An integrative multi-omics analysis based on liquid–liquid phase separation delineates distinct subtypes of lower-grade glioma and identifies a prognostic signature

Currently, the understanding of the pathogenesis of human tumors is incomplete, which dramatically limits the development of effective treatment strategies. Classic perspective is that hallmark characteristics of tumors are acquired through gene-level alterations that disrupt the ‘lock-and-key’ binding type between crucial proteins. However, emerging evidences indicate that a large proportion of tumor malignant phenotypes originate from the intrinsically disordered domains (IDRs) of protein [1‐3]. Notably, the function of proteins with IDRs has been proved to be regulated by the liquid–liquid phase separation (LLPS) process [4‐6]. LLPS refers to the phenomenon that biological macromolecules (protein or nucleic acid) form a droplet like condensate without surrounding membranes through weak polyvalent interactions [7‐9]. LLPS are the formation mechanism of many membraneless organelles, such as stress granule, processing body (P-body) and nuclear speckle. Through forming an independent membraneless compartment by assembling biological macromolecules with a specific function into a specific space, cells could efficiently perform their biological functions, including reshaping chromatin structure, regulating gene transcription and translation, et al. [10, 11]. LLPS is a dynamic process involving the scaffolds, regulators and clients. Scaffolds appear to be essential for the structural integrity of biomolecular condensates. Regulators ensure that biomolecular condensates function properly. Clients reside in condensates only under certain conditions, and often contain components that specifically bind to components in the scaffolds [12]. It has already been confirmed that the LLPS status of many important proteins, including RNA-binding proteins and transcription factors, affect their own biological activities and the regulation of downstream signal pathways13, 14. For example, YTHDC1 is an RNA-binding protein, and the biomolecular condensate of YTHDC1 formed by LLPS is indispensable for protecting target mRNAs from degradation [15]. A growing understanding of the underlying molecular principles of LLPS has created awareness of their diverse functions in various cellular processes, including the stress response, the regulation of gene expression, and the control of signal transduction [8]. Aberrant phase separation of key molecules may result in disturbance of cellular signal pathways, and lead to a further pathology status of individual. Accumulating studies have revealed that the LLPS process is nonnegligible for the development and treatment of human diseases, including tumors [16, 17]. For instance, the deubiquitylase USP42 leads to nuclear speckle mRNA splicing through dynamic LLPS process to promote tumorigenesis [18]. The LLPS of YAP promoted by promoted by interferon-γ induces cancer resistance to anti-PD-1 immunotherapy [19]. Hence, we consider that exploring the role of LLPS should be a fruitful area in oncology research, and will further benefit the understanding of tumor pathogenesis, the prediction of prognosis, and the individualized selection of treatment options.

Diffuse lower-grade glioma (LGG) is the most common primary central nervous system tumor, characterized by high recurrence and progression rates even with the continuing development of multiple treatment modalities [20]. The marked heterogeneities in prognosis and therapeutic response of patients are always major clinical challenges. In the above context, this study attempted to identify and quantify such heterogeneities based on the LLPS patterns of LGG patients. Finally, four LLPS subtypes of LGG patients in The Cancer Genome Atlas (TCGA) cohort were identified with distinct prognosis, clinicopathological features, hallmark characteristics, genomic alterations, tumor immune microenvironment (TIME) patterns and immunotherapeutic responses. The constructed prognostic signature, namely LLPS-related prognostic risk score (LPRS), exhibited robust predictive power in the prognosis and the response to immune checkpoint inhibitor (ICI) therapy. To our knowledge, the present study is the first to reveal the multi-dimensional heterogeneities of LGG patients based on LLPS. Our findings might facilitate individualized prognosis prediction and better immunotherapy options for LGG patients.

Evidence is now mounting that LLPS process plays an integral role in the tumorigenesis and progression [17]. Moreover, the formation of different TIME patterns has also been shown to be correlated with LLPS due to its involvement in the regulation of immune signaling [34, 35]. Therefore, we supposed that a comprehensive exploration of LLPS-related biomarkers held great promise for the identification of novel subtypes of tumors, and the prediction of prognosis and immunotherapeutic response. In this study, we exclusively focused on LGG patients. Based on the expression profiles of 225 prognostic LLPS-related DEGs, we identified four LLPS subtypes of 423 LGG patients by using NMF algorithm. Then, significant differences among four LLPS subtypes were observed regarding prognosis, clinicopathological features, cancer hallmarks, genomic alterations, TIME patterns and immunotherapeutic responses. To make individualized integrative assessments, a prognostic signature, namely LPRS, was constructed via the WGCNA algorithm and LASSO Cox regression. Results revealed that LPRS was correlated with prognosis, clinicopathological features, genomic alterations and TIME patterns of LGG patients. The predictive power of LPRS in response to ICI therapy was also prominent.

Representative hallmarks of tumors include sustained proliferation, angiogenesis, EMT and genome rearrangements, and so on. How do tumors acquire these hallmark characteristics? In recent years, the field of LLPS is changing the way researchers and clinicians are now thinking about the acquisition of malignant characteristics of tumors [17]. For instance, MYC has the potential to form phase-separated transcription condensates by binding to super-enhancers, which can lead to the expression of VEGF and promote angiogenesis [36]. The LLPS of transcriptional coactivators, YAP and TAZ, is involved in the activation of EMT [37‐39]. The abnormal LLPS of ENL is enriched in genomic loci of chromosomes, and recruits large numbers of related transcription complexes, resulting in genome rearrangements in cancer [40, 41]. In this study, different LLPS subtypes of LGG patients exhibited distinct tumor hallmarks characteristics quantized by ssGSEA. The LS1 subtype was characterized by glycolysis, angiogenesis, EMT, hypoxia-responsive activation and regulation of apoptotic signaling pathway. However, the critical tumor hallmarks of LS4 were related to cell cycle and genome stability, which corresponded with the active genomic alteration of LS4. Compared with LS2 and LS3, LS1 and LS4 showed significant malignant progression features, which provided a possible explanation for the worse prognosis of LS1 and LS4. Simultaneously, the constructed LPRS also showed a significant correlation with these well-known hallmarks of tumors. These results provided compelling support for the nonnegligible role of LLPS in conferring specific hallmarks of tumors.

The classical view held that tumors could be divided into three different TIME patterns: immune-inflamed, immune-excluded, and immune-desert. It has long been known that glioma is dominated by immune-excluded and immune-desert patterns, which contribute, to a large extent, to the immune escape of glioma cells and the immunotherapy resistance of patients. The formation of specific TIME pattern is an immensely complex process involving numerous factors. Recent reports about the role of LLPS in innate and adaptive immunity shed new light on this filed. For example, the LLPS of cyclic GMP-AMP synthase (cGAS) promotes the secondary messenger cyclic GMP-AMP (cGAMP) production and innate immune signaling [42]. A large proportion of biomolecules in the transmembrane signaling receptors of T cells might phase separate into clusters to facilitate the transduction of signals and regulate immune responses of tumors [43]. In this study, LS1 had higher immune scores and stroma scores, but lower tumor purity compared with other subtypes, indicating that LS1 was surrounded by more nontumor components. Furthermore, LS1 displayed the activation of adaptive immune pathway and the infiltration of tumor infiltrating lymphocyte infiltration. Thus, it can be considered that LS1 corresponded to the immune-inflamed pattern, and was likely to respond well to immunotherapy. Subsequent prediction of the response to ICI therapy confirmed this speculation. Based on the TIDE algorithm, LS1 presented relatively lower TIDE score and higher MSI score compared with LS2 and LS3. In addition, subclass mapping analysis revealed that LS1 responded remarkably well to PD-1 inhibitor. Taking together, these findings strongly suggested that the established LLPS subtypes would contribute to the differential recognition of TIME patterns, and help to identify patients suitable for ICI therapy. Follow-up studies are warranted to determine the detailed mechanism of how LLPS processes regulate the specific formation of TIME patterns.

Given the multifaceted heterogeneities among four LLPS subtypes, we considered that it was feasible to construct a prognostic signature for the quantification of such heterogeneities, and also for the individualized integrative assessments. As expected, the constructed LPRS not only exhibited a close correlation with clinicopathological features, representative cancer hallmarks, genomic alterations and TIME patterns of LGG patients, but also possessed a prominent power in predicting prognosis and response to ICI therapy. LPRS was composed of twelve selected LLPS-related genes, of which two were regulators and ten were clients. The regulator TNPO1, also known as Karyopherin-β2, has been reported to inhibit the LLPS of an RNA-binding protein Fused in Sarcoma (FUS) and escort it into the nucleus [44]. Another regulator, SFRP2, is required for P-body assembly [26]. Of these ten clients, CADPS, CRTAC1, SCD5 and TPM1 can formed postsynaptic density [45]. FAM204A and PPIF are involved in nucleolus [46‐48]. SMU1 participates in the formation of stress granule [49]. Other three clients can form variety types of biomolecular condensates (FAM110B: centrosome, spindle pole body; TOP2A: nucleolus, centrosome, spindle pole body and P-body; XRN2: nucleolus, P-body and stress granule) [48, 50‐58]. As is already evident, what we currently know about these LLPS-related genes is almost exclusively confined to the forming types of biomolecular condensates that they are involved in. Thus, a more in-depth mechanism of how the LLPS processes that underlie the assembly of various biomolecular condensates affect the occurrence and development of tumors needs to be investigated in the future.

Up to now, there have been so many classifications of LGG patients based on classical biomarkers or star molecules related to a specific topic. For example, the IDH1 mutation status has long been recognized as a very important prognostic biomarker for glioma. Although the LLPS subtypes showed different distribution of IDH1 mutation status, we didn’t think the survival differences among LLPS subtypes were due to such difference. It could be seen that LS2 and LS4 had similar IDH1 mutation ratio, but the prognosis of LS4 was significant worse than that of LS2. Compared with previous classifications of LGG patients, the advantage of our LLPS subtyping was showing multi-dimensional heterogeneities, including prognosis, clinicopathological features, cancer hallmarks, genomic alterations, TIME patterns and immunotherapeutic responses, especially immunotherapeutic responses, a topic which is of great clinical interest. Nonetheless, several limitations of this study should be addressed. First, all analyses were performed based on the retrospective data of public databases, using the prospective multi-center cohorts will produce more reliable results. Second, due to the limitation of immunotherapy cohorts with publicly available transcriptional data and clinical information, we could only assess the predictive power of LPRS in response to ICI therapy by using the cohorts of urothelial cancer and metastatic melanoma. Finally, bioinformatic analyses are not able to deeply elucidate the molecular mechanisms, experimental evidences are indispensable to further exploration.

Taken together, we divided LGG patients into four LLPS subtypes with distinct prognosis, clinicopathological features, cancer hallmarks, genomic alterations, TIME patterns and immunotherapeutic responses. In addition, a prognostic signature, LPRS, was proposed for individualized integrative assessment. The findings might facilitate individualized prognosis prediction and better immunotherapy options for LGG patients, and further studies are needed to clarify this point.