Cancer Immunology Data Engine (CIDE) (https://cide.ccr.cancer.gov/) integrates 90 omics datasets encompassing 5,957 patients with immunotherapy outcomes from 17 solid tumors types (18). The detailed information on these datasets can be found in Supplementary Table 1. Outcome metrics include response binary (response versus non-response), Response Evaluation Criteria in Solid Tumors (RECIST), overall survival (OS), progression-free survival (PFS), and relapse-free survival (RFS). The association between gene expression and immunotherapy outcomes was evaluated using risk scores according to different types of outcomes. Cox proportional hazard (PH) regression, ordinary least squares (OLS), and two-sided Wilcoxon rank-sum test were implemented for survival outcomes (OS, PFS, and RFS), RECIST outcomes, and response binary outcomes, respectively. A curated list of 35 migrasome-related genes was retrieved from our previous study (19). This covers membrane markers (TSPAN1, TSPAN2, TSPAN3, TSPAN4, TSPAN5, TSPAN6, TSPAN7, TSPAN9, TSPAN13, TSPAN18, TSPAN25, TSPAN26, TSPAN27, ITGA1, ITGA3, ITGA5, ITGB1), protein markers (NDST1, EOGT, PIGK, CPQ), regulator molecules (ROCK1, TGFB2, IL1B, PDGFD, CXCL12, WNT8A, WNT11, MYDGF, BMP1, BMP7, CXCL18, WNT5B, LEFTY1, BMP2), comprehensively contributes to migrasome biogenesis. For migrasome-related covered in CIDE, their risk scores were computed, and the P-values of risk scores across cohorts were calculated using the two-sided Wilcoxon signed-rank test. The negative and positive risk score represent a favorable and adverse outcome, respectively. Genes were ranked by the absolute median risk score, with the top 2 (absolute median risk score > 0.5) selected.

TIMER3 (https://compbio.cn/timer3/) includes TCGA and large cancer cohorts (20). We analyzed the differential expressions of TSPAN6 and BMP1 across the tumor and normal tissues across the TCGA cohort. Deconvolution methods spanning 15 state-of-the-art algorithms were systemically used to estimate the infiltration levels of immune cells in tumor tissues. To ensure the robust of immune infiltration estimation, the correlation results were subjected to tumor purity adjustment. The relationships between infiltration levels of immune cells (including CD8⁺ T cells, CD4⁺ T cells, macrophages, myeloid-derived suppressor cells, dendritic cells, and NK cells) and TSPAN6 expression in tumor tissues were investigated using the partial Spearman’s correlation approach. The relationships between expression patterns of immune checkpoint genes and TSPAN6 in tumor tissues was also explored using the Spearman’s correlation approach.

The protein expressions of TSPAN6 and BMP1 across different types of cancers were explored using the human protein atlas (HPA). The immunohistochemistry (IHC) staining images mapping the protein expressions of TSPAN6 in liver cancer, lung cancer, and prostate cancer were acquired from the HPA (https://www.proteinatlas.org/).

ICRAFT (https://icraft.pku-genomics.org/#/homepage) contains a vast array of genome-scale CRISPR screens focusing on the immune-modulatory effects of genes (21). We explored the TSPAN6 and BMP1 knockout (KO) effects on cancer cells, co-cultured with immune cells (NK cells, CD8⁺ T cells, and CAR-T). The statistical evaluation was performed using permutation of ranks and adjusted by false discovery rate (FDR). Negative z-score normalized Log₂FC values indicate that gene KO will increase sensitivity of cancer cells to immune cell-mediated killing, while positive values indicate that it will decrease the sensitivity.

Tabula of the tumor immune microenvironment (TabulaTIME) (http://wanglab-compbio.cn/TabulaTIME/) integrates 103 tumor single-cell RNA-seq (scRNA-seq) studies, yielding a pan-cancer resource of 4,483,367 cells from 746 patients across 36 types of solid tumor (22). This framework encompasses data processing, MetaCell identification, lineage integration, and subtype characterization. Source code and merged data are available at the TabulaTIME server (http://wanglab-compbio.cn/TabulaTIME/). We obtained a pan-cancer TME atlas comprising 140,072 MetaCells annotated with 15 cell types and 56 cell subtypes. The cell types include CD8⁺ T cells, conventional CD4⁺ T cells, regulatory T cells (Treg), NK cells, proliferating T cells (Tprolif), B, plasma cells, monotypes/macrophages (Mono/macro), dendritic cells (DC), mast cells, endothelial cells, fibroblasts, myofibroblasts, epithelial cells, and malignant cells. The atlas was imported and visualized using the R packages Seurat (23) and plot1cell (24). Expression of TSPAN6 was mapped using the R package Nebulosa (25). Malignant cells were divided into TSPAN6-high and TSPAN6-low groups according to the median expression value of TSPAN6. Expression levels of CD274, NECTIN2, and LGALS9 between TSPAN6-high and TSPAN6-low malignant cells were assessed using the two-sided Wilcoxon signed-rank test.

The R package CellChat (26) was employed to quantify cell-cell communication probabilities between malignant cells and exhaustion T cells (Tex) based on ligand-receptor expression from scRNA-seq data. Significance of ligand-receptor interactions between two cell sub-populations was determined by permutation testing.

Virtual knockout of TSPAN6 in malignant cells using the R package scTenifoldKnk (27). The expression profiling of top 5000 highly variable genes from malignant cells was extracted for constructing gene regulatory network. TSPAN6 was then virtually knockout. The parameters were set as: qc_mtThreshold =0.1, qc_minLSize =500, nc_nNet =10, nc_nCells =200, nc_nComp =3. Differentially regulated genes were then identified before and after virtual TSPAN6 knockout according to the Z-score and FDR, and KEGG enrichment analysis was conducted using the R package clusterProfiler.

Spatial Transcriptomic (ST) data from 7 hepatocellular carcinoma patients with anti-PD-1 treatment, including 4 responders and 3 non-responders, was retrieved from the Gene Expression Omnibus (GEO) database under accession GSE238264 (28). The R package Spatial Cellular Estimator for Tumors (SpaCET) was used to infer cell identities (29). Spots with < 100 genes detected were excluded. Fraction of malignant, immune, and stromal lineages was deconvolved using the reference single-cell transcriptomic of hepatocellular carcinoma. We then mapped the spatial composition of TSPAN6 in malignant cells. The differential expression levels of TSPAN6 between responders and non-responders were assessed by the two-sided Wilcoxon signed-rank test.

A total of 44 tumor patients who were diagnosed pathologically at the First Affiliated Hospital of Naval Medical University were included in this study. The blood samples used were the remaining discarded samples from the routine clinical tests of patients. The information of enrolled patients, including age, gender, tumor type, stage, immunotherapy regimen, treatment initiation time, collection time after treatment and duration (days), was extracted and recorded from the electronic medical record system, as shown in the Supplementary Table 2. Blood samples were collected following the standard venous blood collection procedure. Anticoagulant blood collection tubes containing EDTA were used, and approximately 5 mL of peripheral venous blood was collected from each patient. After blood collection, the samples were kept at room temperature (25 °C) for no more than 30 minutes, and then centrifuged to obtain plasma. Subsequently, plasma samples without hemolysis, lipemia, or jaundice were selected for study, and all samples were completely anonymized. This study was reviewed and approved by the Ethics Committee of the First Affiliated Hospital of Naval Medical University (CHEC2025-198.).

Blood samples were centrifuged within 30 minutes after collection. Centrifuge tubes at 1500g at room temperature for 10 minutes. After centrifugation, carefully aspirate the upper-layer plasma, avoiding the middle white blood cell layer and lower red blood cell layer. Immediately test the plasma or store at -80 °C until ELISA initiation.

TSPAN6 was detected using the Thermo Fisher Varioskan LUX microplate reader and double-antibody sandwich ELISA (Cusabio), strictly following kit instructions. Standards and samples were added to the wells of a 96-well plate. The plate was incubated at 37 °C for 2 hours. Subsequently, the plate was washed 3 times using a washer. Biotinylated detection antibody was added to the wells and incubated at 37 °C for 1 hour, followed by washing. HRP-streptavidin was added and incubated at 37 °C in the dark for 1 hour, followed by another wash. TMB substrate was added and allowed to react at 37 °C in the dark for 15 to 30 minutes. Finally, the stop solution was added, and the absorbance (OD) values of each well were measured at a primary wavelength of 450 nm using the microplate reader. TSPAN6 concentrations (pg/mL) were calculated using the standard curve. All samples were tested in triplicate to ensure repeatability and reliability. The differences in TSPAN6 concentration between pre-immunotherapy and post-immunotherapy were assessed using the paired Wilcoxon signed-rank test.

A total of 1615 FDA-approved drugs were downloaded from the ZINC database (https://zinc.docking.org/) (30). The protein structure of TSPAN6 was predicted using the AlphaFold3 tool (https://alphafoldserver.com/). BatchVinaGUI was used to perform virtual screening to batch docking the 1615 FDA-approved drugs and TSPAN6 (31). The top 5 drugs including dantrolene, mitoxantrone, nitisinone, rabeprazole, and gentamicin were selected based on high affinity. Among these five drugs, mitoxantrone was selected to dock with TSPAN6. the binding site between TSPAN6 and mitoxantrone was implemented using the CB-DOCK2 (https://cadd.labshare.cn/cb-dock2/) (32).

We analyzed data from the Cancer Immunology Data Engine (CIDE), a newly released resource that aggregates 90 omics datasets from 5,957 patients across 17 solid tumor types, all with associated immunotherapy outcomes (7). For 35 migrasome-related genes covered in CIDE, we calculated risk scores to quantify the association between their expression levels and immunotherapy outcomes. As depicted in Figure 2A, these 35 migrasome-related genes were classified based on median risk scores, with scores > 0.5 indicating adverse outcomes and scores < 0.5 indicating favorable outcomes. To identify migrasome-associated targets for immunotherapy sensitization, we overlapped 7,922 proteins upregulated in any CPTAC dataset with migrasome-related genes significantly associated with adverse outcomes (risk score > 0.5). BMP1 and TSPAN6 were thus identified, as shown in Figure 2B. We then investigated the prognostic association of BMP1 and TSPAN6 across the TCGA, ICGC, and CPTAC databases (Figure 2C). Consistently, higher expression of BPM1 and TSPAN6 was associated with adverse prognosis (risk > 0.5). Furthermore, we noted upregulation of TSPAN6 in approximately half of the tumors within the TCGA database (Figure 2D; Supplementary Figure 1). Using the ICRAFT database, we investigated the knockout effect of TSPAN6 and BMP1 in malignant cells co-cultured with immune cells. As demonstrated in Figure 2E and Figure 2F, the negative Z-score normalized log₂FC values observed for TSPAN6 and BMP1 in different malignant cells upon co-culture with immune cells suggested increased sensitivity of cancer cells to immune cell-mediated killing. Therefore, ablation of TSPAN6 or BMP1 could increase anti-tumor immunity. We also explored the protein expression of TSPAN6 and BMP1 in cancers using the HPA database. Interestingly, moderate to strong expression of TSPAN6 was observed in the majority of malignant cells, whereas BMP1 expression was limited (Supplementary Figure 2). Furthermore, upregulated expression patterns of TSPAN6 were observed in liver cancer (Figure 2G), lung cancer (Figure 2H), and prostate cancer (Figure 2I). Altogether, these results indicate that TSPAN6 may serve as a migrasome target linked to adverse immunotherapy outcomes, suggesting its potential as a target for immunotherapy sensitization.

Given the potential of TSPAN6 as a target for immunotherapy sensitization, we explored whether TSPAN6 expression is associated with the tumor microenvironment (TIME). Employing the TIMER3, an integrative computational tool comprising 15 state-of-the-art algorithms, we estimated the infiltration levels of immune cells within the TCGA tumors. Since immune cells are critical determinants of cancer immunotherapy response, we systematically evaluated associations between TSPAN6 expression and the infiltration levels of key immune cells (including CD8⁺ T cells, CD4⁺ T cells, macrophages, myeloid-derived suppressor cells, dendritic cells, and NK cells). Interestingly, we noted that TSPAN6 expression was significantly inversely correlated with higher infiltration levels of immune cells in most tumor types (Figures 3A–G). Furthermore, we explored the expression patterns of immune checkpoints (including CD274, PDCD1, LGALS9, HAVCR2, NECTIN2, TIGIT) and TSPAN6. Consistent with the TIME exclusion pattern, TSPAN6 expression was significantly positively correlated with elevated expression of these immune checkpoints (Figure 3H). All correlation results are summarized in Supplementary Table 3. Collectively, these data indicate that TSPAN6 expression negatively associates with immune infiltration and positively associates with checkpoint expression, suggesting its modulation of immune components within the TIME at a pan-cancer level.

To investigate the mechanistic role of TSPAN6 in modulating the TIME, we employed the Tabula of the tumor immune microenvironment (TabulaTIME). This novel framework integrates a vast collection of scRNA-seq datasets, generating a pan-cancer atlas of 4,483,367 cells from 36 cancer types, clustered into 140,072 MetaCells representing 15 well-annotated cell types (Figure 4A). Mapping TSPAN6 expression density across this atlas revealed its highest enrichment in malignant cells (Figure 4B), with significantly elevated expression levels and cellular prevalence compared to other populations (Figure 4C). After stratifying malignant cells into TSPAN6-high and TSPAN6-low groups based on median expression, we assessed differential immune checkpoint expression. As shown in Figure 4D, key immune checkpoints (CD274, NECTIN2, and LGALS9) were significantly up-regulated in TSPAN6-high malignant cells. Intriguingly, virtual knockout (KO) of TSPAN6 in malignant cells identified differentially expressed genes (Supplementary Table 4; ranked by Z-score and FDR). KEGG enrichment analysis of the top 20 deregulated genes implicated perturbation of proliferation pathways (including p53 signaling and cell cycle) post-KO (Figure 4E). Subsequently, we used the CellChat to analyze the immune checkpoint interactions (CD274-PDCD1, LGALS9-HAVCR2, and NECTIN2-TIGIT) between malignant cells and exhausted T cells (Tex). Strikingly, the strength of all analyzed immune checkpoint interactions was enhanced between TSPAN6-high malignant cells and Tex relative to TSPAN6-low counterparts (Figure 4F). Collectively, these data demonstrate that elevated TSPAN6 expression reinforces immunotherapy resistance primarily by upregulating immune checkpoints (Figure 4G), thereby promoting T-cell evasion within the TIME.

In the aforementioned analyses, we suggest the role of TSPAN6 in modulating malignant resistance at a single-cell resolution. To further investigate the spatial distribution of TSPAN6 in TIME, we analyzed the spatial transcriptomics (ST) datasets from seven hepatocellular carcinoma patients with anti-PD-1 treatment, including four responders and three non-responders. Applying the SpaCET pipeline, we evaluated the fraction of malignant cells, stromal cells, and immune cells. We also investigated the expression distribution of TSPAN6. Herein, we presented the TSPAN6 expressions and malignant fractions in each patient from responder group (Figure 5A) and non-responder group (Figure 5B). Notably, the distributions of TSPAN6 expression were mainly enriched in regions with higher malignant fractions. This data spatially validated the higher expressions of TSPAN6 in malignant cells in the scRNA-seq datasets. We further calculated CD8⁺ T cell signature scores, which were significantly elevated in responders (Figure 5C). Conversely, TSPAN6 expression was markedly reduced in responders relative to non-responders (Figure 5D). Collectively, this spatial analysis confirms TSPAN6 downregulation in immunotherapy responders and its tumor core enrichment.

To validate TSPAN6 down-regulation following immunotherapy, we analyzed paired serum samples from 44 cancer patients obtained pre- and post-treatment. Analysis of our in-house cohort revealed a significant decrease in TSPAN6 levels post-immunotherapy (Figure 6A), supporting its potential as a non-invasive response biomarker. Given the functional role of TSPAN6 in immunotherapy resistance, we virtually screened for clinically approved inhibitors using molecular docking. After curating 1,615 FDA-approved drugs and the TSPAN6 protein structure, we computed binding affinities using the BatchVinaGUI software (Supplementary Table 5). The list of 1,615 FDA-approved drugs can be found in Supplementary Table 6. The top five compounds, including dantrolene, mitoxantrone, nitisinone, rabeprazole, and gentamicin, showed strong binding to TSPAN6 with affinities < -10 kcal/mol (Figure 6B). Notably, we prioritized mitoxantrone—an established topoisomerase II inhibitor with clinical antitumor activity—for detailed analysis. High-resolution docking further revealed its optimal binding pose with TSPAN6 (Figure 6C, left), stabilized by hydrogen bonds, hydrophobic interactions, and ionic interactions (Figure 6C, right).