess RNA enrichment. The RNA enrichment fold change was calculated using the relative quantification 2^–ΔΔCt method, normalized to the Input and corrected by the IgG group. All primers used are listed in Supplementary Data 3.

ESCC patients treated with ICB and their tissue sample collection

A total of 100 patients with ESCC were recruited at the First Affiliated Hospital of Harbin Medical University. All patients received 2–4 cycles of neoadjuvant chemotherapy (nab-paclitaxel and nedaplatin) plus immune checkpoint inhibitor (Camrelizumab) therapy. The ESCC tissue biopsy sample was taken from each patient before the neoadjuvant therapy. The neoadjuvant therapeutic efficacy was assessed by pathological tumor regression grading⁸⁸ in surgical removed post neoadjuvant treatment sample, classified as responder or non-responder. The biopsy specimen taken from each patient before the neoadjuvant therapy was analyzed for RBM28 and dsRNA using multiplex immunofluorescence staining.

Publicly available data used in this study

For enhancing statistical power or further validation, we used the publicly available proteomics data of 786 ESCC patients and somatic mutation data of 90 ESCC patients from a previous study²⁴. We also used the somatic mutation and bulk RNA-seq data of 94 ESCC patients that we published previously²⁰. Spatial transcriptomic data for analysis of RBM28 and immunosuppressive tumor microenvironment were retrieved from our previous study on ESCC¹⁹. Single-cell RNA-seq and bulk RNA-seq data of mouse esophageal multiple stage cancerous lesions and single-cell RNA-seq data of human esophageal multiple stage cancerous lesions were obtained from our previous studies^19,21,25. P53 ChIP-seq data of cancer cell lines were from the CistromeDB database (accession numbers: 75042, 90303, 90300, 88587, 82544, and 50361)³⁶. P53 ChIP-seq data of breast cancer cell lines were from GEO (GSE59176)³⁸.

The ESCC patient sets used for validating the effect of RBM28 expression level on response to ICB therapy were obtained from published studies^28,47,48. The data for analyzing the correlation between RBM28 levels in ESCC and patient overall survival time were obtained from published study^25,49. To validate the effect of RBM28 overexpression on immunotherapy of pan-cancers, multiple data sources were collected and used including melanoma^50–53, non-small cell lung cancer^54–56, head and neck squamous cell carcinoma⁵⁷, gastric cancer^58,59, colorectal cancer⁶⁰, and breast cancer⁶¹ (Supplementary Data 4). We also used TCGA pan-cancer data obtained from the TIMER2.0 website (http://timer.cistrome.org) to analyze gene expression, immune cell infiltration, and patient survival⁸⁹. The gene expression data in cell lines was obtained from the Achilles DepMap dataset (CCLE_DepMap_18Q3 release)⁹⁰. The clinical proteomic data was from CPTAC via LinkedOmicsKB⁹¹.

Statistical analysis

Statistical analysis was performed using R 4.3.1 and GraphPad Prism 10. Detailed statistical methods are described in the last of the figure legends or specific methods. The experimental data are shown as mean ± standard error of mean (S.E.M.) for at least 3 independent biological repeats. Data for WB and multiplex immunofluorescence images were from at least 3 biologically independent experiments with similar results. The trend test refers to the Jonckheere-Terpstra test used to assess ordered differences among ESCC stages. Patients’ survival differences were estimated using Kaplan-Meier method and the significances were examined by the log-rank test. The multivariate Cox proportional hazards model with patient’s age, sex, and tumor stage as covariates was used to compute the hazard ratio and 95% confidence interval. We used the R package ESTIMATE (version 1.0.13) to adjust tumor purity for assessing the RBM28 RNA level in tumor and the specific cut-offs derived from the surv_cutoff function implemented in the survminer R package (v 0.4.9) were used to dichotomize the RBM28 RNA level for survival analysis. All the box plots show median (central line) and 25th to 75th percentile (box edge) with 1.5x quantile range represented by whiskers. The horizontal lines in the violin plots represent mean values.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

D.L. and C.W. conceptualized, supervised, and acquired funding for the study. D.L. oversaw the project administration and compiled research resources. T.X., X.W., Q.L. and W.T. were engaged in sample collection and data curation. T.X., Z.H., S. Zhang, S. Zhu, and X.H. performed most proteomics analysis and functional experiments. T.X., J.Y., X.H., and H.Y. performed the bioinformatics analysis and data visualization. T.X. and Z.H. drafted and D.L. and C.W. reviewed and prepared the final manuscript. All authors approved the manuscript.

Peer review

Peer review information

Nature Communications thanks Ge Zhou and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This project was funded by the National Natural Science Foundation of China (81988101 to D.L. and C.W.) and the Medical and Health Technology Innovation Project of Chinese Academy of Medical Sciences (2021-I2M-1-013, 2022-I2M-2-003 and 2023-I2M-QJ-002 to D.L. and to C.W.).

Data availability

The RNA sequencing data generated in this study have been deposited in the GEO database under accession code GSE303217. The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium (https://proteomecentral.proteomexchange.org) via the iProX partner repository with the dataset identifier PXD066397. All relevant data produced in this study are provided in the Supplementary Information. Any additional information needed to reanalyze the data reported in this paper is available from the lead contact upon request. Source data are provided with this paper.

Code availability

No custom code was generated in this study. Code used for all processing and analysis is available upon request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76940-3.