Molecular classification identifies aggressive gastrointestinal stromal tumor subtype targetable by PARP inhibitors

Abstract

Purpose

Gastrointestinal stromal tumors (GISTs), the most common gastrointestinal sarcomas, experience a high rate of resistance to standard treatment tyrosine kinase inhibitor (TKI) imatinib, leading to tumor progression. Identifying non-responders and offering precise alternatives are crucial, however, multi-omics studies of GIST are limited, which hinders GIST biology understanding and novel drug development.

Methods

We performed whole-exome and transcriptomic sequencing on 106 primary GIST patients. Subtype-specific molecular features and therapeutic responses were further validated in vitro and in vivo.

Results

Unsupervised clustering analysis of RNA data classified them into four subtypes, including immune (G1), stromal (G2), proliferative (G3), and metabolic (G4) subtypes. The G4 subtype had more patients with primary non-gastric tumors, a poor response to neoadjuvant imatinib treatment and synchronous metastasis, with an overall poor prognosis. Notably, G4 subtype exhibited significantly enhanced aerobic metabolism. Moreover, G4 subtype harbored homologous recombination repair deficiency (HRD) signatures, with significantly reduced homologous recombination activity and upregulation of non-homologous repair pathways. By stratifying GIST cell lines into subtypes, we confirmed that the HRD-targeting poly ADP-ribose polymerase (PARP) inhibitor Olaparib suppressed tumor growth in the G4 subtype-representative GIST882 cell line. In vivo experiments demonstrated that Olaparib monotherapy significantly reduced tumor burden in GIST882 xenograft-bearing mice and synergized with imatinib.

Conclusion

We systematically deconstructed the molecular subtypes of primary GISTs by integrated genomic and transcriptomic analysis. A specific GIST subtype characterized by poor treatment responses and prognosis, marked by the activation of aerobic metabolism and HRD features, may be a potential candidate for PARP inhibitors.

Graphical Abstract

Graphical summary. Primary GISTs were classified into four subtypes (G1–G4) with distinct clinical, transcriptomic, and genomic features. The G4 subtype was associated with inferior response to imatinib and poorer prognosis, aerobic metabolism, high genomic instability, and HRD signatures. PARP inhibitors selectively targeted G4 tumors, offering a potential therapeutic alternative

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s13402-026-01247-w.

Introduction

Gastrointestinal stromal tumor (GIST) is one of the most common soft tissue sarcomas, with an annual incidence of approximately 6–22 cases per million [1]. Surgery is the standard treatment for primary GIST [2]. However, postoperative disease progression remains a major concern, with a 5-year disease-specific survival rate of only 54% [3].

A key characteristic of GIST is the presence of KIT and PDGFRA driver mutations, found in approximately 70% and 15% of cases, respectively, enabling the widespread use of targeted tyrosine kinase inhibitors (TKIs) [1]. Among these, imatinib is the standard first-line therapy for neoadjuvant, adjuvant, and advanced-stage treatment. However, secondary resistance to imatinib develops in approximately 80% of patients within 2–3 years, resulting in tumor progression [4]. Although later-line therapies such as sunitinib, regorafenib, and ripretinib can partially overcome resistance caused by secondary KIT exon 13/14/17/18 mutations, their efficacy remains limited [4–6]. Therefore, identifying patients with poor response to imatinib and unfavorable prognosis and further developing alternative precise pharmacological therapy, is essential for advancing GIST treatment. A previous proteomic analysis identified a poor-prognosis subtype characterized by SRSF3 upregulation, though its association with imatinib resistance was not directly explored [7]. Xie F. et al. proposed a GIST classification based on transcriptomic data, but it did not help identify patients with a potentially poor response to imatinib [8]. Additionally, single-cell sequencing revealed an enrichment of Treg cells in imatinib-resistant tumors, where the TIGIT-NECTIN2 interaction between immunosuppressive cells and tumor cells was associated with resistance [9]. Yet, actionable management of immunosuppressive imatinib-resistant tumors remains undefined. Several clinical trials have investigated immunotherapies and targeted treatments other than TKIs. Nevertheless, alternative therapies such as immune checkpoint inhibitors or MEK/PI3K inhibitors have shown limited efficacy for GIST [10–13]. The limited success of these therapies in GIST further emphasizes the importance of identifying GIST subtypes and applying precise treatments. However, a comprehensive understanding of the molecular features of GIST remains lacking, which hinders the implementation of effective clinical management strategy.

Omics technologies are critical for exploring the molecular characteristics of tumors. Currently, whole-exome sequencing (WES) has made progress in identifying additional driver mutations and copy number alterations, such as DIAPH and PI3K pathway genes, which are found to be associated with poor prognosis in GIST [14, 15]. Bulk RNA-seq analyses of different mutation types in GIST have shown that tumors with PDGFRA mutations exhibit higher immune cell infiltration than those with KIT mutations [16]. Transcriptomic studies of TKI-treated GISTs have further revealed enrichment of immunosuppressive cells, including Tregs and M2, in tumors progression after imatinib treatment [17]. Despite these insights, comprehensive multi-omics profiling of primary GIST remains limited, and the lack of a thorough understanding of its characteristics continues to constrain the development and application of precision therapies.

This study performed RNA-seq and WES on primary GIST samples from 106 patients who underwent surgical resection. Unsupervised clustering identified four subtypes with distinct intrinsic molecular characteristics. We identified that the subtype with the worst clinical prognosis and the poorest response to imatinib exhibited homologous recombination repair deficiency (HRD). Further in vitro and in vivo experiments demonstrated that the poly ADP-ribose polymerase (PARP) inhibitor effectively suppressed tumor proliferation and reduced tumor burden in this subtype, providing the theoretical foundation for its practice in GIST.

Results

Unsupervised transcriptomic clustering identifies four subtypes of primary GISTs

This study included 106 primary GIST samples, all with bulk RNA sequencing data and 94 of which had paired WES data. Clinical information of 106 patients was presented, including NIH risk factors, the use of neoadjuvant imatinib treatment and its effectiveness, synchronous metastasis status and survival data (Fig. 1A, Supplementary Table S1).

Fig. 1.

Transcriptomic clustering and associated clinical features of primary GISTs. (A) Schematic workflow of data collection, data analysis, and validation in this study. (B) Integrated plots of unsupervised clustering analysis based on transcriptomic data in 106 primary GIST patients with corresponding clinical information. The 106 GIST cases were classified into four subtypes: G1 (immune), G2 (stromal), G3 (proliferative), and G4 (metabolic). Clinical information of patients is shown above (upper panel). Heatmap showing top differentially expressed genes in each subtype (lower panel). The bar plot illustrating the results of GSEA pathway enrichment analysis for each subtype, with different colors representing the representative pathways enriched in G1, G2, G3, and G4 subtypes (right panel). Abbreviations: LL, very-low-risk; L: low-risk; M: intermediate-risk; H: high-risk; neg, negative; pos, positive; PD, progressive disease; PR, partial response; SD, stable disease. Statistical significance is calculated by Chi-square test. (C) Kaplan–Meier curves of postoperative progression-free survival in GIST patients, comparing the G4 subtype with the other three. p value is calculated by log-rank test. (D) Principal component analysis (PCA) showing transcriptomic distribution of the external validation and experimental cohorts. (E) Box plots showing the enrichment of signatures of the four subtypes in the external validation cohort. p value is calculated by wilcoxon test. *p < 0.05, **p < 0.01, ***p < 0.001

To uncover the molecular characteristics of GISTs, we first performed unsupervised clustering of RNA-seq data from 106 primary GISTs and identified four transcriptional clusters with distinct signatures (Fig. 1B). The first group (G1) was enriched in immune/inflammatory response-related pathways and was identified as the “immune subtype”. The second group (G2), in addition to immune-related pathway enrichment, was particularly characterized by the enrichment of stromal signatures such as epithelial mesenchymal transition and was defined as the “stromal subtype”. Proliferative-associated signatures including pathways associated with DNA replication were enriched in the third group (G3), which was identified as the “proliferative subtype”. Meanwhile, the fourth group (G4) was enriched with metabolism-related pathways and was defined as the “metabolic subtype” (Fig. 1B).

We further explored the differences in the clinical features among these four subtypes. For the NIH risk assessment and its specific criteria, no significant differences were observed in tumor size and mitotic count among the four subtypes, while there were more primary tumors located in non-gastric sites in the G4 subtype and fewer high-risk tumors in the G3 subtype (Fig. 1B, Fig. S1 A-D). The G4 subtype also included more patients with synchronous metastases (Fig. 1B, Fig. S1E). Sixty-three of the 106 patients underwent neoadjuvant imatinib therapy, and the G3 subtype included more patients who did not receive neoadjuvant treatment. The G1 and G2 subtypes comprised a higher proportion of patients with PR and SD to imatinib therapy. In contrast, G4 had relatively more patients with ineffective responses to imatinib (PD) (Fig. 1B, Fig. S1F). Consistently, G4 subtype also had a worse prognosis, with shorter post-surgery progression-free survival (PFS) and a higher rate of tumor progression compared to the other three subtypes (Fig. 1B-C, Fig. S1G). Moreover, multivariable Cox regression confirmed that G4 was a significant independent prognostic factor (HR = 5.01; 95% CI: 1.42–17.71; p = 0.012) after adjusting for potential confounders including risk classification and treatment effect (Fig. S1H). By analyzing 15 subsequent surgical samples from three patients initially classified as non-G4 subtypes who experienced tumor progression after TKI treatment, we found that most of these samples shifted toward the G4 subtype, as predicted by the glmnet elastic net model [18] (Fig. S1I). The identification of G4 as a subtype that integrates aggressive clinical features demonstrates the robustness of our transcriptomic classification and points to the need for further investigation of its molecular characteristics.

To further validate the reproducibility of our clustering results, we included an external cohort (Xie F. et al. [8]) with primary GISTs. Considering that the majority of the external cohort consisted of untreated GISTs, we included 77 untreated primary tumors in this cohort for validation analysis. As expected, principle component analysis (PCA) revealed that the external cohort clustered more closely with the G3 subtype, which had the highest proportion of untreated GISTs in our cohort (Fig. 1D). Concordantly, the glmnet elastic net model also confirmed that this cohort had the highest similarity to the G3 subtype (Fig. S1J). Moreover, pathway enrichment analysis using the signature gene sets of the four subtypes (Supplementary Table S2) showed that the G3 signature ranked significantly higher in the external cohort (Fig. 1E). Altogether, the above findings confirm that our transcriptomic subtyping is reliable.

Enhanced aerobic metabolic features in the G4 subtype

To further characterize the metabolic characteristics of the G4 subtype, we performed GSEA using metabolism-related pathways from KEGG and HALLMARK databases. The results showed that aerobic metabolic pathways, including the citric acid cycle (TCA cycle) and oxidative phosphorylation (OXPHOS), were most significantly enriched in G4 subtype (Fig. 2A). We also utilized the Compass algorithm [19] to characterize the metabolic states of individual samples. Notably, every metabolic reaction within the TCA cycle and OXPHOS was upregulated in the G4 subtype (Fig. 2B). Next, we intended to explore the correlation between transcriptomic features and clinical manifestations of the G4 subtype. Stratifying all samples based on GSVA enrichment results, we found that the subgroup with higher enrichment scores of TCA cycle and OXPHOS had shorter PFS time (Fig. 2C-D), while higher enrichment scores of propanoate metabolism or valine/leucine/isoleucine degradation were not positively correlated with poorer prognosis (Fig. S2A-B). Additionally, the activity of the aerobic metabolic pathways was correlated with the worse imatinib response (Fig. 2E).

Fig. 2.

Metabolic Features of the G4 GIST Subtype. (A) Bar plot showing GSEA enrichment of KEGG and HALLMARK metabolic pathways in G4 compared to other subtypes. (B) Volcano plots illustrating COMPASS-predicted differences in metabolic pathway activity between the G4 subtype and others, with the x-axis reflecting the effect size of metabolic activity (Cohen’s d) and the y-axis reflecting -log₁₀ adjusted p-values. (C-D) Progression-free survival Kaplan–Meier survival curves stratified by TCA cycle enrichment scores (C) and oxidative phosphorylation enrichment scores (D). p value is calculated by log-rank test. (E) Heatmap showing the correlation between G4-enriched metabolic pathways and the efficacy of neoadjuvant imatinib treatment. p values were determined by the Spearman’s correlation test. (F) Representative IHC staining images (left) and quantification of positive cell proportions (right) for the glycolysis marker MCT4 and the oxidative phosphorylation marker NDUFS1 across different subtypes. p value is calculated by t-test. Scale bar: 50 μm. (G) Heatmap of NTP classification for three GIST cell lines (GIST430, GIST-T1, and GIST882), with molecular subtypes indicated above and cell line–specific signature genes annotated on the right. (H) Western blot assays and quantification of NDUFS1 and MCT4 protein expression in three GIST cell lines. p value is calculated by t-test. (I) Cellular metabolism assays showing oxygen consumption rate (OCR), ATP synthesis, maximal respiration, and spare respiratory capacity in the GIST430, GIST-T1, and GIST882 cell lines. p value is calculated by t-test. *p < 0.05, **p < 0.01, ***p < 0.001

To validate these findings at protein level, we examined the expression levels of related proteins and found that monocarboxylate transporter 4 (MCT4) was significantly downregulated and NADH: ubiquinone oxidoreductase core subunit S1 (NDUFS1) was significantly upregulated in G4, which was confirmed by immunohistochemical (IHC) staining (Fig. 2F). MCT4 is a transmembrane protein primarily responsible for the transport of lactate and pyruvate across membranes, which is recognized as a symbol of anaerobic respiration [20, 21]. In contrast, NDUFS1, a core subunit of mitochondrial complex I and involved in redox reactions within the electron transport chain, typically reflects the activity of aerobic respiration [22, 23]. IHC results demonstrated a higher proportion of NDUFS1 + GIST cells and a lower proportion of MCT4 + GIST cells in G4 samples compared to other subtypes (Fig. 2F).

Subsequently, we tried to confirm our findings regarding the G4 subtype in a more homogenous experimental condition, and thus we validated our results in GIST cell lines. First, based on the different signature genes of the above-mentioned transcriptomic subtypes (Supplementary Table S2), we applied the Nearest Template Prediction (NTP) method to classify GIST430, GIST-T1, and GIST882 cell lines using their transcriptome sequencing data (GSE95864). It showed that GIST430, GIST-T1, and GIST882 could be classified as subtypes G2, G3 and G4, respectively (Fig. 2G, Supplementary Table S3). Typical subtyping genes were highly expressed in certain cell lines (Fig. 2G). GSEA also confirmed that GIST430 was enriched in immune-related pathways, GIST-T1 in proliferative pathways, and GIST882 in aerobic metabolic pathways, correspondingly (Fig. S2C-D). Consistently, western blot demonstrated that the G4 subtype cell line GIST882 expressed significantly higher levels of NDUFS1 and lower levels of MCT4 (Fig. 2H). To further clarify these findings, a Seahorse assay was conducted on these cell lines, which showed that GIST882 exhibited the highest oxygen consumption rate (OCR), the most robust ATP production and the greatest respiratory potential among all GIST cell lines tested, indicating that the G4 cell line had higher mitochondrial respiratory capacity (Fig. 2I). Collectively, we confirmed that G4 subtype was characterized by enhanced aerobic metabolism in vitro and in patient samples.

G1 subtype exhibits immune activation while G2 subtype enriches stromal features

Subsequently, we also explored the characteristics of the other subtypes in addition to G4. In G3, IHC confirmed high MKI67 expression, consistent with its proliferative signature (Fig.S3A). While both G1 and G2 subtypes were enriched in immune-related pathways, we initially analyzed the immune composition in different subtypes using MCP-counter and Xcell. The results of both algorithms showed a higher level of immune cell infiltration in the G1 and G2 subtypes than in the G3/G4 subtypes (Fig. S3B-C). Furthermore, we observed that G2 exhibited elevated fibroblast proportions alongside immune cells, consistent with the enrichment of stromal-related signatures (Fig. S3B-C). Considering the similar immune pathway enrichment between these two subtypes but different transcriptomic subtyping results, we next used various well-established gene sets to assess the differences in the immune microenvironment between G1 and G2. Overall, G1 exhibited immune activation, characterized by increased T cell chemotaxis, activation, and cytotoxicity, while G2 displayed a more immunosuppressive state, marked by the relative enrichment of tumor stromal genes, TGF-β response and other related pathways (Fig. S3D-I). These findings were further characterized in 20 (five per subtype) patient-derived samples using multiplex immunohistochemistry (mIHC) staining. It showed that the G1 subtype had a higher infiltration of CD8⁺ T effector cells, whereas the G2 subtype exhibited increased infiltration of FOXP3⁺ T regulatory cells (Tregs) and CD11b⁺CD15⁺ myeloid-derived suppressor cells (MDSCs) (Fig. S3J-K). Aligning with the tumor microenvironment states, the G2 subtype had significantly shorter PFS time after the first postoperative year (Fig. S3L) and a relatively higher incidence of synchronous metastasis when compared to the G1 subtype (p = 0.0958). Taken together, both G1 and G2 subtypes exhibited increased immune infiltration, with G1 showing immune activation and G2 marked by immunosuppressive cells and stromal enrichment, which corresponded to its adverse clinical features and likely accounted for its separation from G1 in unsupervised clustering.

G4 subtype was characterized by genomic instability

To further investigate the mechanisms underlying GIST subtype classification, we analyzed WES data of the tumor and paired non-tumor samples from 94 patients to explore the link between genomic alterations and transcriptomic subtypes (Fig. 1A). The mean coverage depth of WES was 331× for tumor samples and 329× for adjacent non-tumor samples. A total of 7258 single nucleotide polymorphisms (SNPs) (range: 1–1370, averaging 77.2 per sample) and 104 indels (range: 0–11, averaging 1.1 per sample) were identified.

Consistent with prior studies, KIT was the most frequently mutated gene in our cohort (67.0%, 63/94), while PDGFRA mutations were rare (5%) (Fig. 3A). For the remaining KIT/PDGFRA wild-type GISTs, one case (85_T) was identified as SDH-deficient based on postoperative pathological SDHB staining. The others harbored driver gene mutations previously reported in GIST [24, 25], including NF1 (3%), TP53 (2%), PIK3CA (2%), MEN1 (4%), FGFR1 (1%), CHD4 (4%), and ARID1A (12%), while no mutations were detected in BRAF or RAS genes. Other frequently mutated genes included TTN (10%) and MUC16 (10%), etc. Previous studies have shown that KIT mutation sites are correlated with imatinib effectiveness. Specifically, mutations in exons 9 and 11 of KIT are associated with imatinib sensitivity, whereas mutations acquired in other exons are linked to imatinib resistance [1, 2]. The distribution of exon9/11 and other exons of KIT in KIT mutated samples across different subtypes and between imatinib PR/SD and non-responsive patients showed no significant differences, indicating that the known mechanism alone could not fully explain their diverse manifestations (Fig. 3B, Fig. S4A). We also observed that the tumor mutational burden (TMB) in our cohort was generally low, with a median of 0.13/Mb, in line with the findings of earlier studies [8] (Fig. 3A). Nonetheless, differences among subtypes were present, as G2 and G4 exhibited relatively higher TMB levels (Fig. 3C).

Fig. 3.

Genomic landscape and molecular alterations across GIST subtypes. (A) Genetic profiles and clinicopathologic features of GISTs in our cohort. Waterfall plot showing mutations in KIT, PDGFRA, and the other most frequently altered driver genes, with mutation frequencies displayed in the right panel. Bar plot indicating tumor mutation burden (TMB) for each sample in the upper panel. (B) Stacked bar chart showing differences in the proportions of KIT exon 9/11 mutations and other exon mutations across the subtypes. Statistical significance is calculated by Chi-square test. (C) Box plot showing differences in TMB levels across the subtypes. p value is calculated by wilcoxon test. (D) Landscape of copy number variations (CNVs) across the four subtypes. (E-F) Box plots showing differences in the number of CNV events (E) and the genomic instability index (GII; F) across the subtypes. p value is calculated by wilcoxon test. (G) Pie charts showing the incidence of HLA loss of heterozygosity (HLA-LOH) events in the entire cohort and stratified by subtype. Abbreviations: neg, negative; pos, positive. (H) Box plot indicating immunoediting scores across different subtypes. The dashed line represents an immunoediting score of 1. *p < 0.05, **p < 0.01, ***p < 0.001

In addition to alterations at mutational level, we also analyzed the characteristics of copy number variations (CNVs). In total, 4035 copy number amplification events and 10,101 deletion events were detected in our data. The high burden of CNV in GIST was in accordance with previous research [8]. Generally, the most frequent gains occurred in chromosomes 19q11 (46.8%) and 5q11.1 (45.7%), with the most frequent losses were observed in chromosomes 1q36.32 and 14q32.33 (55.3%), some of which were validated in former studies [15, 26] (Fig. S4B-C). We quantified the number of total CNV events and found that subtypes G2 and G4 had a higher CNV burden, especially G4 subtype (Fig. 3D-E). Furthermore, by calculating the Genomic Instability Index (GII) based on the gene-specific copy number values across samples, we found an increase in GII for the G2 and G4 subtypes, with a more pronounced increase in G4 (Fig. 3F). We further analyzed specific cytoband alterations across subtypes and identified several changes that were more frequently enriched in the G4 subtype (Fig. S4E–H), including losses at 1p11.2, 1p36.32, and others, of which loss of 1p36 has previously been associated with shorter disease-free survival [15].

Genetic immune escape (GIE) is a crucial mechanism involving genetic alterations that enables tumor cells to evade elimination by immune system [27, 28]. In GIE, decreased neoantigen presentation or reduced neoantigen recognition compromises an important part [27, 29, 30]. Thus, we first analyzed the loss of heterozygosity (LOH) in class I human leukocyte antigen (HLA) locus, as HLA governs the presentation of neoepitopes for T cell recognition [29]. The total prevalence of HLA-LOH in our cohort was 7% (9/93), with most events occurring in the G1 and G2 subtypes (Fig. 3A and G). Notably, no LOH events were observed in the G3 subtype, while G1 had the highest prevalence (20%). In addition, we calculated the immunoediting score by analyzing the ratio of observed to expected neoantigens, where a value below 1 indicates the occurrence of immunoediting and subsequent neoantigen loss [30]. It showed that median score of G1 was markedly lower than 1, while that of G4 was above 1 (Fig. 3H). Taken together, immune escape was not a common phenomenon in GIST. The absence of HLA-LOH and an immunoediting score close to 1 in G3 suggested a potential association between imatinib treatment and immune escape. In G1, which exhibited an activated immune microenvironment, tumor cells were subjected to greater immune pressure and might therefore evolve immune evasion. While GIE analysis in the G4 subtype suggested that GIE might not account for its malignant phenotype, the previously observed mutation and CNV results, on the other hand, pointed to increased genomic instability, highlighting the need to further investigate genomic features associated with this in G4.

Homologous recombination deficiency and non-homologous recombination pathway activation in G4 subtype

Genetic alterations induced by environmental and endogenous sources of mutagenesis can leave distinct imprints in the genome, which can be detected as mutational signatures [31]. We used Palimpsest to de novo identify 6 single-base substitution (SBS) signatures, designated SBS_denovo_1 to SBS_denovo_6, which highly resembled the reported COSMIC signatures (Fig. 4A, Fig. S5A). Notably, SBS_denovo_1 was highly similar to the widely known SBS3 signature (cosine similarity: 70%), a hallmark of homologous recombination repair deficiency (HRD) (Fig. 4A). We recognized that the SBS_denovo_1 signature was enriched in G4 subtype, whereas its presence was relatively limited in G3 subtype (Fig. 4B). Considering that the G3 subtype included more patients without neoadjuvant therapy, we speculated that this signature might be linked to imatinib exposure. As expected, the prevalence of SBS_denovo_1 was significantly higher in imatinib-treated patients (Fig. 4C). We then compared the prevalence of SBS_denovo_1 in imatinib-treated samples across the four subtypes and included the external cohort for further validation. In treated patients, this signature was markedly enriched in G4, indicating that HRD played a key role in driving genetic alterations in G4 (Fig. 4D). Other signatures were also analyzed, among which SBS_denovo_2, corresponding to the COSMIC SBS5 signature, was enriched in G1 and G2 subtypes (Fig. S5B). SBS5 is a clock-like signature found in both tumor and normal cells, but its underlying cause remains unclear. SBS_denovo_6, representing SBS40a, was more frequently observed in the G3 and G4 subtypes, yet its proposed etiology remains unknown (Fig. S5F). Other signatures showed no significant differences among subtypes (Fig. S5C-E).

Fig. 4.

Genomic signatures and alterations in DNA damage repair pathways across GIST subtypes. (A) SBS spectrum illustrating SBS signatures (SBS_denovo_1 to SBS_denovo_6) identified through de novo decomposition of mutational features. Cosine similarity between each identified signature and known COSMIC signatures is indicated in parentheses. (B) Box plot showing differences in the distribution of the SBS_denovo_1 signature across the subtypes. p value is calculated by wilcoxon test. (C) Box plot showing differences in the distribution of the SBS_denovo_1 signature between patients with and without imatinib treatment. p value is calculated by wilcoxon test. (D) Box plot showing the distributional differences of the SBS_denovo_1 signature in post-imatinib patients across different subtypes and in the external validation cohort. p value is calculated by wilcoxon test. (E) Stacked bar chart showing differences in the distribution of CNV48E signature across the subtypes. Statistical significance is calculated by Chi-square test. (F) CNV spectrum illustrating the de novo identified CNV48E signature. Cosine similarity between the CNV48E signature and known CN17/CN19 signatures is shown in parentheses. (G) Schematic illustrating the main DNA double-strand break repair pathways, with embedded heatmaps showing expression levels of key genes involved in each pathway across the GIST subtypes. (H-J) Box plots showing differences in GSVA enrichment scores of the HR (H), NHEJ (I), and aEJ (J) pathways across the subtypes. p value is calculated by wilcoxon test. (K-M) Correlation analyses between enrichment scores of the HR (K), NHEJ (L), and aEJ (M) pathways and oxidative phosphorylation. Each dot represents one tumor (N = 106). p value is determined by the Spearman correlation test. ns, not significant, *p < 0.05, **p < 0.01, ***p < 0.001

Signatures at the mutational level hinted at HRD, prompting us to further explore its imprints at the CNV level [32]. SigProfilerExtractor de novo identified six signatures in the recommended solution, named CNV48A–F, respectively (Fig. 4F, Fig. S5G-I). CNV48E, which showed high similarity to CN17 and CN19 (cosine similarity: 60% and 75%, respectively), was enriched in the G4 subtype (Fig. 4E). CN17 has been established as a signature of HRD, while CN19 is associated with high scarHRD score, a measure of genomic scars of HRD [32, 33]. Based on the de novo genomic signatures identified in our data, we proposed that HRD might be a characteristic trait of G4.

Homologous recombination (HR) is one of the major repair pathways for DNA double-strand breaks (DSB). It utilizes a homologous DNA template to accurately repair the break, with key enzymes including BRCA1, BRCA2, RAD51, and others [34, 35]. When HRD occurs, cells rely on non-homologous (non-HR) pathways to repair DSBs, with non-homologous end joining (NHEJ) and alternative end joining (a-EJ) playing major roles [36]. Thus, we further examined the expression patterns and enrichment of key enzymes involved in HR and non-HR pathways [34, 37]. As the results illustrated, HR was highly enriched in the G3 subtype, while significantly less enriched in the G1, G2, and G4 subtypes, suggesting that HR damage occurred in these subtypes (Fig. 4G-H). However, unlike the G1 and G2 subtypes, non-HR pathways were notably activated in the G4 subtype (Fig. 4G and I-J). The lower repair fidelity of non-HR repair might be a potential source of genomic scars observed in the G4 subtype. Supporting this, both correlation and multivariable regression analyses showed that the genomic HRD signatures were significantly associated with aEJ pathway activation (p < 0.05), with a-EJ pathway identified as a significant independent predictor of the genomic features (p < 0.001, Fig. S5J-K).On the other hand, it was reasonable to hypothesize that due to the activation of non-HR pathways in G4, tumors under imatinib pressure would have a means to repair genome damage and ensure survival, leading to reduced drug efficacy and poorer clinical prognosis.

We also explored the potential relationship between the genetic characteristics of G4 and its transcriptomic features. It has been reported that tumor cells rely on OXPHOS to supply NAD + and ATP, which are essential for non-HR repair mediated by poly(ADP-ribose) polymerase (PARP) [38, 39]. Consistent with expectations, HR enrichment was negatively correlated with the enrichment of aerobic metabolic pathways, whereas non-HR pathways showed a positive association with aerobic metabolism (Fig. 4K-M, Fig. S5L). Taken together, HRD-related signatures were enriched in the G4 subtype, with downregulation of the HR pathway and upregulation of non-HR pathways, potentially contributing to its aggressive clinical behavior.

PARP inhibitor olaparib suppresses tumor growth in G4 subtype and synergized with imatinib

HRD in G4 implied an emerging therapeutic target for this subtype. PARP inhibitors, targeting HRD, have been approved for clinical use and demonstrated significant efficacy across various cancer types [40, 41]. It operates through synthetic lethality, where loss of either gene alone is non-lethal, but loss of both leads to cell death. In HRD tumor cells, the inability to repair DSBs via HR leads to reliance on non-HR pathways, with PARPs serving as key enzymes. Inhibition of PARPs disrupts these repair processes, resulting in the accumulation of unresolved DSBs, ultimately leading to chromosomal instability and cell death [42–45]. Yet, no study has explored the application of PARP inhibitors in GISTs [42, 46].

We first validated HR deficiency and non-HR pathway activation using IHC analysis of key biomarkers, as evidenced by decreased BRCA2 and increased PARP1 expression respectively (Fig. 5A). These findings were corroborated at the protein level in GIST cell lines, as evidenced by the concordant expression patterns of BRCA2, PARP1, and other key enzymes (Fig. 5B, Fig. S6A). Based on the validation of the GIST cell line phenotypes, we evaluated their drug sensitivity to the classical PARP inhibitor Olaparib and the standard GIST treatment imatinib via half-maximal inhibitory concentration (IC50) determination. GIST882 showed the lowest sensitivity to imatinib (IC50 = 4790nM) and was substantially sensitive to Olaparib (IC50 = 303.8 μm) (Fig. 5C-D).

Fig. 5.

Functional validation of PARP inhibitor efficacy in GIST in vitro and in vivo. (A) Representative IHC images (left) and corresponding quantification of positive cell proportions (right) for the HR marker BRCA2 and the non-HR marker PARP1 across different subtypes. p value is calculated by t-test. Scale bar: 50 μm. (B) Western blot assays and quantification of BRCA2 and PARP1 protein expression in three GIST cell lines. p value is calculated by t-test. (C-D) Dose–response curves of imatinib (C) and Olaparib (D) in three GIST cell lines. (E) Comparison of 8-OHdG levels measured by ELISA in three GIST cell lines treated with Olaparib at different concentrations (top) and in combination with imatinib (bottom; imatinib: 0.01 µmol, Olaparib: 0.5 µmol). p value is calculated by t-test. (F) Colony formation assay showing the effects of different drug treatments on the growth of three GIST cell lines. Representative colony images (left) and box plots comparing colony numbers (right) are shown. p value is calculated by t-test. (G) Mouse xenograft experiments. Schematic of xenograft establishment and treatment regimens (top). Representative images of tumors under each treatment condition (bottom). GIST882-derived xenograft models were treated with imatinib, Olaparib, or combination therapy (n = 4 per group). (H) Tumor growth curves of xenografts under different treatments. p value is calculated by two-way ANOVA test. (I) Tumor volumes of xenografts under different treatments. p value is calculated by t-test. (J) Representative IHC images (left) and quantification of MKI67-positive cells (right) in mouse xenografts. p value is calculated by t-test. Scale bar, 50 μm. ns, not significant, *p < 0.05, **p < 0.01, ***p < 0.001

Colony formation assay was performed to evaluate the antitumor efficacy of Olaparib in GIST in vitro. We found that Olaparib significantly inhibited the proliferation of GIST882 cells compared to that of the control at a concentration of 1 µM, whereas the other two cell lines showed no apparent inhibition (Fig. 5F). Moreover, at 0.005 µM, a concentration at which imatinib alone showed no inhibitory effect on any cell line, its combination with Olaparib significantly enhanced the suppression of GIST882 proliferation compared to imatinib monotherapy (Fig. 5F). Additionally, the inhibitory effect of Olaparib increased progressively with its concentration across all cell lines, suggesting a potentially promising antitumor effect of PARP inhibitors in GISTs (Fig. S6B-F). To validate that Olaparib inhibited GIST proliferation by interfering with the DNA repair process, we evaluated the level of DNA damage by measuring the concentration of 8-hydroxy-2’-deoxyguanosine (8-OHdG) in the supernatant of GIST cell lines following 7-day co-culture with various drugs. Expectedly, Olaparib monotherapy at low concentrations resulted in elevated 8-OHdG levels in GIST882. As the concentration increased, 8-OHdG levels increased across all cell lines, with a more pronounced increase in GIST882 (Fig. 5E). The combination of Olaparib and imatinib elevated 8-OHdG levels in both GIST-T1 and GIST882 cell lines compared to the control, whereas the increase was more significant in GIST882 cell line (Fig. 5E).

Based on in vitro experiments, we established a GIST882 subcutaneous xenograft model in NCG mice to investigate the antitumor effects of Olaparib alone and in combination with imatinib in vivo (Fig. 5G). After four weeks of drug treatment, we observed significant inhibition of tumor growth in the Olaparib treatment group and additional inhibition in the combination treatment group, indicating that Olaparib monotherapy showed comparable effects to imatinib, and the combination treatment achieved a markedly enhanced antitumor effect (Fig. 5G-I). In keeping with this, cell proliferation was significantly decreased in tumor tissues from mice treated with Olaparib and the combination (Fig. 5J). Furthermore, to evaluate the in vivo safety of the treatments, we monitored the body weight of the mice throughout the treatment period. No statistically significant differences were observed among the treatment groups or across different time points, indicating that the combination regimen was well tolerated (Fig. S7A).

Finally, we tried to screen for specific molecular alterations within non-HR pathways that could predict olaparib sensitivity. To achieve this, we developed an imatinib-resistant GIST430 cell line and evaluated its altered drug responsiveness using oncoPredict. The resistant cells showed higher sensitivity to olaparib compared to the imatinib sensitive controls (Fig. S7B). By taking the intersection of 3 gene sets: differentially expressed genes between the resistant and sensitive cells, olaparib-sensitivity-associated genes from oncoPredict [47], and known non-HR pathway genes, we identified XRCC5 as a key altered molecule (Fig. S7C). XRCC5 expression was significantly upregulated in the resistant cells (Fig. S7D). This specific change within non-HR pathways might be associated with the mechanistic vulnerability of the imatinib resistant tumors to PARP inhibition, and could serve as a potential biomarker for its sensitivity.

Cumulatively, these results suggest that the PARP inhibitor effectively inhibits the proliferation and tumor growth of G4 subtype GIST, highlighting its potential as a candidate strategy for GIST treatment.

Discussion

GIST is the most common mesenchymal tumor in the gastrointestinal tract, characterized by a consistent molecular feature of KIT/PDGFRA mutations [2]. However, despite this shared hallmark, GISTs exhibit significant heterogeneity in tumor behavior, indicating additional molecular complexities. Based on large-scale multi-omics characterization of 106 GISTs, we identified four transcriptomic subtypes defined by immune activation, stromal enrichment, proliferation, and aerobic metabolism signatures. The G1 subtype exhibited activation of the immune microenvironment, potentially reflecting the effectiveness of imatinib, however, the emergence of immuno-editing events under treatment pressure remains a concern. Maintaining neoantigen presentation and preserving the infiltration of effector immune cells may represent a potential strategy for sustaining the efficacy of imatinib in this subtype. While the G2 subtype demonstrated enrichment of stromal components, particularly cancer-associated fibroblasts (CAFs), which have been implicated in tumor migration and drug resistance [48]. Given the relatively higher metastasis rate and poorer long-term survival of G2 compared to G1, therapeutic strategies targeting CAFs may offer clinical benefits for this subtype. Meanwhile, the G3 subtype, composed predominantly of imatinib-untreated patients, was characterized by the enrichment of proliferative signatures and may therefore benefit from subsequent tyrosine kinase inhibitor (TKI) therapy. Of particular note, the G4 subtype, with the worst prognosis and poorest response to imatinib, exhibited upregulated aerobic metabolism. Moreover, HRD was identified as a novel genetic alteration in the G4 subtype, highlighting PARP inhibitors as a potential therapeutic strategy. The molecular signatures of G4 may serve as biomarkers for identifying GIST patients with unfavorable clinical outcomes. In line with this, ROC analysis showed that both a-Ej (AUC = 0.722) and TCA cycle (AUC = 0.718) pathways have potential in predicting imatinib response (Fig. S7E), and combining these two pathways would slightly improve the predictive accuracy (AUC = 0.765; Fig. S7E). On the other hand, to make G4 identification more practical for clinical use, we developed a simplified panel of three genes (ANKRD45, NUDT11, and VIM) that showed high reliability in our cohort (AUC = 0.905, Fig. S7F). This panel may help identify patients with unfavorable outcomes in the future, although further validation is required.

Different from the transcriptomic classification proposed by Xie F. et al., which was based mainly on untreated primary and metastatic tumors, our cohort included primary tumors that had received neoadjuvant imatinib therapy, thereby enabling better characterization of molecular signatures associated with imatinib-resistant subtype. In Xie F.’s study, the relatively more aggressive subtypes C2 and C3 were characterized by enrichment of immune-active and cell cycle-related gene sets, for which the authors suggested immunotherapy and CDK4/6 inhibitors as potential therapeutic options. In relation to these findings, our aggressive subtype G4 showed enrichment of oxidative phosphorylation, immune-desert features, and high genomic instability, suggesting that these may represent common hallmarks of aggressive GIST.

Studies have reported heterogeneous metabolic states in GIST cells. While one has suggested that imatinib-resistant cells shift from the TCA cycle to the pentose phosphate pathway with downregulated OXPHOS, another study has found that OXPHOS upregulation is associated with increased cell proliferation and growth in GIST [8, 49]. Our findings supported the diverse metabolic reprogramming of GIST cells. As in some other tumors in which upregulated OXPHOS is associated with TKI resistance and tumorigenesis, the metabolic abnormalities observed in the G4 subtype provide insights into potential metabolically targeted therapies for further investigation [50–53].

TKIs are currently the primary approach for GIST management, with imatinib being the most established treatment [54, 55]. However, TKIs developed based on KIT/PDGFRA mutations have yet to bring significant breakthroughs in GIST treatment, and the inevitable emergence of drug resistance underscores the urgency of advancing precision therapy [1]. Through in-depth analysis of genomic alterations in our subtypes, we identified an enrichment of HRD signatures in the G4 subtype that was most correlated with imatinib resistance. Integration with transcriptomic data revealed that these genomic scars reflect the activation of non-HR pathways and the dysfunction of HR repair. Research findings have revealed that alterations in HR pathway genes, such as BRCA2 and ATM, occur in GIST, with frequencies ranging from 3.7% to 9% [56, 57]. However, to the best of our knowledge, this is the first time that HRD features have been identified in GIST as being associated with a worse clinical phenotype. The HRD features prompted us to explore the potential use of PARP inhibitors in this GIST subtype.

Though PARP inhibitors have been proven safe and effective in various solid tumors, their application in GIST has not been investigated to date. Our study suggests that the PARP inhibitor Olaparib effectively controlled tumor proliferation and reduced tumor burden in the G4 subtype. Besides the G4 cell line, we also noticed in the colony formation assay that Olaparib exerted a modest anti-proliferative effect on the GIST430 and GIST-T1 cell lines. Notably, during past clinical trials, PARP inhibitors also demonstrated antitumor effects in tumors without HRD and have been approved for population-wide use in advanced ovarian cancer [58–60]. We suspect several factors may contribute to this phenomenon. In the G2 subtype cell line, GIST430, we observed downregulation of DNA damage response (DDR)-related proteins, which may lead to increased genomic instability under external stress. In this context, the addition of Olaparib could further exacerbate this instability by inhibiting both single- and double-strand DNA break repair [61]. Moreover, Olaparib treatment may lead to the downregulation of multiple DDR proteins that are upregulated in proliferative tumor cells to support rapid DNA replication [62, 63]. Broadly speaking, genomic instability beyond traditional HRD caused by oncogenic events may create opportunities for the application of PARP inhibitors [64, 65]. Overall, Olaparib exhibited antitumor effects in GIST, particularly in the subtype with HRD signatures.

We found that Olaparib showed a synergistic effect with imatinib both in vitro and in vivo. In fact, imatinib treatment has been reported to possess the ability to inhibit the key HR enzyme RAD51, thereby inducing HRD [66–68]. The combination of imatinib and Olaparib has been tested in ovarian cancer cells, where imatinib was found to enhance Olaparib-induced growth suppression [69]. In chronic myeloid leukemia, co-treatment with imatinib and a PARP inhibitor has also been shown to more effectively eradicate clonogenic cells than monotherapy [70]. Our findings demonstrated the value of combining standard GIST therapy with other targeted agents and offered new insight into the efficacy of PARP inhibitors in GIST therapy, although further studies are required to elucidate the underlying mechanisms.

Our study had some limitations. Due to the lack of public sequencing data for GIST, especially samples with complete treatment information, our findings still require validation in external cohorts. Additionally, the mechanistic relationship between the metabolic characteristics and HRD features of the G4 subtype remains to be explored in future studies.

In summary, our study provided a systematic and precise molecular classification of GISTs and identified a previously undefined subtype, G4, with poor prognosis and drug response characterized by enriched HRD features. Based on this, we provided the preliminary evidence for the potential role of PARP inhibitors in GIST management.

Materials and methods

Patients and specimens

A total of 106 patients with primary GIST who underwent surgical resection at The First Affiliated Hospital of Sun Yat-sen University (FAH-SYSU) between 2017 and 2022 were pathologically diagnosed and consecutively enrolled in this study. This study was centrally approved by the Ethical committee of the First Affiliated Hospital of Sun Yat-sen University (No. IIT-2023-840) and complied with all relevant ethical regulations. All patients provided written informed consent of the samples for research. The clinicopathological information was summarized in Supplementary Table S1. Two experienced pathologists confirmed the diagnosis through histology examination. All patients who received neoadjuvant therapy were treated exclusively with imatinib. Response to therapy was assessed using CT and measured using RECIST 1.1. In detail, bulk RNA sequencing was performed on GIST tissues from 106 patients preserved in liquid nitrogen and 15 formalin-fixed, paraffin-embedded (FFPE) samples. Liquid nitrogen-frozen GIST and peritumoral normal tissues from 94 patients were used for whole-exome sequencing. (Fig. 1A.)

Whole-exome sequencing (WES)

Whole-exome sequencing was conducted on 94 tumor samples and their matched non-tumor samples. Genomic DNA was extracted from frozen tissue using the DNeasy Blood & Tissue Kit (QIAGEN, Hilden, Germany) following the manufacturer’s protocol. The extracted DNA was then randomly fragmented into an average size of 180-280 bp by Covaris S220 ultrasonicator. DNA fragments then underwent end-repair, phosphorylation, A-tailing, and ligation with paired-end adapters at the 3′ ends (Illumina). Adapter-ligated DNA fragments were selectively enriched by PCR and hybridized with biotin-labeled probes to capture exonic regions. Captured libraries were further enriched to add sequencing indices and purified using the AMPure XP system (Beckman Coulter, Beverly, USA). After quality and quantity assessment, the libraries were sequenced on the Illumina NovaSeq 6000 platform (Illumina, San Diego, CA, USA) to generate paired-end 150 bp reads. Raw sequencing reads were subjected to quality control using FASTQ format standards and processed with Trim-Galore (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/) [71] to remove adapter sequences. Filtered reads were aligned to the human reference genome (GRCh38) using Burrows-Wheeler Aligner (BWA, v0.7.17; RRID: SCR_010910) [72] with default settings. Duplicate reads were subsequently marked and removed using Sambamba (v 0.7.0) [73].

Somatic mutation calling

Somatic variant calling was performed following the Genome Analysis Toolkit (GATK 4.1.2.0) Best Practices pipeline [74]. Specifically, MuTect2 was used to identify somatic mutations by comparing each tumor sample with its matched non-tumor counterpart. Variant annotation and filtering were conducted using public databases, including gnomAD, avSNP, ExAC, ICGC, COSMIC, ClinVar, InterVar, dbNSFP, and dbscSNV. All identified somatic mutations were subsequently annotated using ANNOVAR [75] for downstream analyses. Mutation patterns were further characterized and visualized using the coOncoplot function from the maftools R package [76], incorporating group classifications and clinical information. Additionally, the distributions of mutations and tumor mutational burden (TMB) were compared across different GIST subtypes.

HLA typing and loss-of-heterozygosity (LOH) identification

POLYSOLVER [77] was used to infer class I HLA alleles from each patient’s non-tumor sample. Subsequently, LOHHLA [29] was employed to detect loss of heterozygosity (LOH) events in MHC class I HLA genes in tumor samples, based on the inferred HLA typing with default settings.

Neoantigen prediction

Tumor neoantigens were predicted based on non-synonymous missense mutations identified from WES data. Amino acid substitutions were introduced into protein sequences to generate mutated peptide sequences. Neoantigen prediction was performed using NetMHCpan (v4.0) [78] using mutated peptides. Candidate neoantigens were selected according to the following criteria: (1) peptide length between 9 and 11 amino acids; (2) predicted binding affinity (IC50) < 500 nM; and (3) percentile rank < 0.5%.

To calculate the immune-editing score, we followed the method developed by Rooney et al. [30]. Briefly, the expected numbers of nonsynonymous mutations (Npred) and neoantigens (Bpred) for our GIST cohort were estimated based on TCGA pan-cancer data. The observed numbers of nonsynonymous mutations (Nobs) and neoantigens (Bobs) were obtained from our data. The immune-editing score (R) was then calculated as R = (Bobs/Nobs)/(Bpred/Npred), where a value of R < 1 indicates the presence of immune editing in tumors.

Copy number analysis

Somatic copy number profiles and tumor purity were estimated using Sequenza (v3.0.0) [79]. To identify focal genomic regions that were recurrently amplified or deleted across samples, copy number segments from individual patients were uploaded to the GenePattern public server and analyzed using the GISTIC 2.0 module [80], with amplification and deletion thresholds set at 0.3 and -0.3, respectively. Copy number variation (CNV) visualizations were generated using the MesKit R package [81]. For each sample, the total number of copy number gain (amplification) and loss (deletion) events was calculated.

To quantify genomic instability, the genomic instability index (GII) [82–84] was calculated based on segmentation files generated by Sequenza. GII was defined as the proportion of the genome affected by copy number alterations relative to the total genome size and was computed as the average fraction of altered regions across all 22 autosomal chromosomes.

Mutation signature and CNV signature analysis

Mutation signature analysis was conducted using Palimpsest (v2.0.0) [85]. Single base substitution (SBS) signatures were de novo extracted from the samples and compared to known mutational signatures in the Catalogue Of Somatic Mutations In Cancer (COSMIC) database (v3) [86] by calculating cosine similarity. The relative contribution of each de novo signature was then estimated for each sample. CNV signatures were de novo extracted using SigProfilerExtractor [87] with the recommended solution, which also output the relative contribution of each signature for individual samples. The similarity between de novo and known CN signatures was assessed using Palimpsest.

Bulk RNA sequencing

RNA sequencing (RNA-seq) was performed on 106 primary tumor samples and 15 tumor samples obtained from subsequent surgeries following imatinib treatment. Total mRNA was extracted from fresh frozen tissue for the primary samples and from formalin-fixed, paraffin-embedded (FFPE) tissue for the subsequent samples. The extracted RNA was fragmented into short fragments of appropriate length for complementary DNA (cDNA) synthesis. cDNA fragments of suitable size were then selected as templates for PCR amplification to construct sequencing libraries. Final libraries from primary samples were sequenced on the Illumina NovaSeq 6000 platform, while those from FFPE samples were sequenced on the Illumina NovaSeq X Plus platform, both using paired-end 150 bp sequencing.

Quality control of raw reads was performed using FastQC [88], and adaptor sequences were removed with Trim-Galore. Sequencing reads were aligned to the human reference genome (GRCh38) using STAR (2.7.9a) [89]. Gene expression quantification was conducted with RSEM (1.3.3, default parameters) [90], generating both read counts and fragments per kilobase of exon per million mapped reads (FPKM) for all genes and isoforms.

Unsupervised transcriptomic clustering

Molecular subtypes were identified by unsupervised hierarchical clustering of RNA expression profiles from 106 tumor samples. Gene filtering was performed based on the following criteria: (1) expression > 0 in more than 90% of samples; (2) expression > 1 in at least 50% of samples; and (3) a coefficient of variation (CV) > 0.8 across all samples. Unsupervised clustering was subsequently conducted using the consensus non-negative matrix factorization (NMF) algorithm based on the selected genes. Based on the cophenetic correlation coefficients and inter-group feature differences, a four-cluster solution was identified as the optimal NMF decomposition. To enhance clustering robustness, NMF clustering was repeated 100 times to generate a 106 × 106 consensus similarity matrix. K-means clustering (k = 4) was then applied to this matrix, yielding four stable transcriptomic subtypes. The representative genes for each subtype are listed in Supplementary Table S2.

Gene set enrichment analysis (GSEA) and gene set variation analysis (GSVA)

To explore differences in biological characteristics and processes among the four defined GIST subtypes and across various GIST cell lines (GSE95864), we performed gene set enrichment analysis (GSEA, clusterProfiler v4.8.2) [91] using gene sets from the Molecular Signatures Database (MSigDB v7.5.1; KEGG and Hallmark collections) [92].

GSVA was performed using the GSVA R package [93] to assess the activity of specific signatures in bulk RNA-seq samples. GSVA scores reflected the relative activity of each gene signature or pathway. The gene sets used in this study are listed in Supplementary Tables S4 and S5.

Cell type deconvolution prediction

To assess tumor microenvironment cell enrichment across different GIST subtypes, we applied the immune deconvolution tools xCell [94] and MCP-counter [95] to estimate immune and stromal cell populations from bulk RNA-seq data.

Identification of feature genes characterizing GIST subtypes identification of feature genes for GIST subtypes and phenotypic prediction for GIST cell lines

To identify feature genes for each GIST subtype, we used the random forest algorithm from the randomForest R package [96]. Differentially expressed genes (DEGs) specific to each subtype were used as input features for model training. Genes with higher mean decrease accuracy were selected as subtype-specific feature genes (Supplementary Table S2).

We then applied the Nearest Template Prediction (NTP) [97] method using the feature genes of subtypes and bulk RNA data from GIST cell lines (GSE95864) to assign phenotypes to three different GIST cell lines (GIST430, GIST-T1, and GIST882) with default parameters.

Metabolic activity analysis

To further investigate the metabolic characteristics of the G4 subtype, we used the COMPASS metabolic prediction tool [19]. Metabolic activity in key enzymes of individual samples was analyzed, generating likelihood values for each metabolic reaction. Cohen’s d values were calculated to assess the differences between G4 and other subtypes.

Human GIST cell lines

The human GIST GIST882, GIST430 cell lines were kindly gifted by Professor Jonathan Fletcher at Harvard Medical School. GIST-T1 cell line was purchased from MEISEN CELL Company (#CTCC-001-0436, Hangzhou, China). All cell lines were identified by short tandem repeat (STR) analyses and confirmed to be mycoplasma free. GIST-T1 was cultured in DMEM medium (Gibco) supplemented with 15% fetal bovine serum (FBS, Gibco) and 1% PenStrep (Gibco). GIST-430 was cultured in IMDM medium (Gibco) supplemented with 25% FBS and 1% PenStrep. GIST882 was cultured in RPMI-1640 supplemented with 20% FBS and 1% PenStrep. Cells were grown in a 5% CO2 cell culture incubator (Thermo Scientific, USA) at 37 ℃. The imatinib-resistant GIST430 cell line was developed by exposing parental GIST430 cells to step-wise increasing concentrations of imatinib until the cells exhibited stable growth under imatinib selection.

Cell proliferation assay

We determined the half-maximal inhibitory concentration (IC50) of GIST-T1, GIST882 and GIST430 cells to Imatinib, Olaparib and the combination. Cells were seeded in a 96-well plate and treated with the presence of a dilution series of Imatinib, Olaparib or, and 0.1% DMSO was used as control. After cultured for 2 days, the number of viable cells were determined by MTS assay (Promega, China).

For colony formation assay, GIST-T1, GIST882 and GIST430 were seeded in 6-well plates (2000 cells/well) and cultured in the presence of a dilution series of Imatinib, Olaparib or the combination, and 0.1% DMSO was used as control. After cultured for 2 weeks, colonies were stained with crystal violet (Beyotime, C0121) and counted.

Western blot

Western blot were performed as previously described [98]. In brief, for Western blot, 20 µg proteins were separated with 10% PAGE gel and transferred to polyvinylidene difluoride (PVDF) membrane. Next, the membrane was incubated with indicated antibody for overnight and conjugated with HRP-linked secondary antibody to detect signals. Antibodies used in this study are shown in Supplementary Table S6.

ELISA

Conditioned medium was collected from cell lines treated with Imatinib, Olaparib or the combination for 7 days. The 8-OHdG was quantified by ELISA (Elabscience, China) according to the manufacturer’s protocol.

Immunohistochemistry and multiplex immunohistochemistry

Immunohistochemical (IHC) or Multiplex immunohistochemistry staining (mIHC) were carried out using 4 μm paraffin embedded sections of mouse or human tumor. For IHC, the sections were performed heat-induced epitope retrieval and blocked in 10% goat serum for 30 min to reduce background. Then primary antibodies were used for overnight at 4 ℃ and HRP conjugated secondary antibody (Dako REAL™ EnVision™ Detection System, Peroxidase/DAB+, Rabbit/Mouse, K5007) was incubated for 30 min. Finally, 3,30-diaminobenzidine chromogen substrate solution (Dako REAL™ EnVision™ Detection System, Peroxidase/DAB+, Rabbit/Mouse, K5007) was used for chromogenic detection. The sections were scanned using KF-PRO-020 Digital Slide Scanner (KFBio, China).

mIHC was performed according to the manufacturer’s protocol using the TSA Multiplex IHC Assay Kits (TissueGnostics, TGT5C100). Briefly, sections were baked, dewaxed, rehydrated and fixed in 10% neutral formalin. Antigen retrieval was performed in EDTA buffer by microwave heating. After cooling, protein blocking was performed using 10% goat serum, followed by primary antibody for 30 min at 37 ℃. Then, HRP-conjugated secondary antibodies were incubated for 10 min at room temperature. Tyramide signal amplification (TSA) with fluorophores Opal followed for 10 min. After that, antigen retrieval was repeated to remove the antibody and TSA complex, starting a new cycle. After 5 cycles for planed antibodies, the TSA-stained sections were stained with DAPI. Stained sections were scanned using the TissueFAXS Spectra imaging system (TissueGnostics, Austria). Images processing, analysis of cell density and cell mutual distance were performed using StrataQuestv.7.0.165(TissueGnostics). The staining sequence, antibodies and TSA buffers are listed in Supplementary Table S6.

Subcutaneous implantation mouse model and in vivo treatment studies

All mouse experiments have been regulated under the Animals (Scientific Procedures) Act 1986 Amendment Regulations 2012 following ethical review by Sun Yat-sen University Animal Welfare and Ethical Review Body and have been performed in accordance with the license. For subcutaneous grafts, 5 × 10^6 GIST882 cells and 100 µL Matrigel were mixed and subcutaneously injected into the right flank of 4–5 weeks-old female NCG mice. Visible tumors developed in 4-6 weeks approximately. Mice were sacrificed and tumors were culled when tumor size reached 1500 mm3. We cut the tumors into 10 mm3 pieces and transplanted them into right axilla of NCG mice. After 6 weeks observation, mice with established subcutaneous tumors in approximate size were randomly divided into four groups (n = 4), which consists of control group, Imatinib group, Olaparib group and combination group. Mice were gavaged with Imatinib (50 mg/kg every 2 days) or injected intraperitoneally with Olaparib (30 mg/kg, 5 times a week). The drug dosage was administered in accordance with the previously established protocol [99–101]. Volumes were calculated by applying the formula v = 0.5 × L × w² in vivo and v = 0.5 × L × w × h in vitro, where v is volume, L is length, w is width and h is height. At the experimental endpoint, tumor tissues were harvested.

Oxygen consumption rate (OCR) measurement

At 24 h before the experiment, cells were cultured on XF-96 plate (2 × 10^4 cells/well). Before the analysis, the medium was replaced with Assay Medium and incubated at 37 ℃ in a non-CO2 incubator for 1 h. Subsequently, 1.5 µM oligomycin, 0.5-2 µM FCCP and 0.5 µM rotenone were incorporated sequentially to measure mitochondrial OCR parameters (the above reagent is the final concentration). The OCR was automatically calculated by the Seahorse XF-96 analyzer (Seahorse Bioscience, CA, USA).

Statistical analysis

Statistical analysis was conducted using R (v4.3.0). Categorical data were compared using Fisher’s exact test or Chi-squared test, while continuous data were compared using the Wilcoxon test or two-tailed Student’s t-test. Survival analysis was performed using the Kaplan-Meier method and compared with the log-rank test. Correlations were evaluated using Spearman correlation analysis, unless otherwise specified. All statistical tests were two-sided, and a p-value < 0.05 was considered statistically significant, unless otherwise indicated.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

XZ was responsible for the overall content as the guarantor. XZ, KS, and XW conceived and supervised this study. YD, XR, YY1, PC, MT, XL, SH and YY2 performed sample preparation, experimental operation and data generation. SC, YX and KS commented on the study and revised the manuscript. YD and YX performed bioinformatics analyses. YD, XR, YY1 and JC wrote the manuscript. YY1: Ye Yang; YY2: Yu Yang.

Funding

This study was supported by the Guangzhou Technology Research Plan (No. 2023A04J2212) and the National Natural Science Foundation of China (No. 82203642).

Data availability

Processed bulk RNA sequencing data of the 106 GIST tumors have been deposited in the Genome Sequence Archive (http://bigd.big.ac.cn/omix, accession number OMIX010362; preview link: https://ngdc.cncb.ac.cn/omix/preview/KTvM2e50). The raw RNA sequencing data of Xie F. et al. were obtained from the Genome Sequence Archive at HRA005970. The expression data of the GIST cell lines used in this study was obtained from Gene Expression Omnibus (GEO) at GSE95864. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Declarations

Ethical approval and consent to participate

This study was performed at the First Affiliated Hospital of Sun Yat-sen University. It was centrally approved by the Ethical committee of the First Affiliated Hospital of Sun Yat-sen University (No. IIT-2023-840) and complied with all relevant ethical regulations. All patients provided written informed consent of the samples for research.

Animal experiment

All mouse experiments have been regulated under the Animals (Scientific Procedures) Act 1986 Amendment Regulations 2012 following ethical review by Sun Yat-sen University Animal Welfare and Ethical Review Body and have been performed in accordance with the license.

Competing interests

The authors declare no competing interests.