The effect of immunotherapy varies across tumor types and patient populations (Okazaki et al., 2013; Weiss et al., 2019; Chamoto et al., 2020), highlighting the necessity of developing personalized immunotherapies based on tumor-specific antigens, which are absent in normal tissues (Jhunjhunwala et al., 2021). Studies have shown that patients with more tumor-specific antigens receive increased sensitivity to neoantigen-based immunotherapies, as well as activate more potent anti-tumor immune responses under ICB therapy. Therefore, individuals with a high mutation burden tend to produce more neoantigens, which makes them benefit more from immunotherapy (Ott et al., 2017; Yarchoan et al., 2019; Blass and Ott, 2021).

In addition to the somatic mutation, aberrant splicing is also an essential origination of neoantigens. Mis-splicing in various tumor types leads to the production of tumor-specific peptides (Frankiw et al., 2019). Previous studies have shown that peptides derived from mis-splicing can bind to major histocompatibility complex class I (MHC I) for T-cell recognition (Jayasinghe et al., 2018; Kahles et al., 2018; Smart et al., 2018; Frankiw et al., 2019). Although most aberrant splicing may introduce early stop codons leading to nonsense-mediated mRNA decay (NMD), some of these RNAs can still undergo a pioneer round of translation to produce some peptides to activate the immune system (Apcher et al., 2011). Compared to mutations that typically affect a single amino acid, mis-splicing, especially intron retention, often inserts a non-coding sequence into the transcript to generate more neoepitopes (Smart et al., 2018). Furthermore, many tumors often accompany the dysfunction of splicing factors (SFs), which exhibit widespread mis-splicing across the whole transcriptome (Wang and Aifantis, 2020). This global splicing change can produce more neoantigens than somatic SNV in many tumor types. For instance, in breast and ovarian cancers, mis-splicing produces at least twice as many neoantigens as those generated by nonsynonymous mutations (Jayasinghe et al., 2018; Kahles et al., 2018).

The splicing dysregulation in tumors can accelerate the development of neoantigen-based immunotherapy, as it provides an expanded candidate pool of antigens for positive selection. For example, fibronectin (FN) encoded by FN1 is a valuable AS-derived antigen resource (Villa et al., 2008; Huijbers et al., 2010; Jailkhani et al., 2019; Xie et al., 2019; Wagner et al., 2021). Through alternative splicing, FN1 can generate three distinct adhesive extracellular matrix isoforms, each with unique structural regions: V (IIICS), EIIIA (EDA), and EIIIB (EDB) (Dubin et al., 1995). Studies show that fibronectin containing EDA and EDB segments were significantly upregulated during tumor angiogenesis; while displaying low expression levels in normal adult tissues (Khan et al., 2005; Su et al., 2020). Based on this unique AS pattern, the CAR T-cell, mAbs, and cancer vaccine against EDA or EDB have been developed and shown to reduce tumor growth in several solid tumors, including melanoma and lung adenocarcinoma (Villa et al., 2008; Huijbers et al., 2010; Xie et al., 2019; Wagner et al., 2021). It is worth noting that these two splicing derivatives are accumulated in neovasculature, which is present in most solid tumors. Thus, immunotherapies based on EDA and EDB hold promise for extensive applications across tumors.

CD44 is another important gene whose AS-derived antigens have been targeted by many immunotherapies. For example, the mAb RG7356 targets the CD44s isoform and has shown efficacy in clinical trials for acute myeloid leukemia (Vey et al., 2016; D'Arena et al., 2014). Another mAb selectively targeting CD44v6 has also shown success in treating many cancers, including squamous cell carcinomas and a subset of adenocarcinomas (Heider et al., 2004). In ACT therapies, CD44v6-targeted CAR-T cells coexpressing a suicide gene eradicate autologous leukemia in vivo (Casucci et al., 2013). Besides targeting the FN and CD44 isoform, there also exist many successful immunotherapy designs targeting AS-derived other neoantigens (Table 2), such as a mAb against the splicing structural domain D of tenascin C (Nadal et al., 2020), mAb designed to bind to CLDN18.2 peptide (Sahin et al., 2008), and a cancer vaccination based on D393-CD20 peptide (Vauchy et al., 2015). Although additional function and safety tests are required, these designs hold an excellent prospect for clinical usage.

Typically, developing an immunotherapy strategy targeting splicing-derived neoantigen involves the following steps (Figure 1). First, bioinformatics approaches are used to detect cancer-specific AS events, screen AS-derived peptides, and predict the immunogenicity of peptide candidates. Then, based on the prioritized list of peptide candidates, peptides are synthesized to mimic potential immunogenic epitopes. Next, in vitro validations, such as peptide-MHC stability assay and T-cell function assay, are conducted to evaluate the immunogenicity of candidate peptides. Finally, in vivo validations with animal models are carried out to examine the effect of neoantigen-based immunotherapies. The efficacy of the designed therapy can be evaluated by phenotypes like changes in tumor size and immune cell infiltration (Ott et al., 2017; Chen et al., 2019; Arnaud et al., 2020; Cheng et al., 2022).

Tools have been developed for neoantigen screening by evaluating the binding affinity between splicing-derived peptides and various MHC-I (MHC-II) allotypes (Bayer, 2019). These Peptide-MHC immunogenicity prediction tools generally fall into two categories: scoring-based, for instance, PSSMHCpan (Liu G. et al., 2017) and MixMHCpred (Bassani-Sternberg et al., 2017), and machine learning-based, such as POLYSOLVER (Shukla et al., 2015). Upon extraction of AS-derived antigens that bind strongly to MHC molecules, it is also necessary to use computer tools to assess the immunological activity against the identified antigens. Agent-Based Models (ABMs) are the main computational approaches for such analyses (Butner et al., 2022). In ABM, each cell is represented as a discrete object (agent). These agents interact with their environment following predefined biological rules over discrete time steps (Bonabeau, 2002). One of the most successful ABM methods is IMMSIM (Puzone et al., 2002), which simulates T-cell responses, including T-cell activation, proliferation, differentiation, and antigen recognition. Following this approach (Celada and Seiden, 1992), to date, many ABMs are designed to predict a more comprehensive immune system response, not just the T-cell responses (Celada and Seiden, 1992; Chavali et al., 2008; Pennisi et al., 2013; Madonia et al., 2017; Shou et al., 2022). For example, One ABM model SimB16 was utilized to predict the immune responses of immunotherapy in B16 melanoma (Pappalardo et al., 2011). Another ABM model NetLogo has been successfully adapted to describe the interactions between the immune system and tumor cells (Chiacchio et al., 2014). In addition to ABMs, there are also several alternative approaches, such as Virtual Cell (Resasco et al., 2012) and PySB (Lopez et al., 2013), which use differential equation models to simulate changes in immune cells and cytokines to test the immunological activity. Collectively, these models will greatly speed up the design of immunotherapy targeting splicing-derived neoantigen.

However, the time-consuming development of animal models may not keep pace with the computational identification of antigens for a large-scale candidate screening analysis (Cheng et al., 2022). In future studies, tumor organoids are potential alternative models for the optimization of antigen-based immunotherapy, which will be discussed in subsequent sections.

As discussed above, the dysregulation of SFs may enhance immunogenicity by inducing widespread splicing defects, suggesting that combining immunotherapy with modulations of splicing factors can improve the therapeutic effect of immunotherapy (Figure 2). In mouse models of several solid tumors, degradation of the splicing factor RBM39 generates numerous AS-derived neoantigens, subsequently stimulating anti-tumor immunity and augmenting the effect of ICB therapy (Lu et al., 2021). In another study of colorectal cancer, the chemical activation of the splicing factor SRSF family achieves a similar effect (Matsushima et al., 2022). Furthermore, splicing manipulation can also boost immunogenicity by generating double-stranded RNA (dsRNAs) (Bowling et al., 2021; Ishak et al., 2021). Recent studies have revealed that spliceosome-targeted therapies can cause widespread cytoplasmic accumulation of mis-spliced mRNAs, a significant portion of which can form dsRNAs. These dsRNAs can be recognized by intracellular immune sensors and induce apoptosis in breast cancer cells (Bowling et al., 2021). Although only confirmed in mouse models, these examples provide compelling evidence that splicing modulation can activate immunogenicity and enhance the response to immunotherapy by potentially turning “cold” tumors into “hot” tumors (Galon and Bruni, 2019).

Besides the above contributions of AS-induced immunogenicity, on the other side, many mis-splicing events have also been found to associate with immunotherapy resistance (Gamonet et al., 2015a; Gamonet et al., 2015b; Sotillo et al., 2015; Fischer et al., 2017; Wang and Lee, 2018; Gong et al., 2019; Sciarrillo et al., 2020; Troiani et al., 2020; Deng et al., 2021; Qu et al., 2021; Bourcier and Abdel-Wahab, 2022; Zheng et al., 2022). For example, (Figure 2), in non-small cell lung cancer, secreted PD-L1 splicing isoforms can compete for binding PD-L1 antibodies, resulting in resistance to PD-L1 blockade therapy (Gong et al., 2019). In lung adenocarcinoma, aberrant splicing of the PD-L1 gene can produce a long non-coding RNA (lncRNA), which promotes resistance by enhancing c-Myc activity (Qu et al., 2021). Although CART-19 treatment targeting CD19 achieves a 70% response rate in patients with B cell acute lymphoblastic leukemia, tumors can also evade treatment via generating CD19 isoforms lacking exon 2 (Sotillo et al., 2015). These observations suggest that targeting harmful AS derivates is a complementary strategy for immunotherapy.

It should be noted that AS-driven resistance to immunotherapy works by different mechanisms. Therefore, the approaches to overcome these resistances vary accordingly (Figure 2). The secreted PD-L1 isoforms lead to drug resistance through competition for PD-L1 antibodies. Thus, using PD-1 antibodies or depleting the AS derivatives by antisense oligonucleotides (ASOs) are practical options (Gong et al., 2019). However, CD19 variants alter the antigen structure and escape from recognition by conventional CART-19. To overcome this resistance, depletion of the CD19 splicing variants cannot rescue the lack of CD19 antigen, and a newly engineered CAR T-cell targeting the CD19 AS-derived antigen is a better choice (Sotillo et al., 2015).

Inter-tumor heterogeneity presents a major reason for variable responses to immunotherapy (Saito et al., 2019; Chen et al., 2021). Distinct immune subtypes based on transcriptome successfully predict prognosis and immunotherapy responses in some cancer types (Chen et al., 2021). Alternative splicing is an essential source of transcriptomic and proteomic heterogeneity, which can further help to improve patient stratification. Mis-splicing can remodel the immune microenvironment in tumors. Studies have shown that AS-derived neoantigens can participate in immune reprogramming and directly influence the formation of the tumor microenvironment (TME) (Zhang Y. et al., 2021), suggesting that patients who harbor different splicing substrates may respond differently to the same immunotherapy (Zhang Y. et al., 2021; Su et al., 2022). Another study has validated this hypothesis in lung adenocarcinoma. This study showed that a specific subset of patients with a particular splicing pattern had been found to have higher immunogenicity, leading to better response rates to ICB therapy than other patients (Wu et al., 2022).

Furthermore, splicing alterations typically occur in a cancer-specific or stage-specific manner (Bonnal et al., 2020), indicating that the AS can act as biomarkers to optimize the therapeutic strategy (Le et al., 2015; Chandrakesan et al., 2020; Yang et al., 2022). Mis-splicing can serve as biomarkers for immunotherapy response. Similar to the correlation observed between ICB response and tumor mutational burden, patients with a higher mis-splicing disorder in the tumor tissues may present a higher response rate of ICB therapy (Frankiw et al., 2019; Lu et al., 2021; Matsushima et al., 2022). Besides, another study found that specific AS signatures can indicate immune activity and can be used to predict the response to immunotherapy (Chen et al., 2022). Moreover, some splicing variants can lead to resistance to particular immunotherapy. Besides mentioned AS derivates related to resistance, there are also many other AS events related to resistance, highlighting their critical role as therapeutic biomarkers. For example, skipping exons 5 and 6 of CD22 leads to resistance to CD22 CAR T-cells (Zheng et al., 2022), and D393-CD20 can lead to resistance to CD20 mAb therapy (Gamonet et al., 2015b). These observations demonstrate that splicing biomarkers can potentially assist patients in determining immunotherapy choices (Troiani et al., 2020; Qu et al., 2021; Zhao et al., 2021; Weng et al., 2022).

In the above studies, the experimental models play a crucial role in screening AS-derived antigens, evaluating AS-related immunotherapy, and exploring the biological functions of AS events. But the commonly used experimental models in current studies are mainly cancer cell lines and mouse models (Table 2), which have some limitations. First, in vitro culture of cells cannot accurately replicate the interactions with other cell types or the extracellular matrix. Second, differences in the genome and microenvironment between species make it challenging to translate findings from mice to humans. Especially, the nature of poor conservation of intronic sequences and minimal overlap of mis-spliced transcripts between mice and humans (Lieu et al., 2022) suggest that mice may be less useful in modeling patients with AS dysregulation. Last, it is complex and time-consuming to manipulate genes in mice by multigenerational hybridization, which may result in patients missing optimal treatment periods (Shang et al., 2022; Stribbling and Ryan, 2022). Recent studies have shown that organoids are highly effective in disease modeling and are widely utilized for basic research, drug selection and personalized medicine, which may also benefit the splicing-targeted immunotherapies (Drost and Clevers, 2018).

In cancer modeling, tumor organoids can recapitulate the (epi)genetic and phenotypic diversity of distinct tumor cell subclones, as well as their morphological features (LeSavage et al., 2022). Furthermore, tumor organoids also enable the modeling of TME, including the functions of non-neoplastic cells, the signaling of niche-specific soluble factors, and the altered extracellular matrix (Neal et al., 2018; Yuki et al., 2020). To date, many tumor organoids have been developed and reproduced pathological features (Yeung et al., 2010; Boj et al., 2015; Drost et al., 2016; Hubert et al., 2016). For example, tumor organoids with 3D microfluidic devices maintain immune cell composition of the donor tumors and are used to evaluate the response of ICB treatment (Jenkins et al., 2018). Another approach to culture organoids with an air-liquid interface system achieves similar success (Homicsko, 2020). Additionally, tumor organoids can be grown for an extended period, modified to investigate specific genetic alterations and maintain their features across multiple passages, making them wildly used for basic research.

Tumor organoids are a reliable model for the functional study of alternative splicing events (Rossi et al., 2018; Artegiani et al., 2020; Dekkers et al., 2021). For instance, a study uses cortical organoids to demonstrate that the reintroduction of the archaic splicing variant of NOVA1 alters neurodevelopment (Trujillo et al., 2021). Another example of co-culturing organoids and autologous lymphocytes proved that interactions between melanoma tumor cells and T-cells can induce splicing isoform FKBP51s, which is related to the resistance of anti-PD1 blockade therapy (Troiani et al., 2020).

Organoids also provide an ideal model for large-scale drug screening. And specially, drug screening using patient-derived organoids can further guide personalized treatment options. One successful example is ‘eribulin’, selected by cancer xenografts and organoid platforms. Individuals with this treatment achieved complete remission for nearly 5 months (Guillen et al., 2022). More importantly, organoids are valuable for exploring treatment combinations. A previous study has shown that the combination of KRAS inhibitor AMG501 and EGFR inhibitor cetuximab achieves a synergistic effect for treating colorectal cancer organoids with KRAS ^G12C mutation (Amodio et al., 2020). Another tumor organoid from circulating tumor cells of patients proved that GKB202 is a promising adjuvant for 5-FU-based treatment (Li et al., 2021). Furthermore, the organoid is particularly effective in immuno-related drug selection, as the response to immunotherapy is shaped by both cancer cells and the TME (Xu et al., 2018). To date, many organoid platforms have been developed to evaluate the effects of CAR T-cell therapy (Dijkstra et al., 2018; Schnalzger et al., 2019), ICB therapy (Jenkins et al., 2018), or other neoantigen-based immunotherapies (Courau et al., 2019; Gonzalez-Exposito et al., 2019).

Currently, though, there are no organoid platforms for investigating the functional role of alternative splicing, screening AS-derived neoantigens, or directly evaluating splicing-based immunotherapy. The establishment of these platforms in the future will definitely accelerate the application of alternative splicing in immunotherapy (Raue et al., 2023).

The priority in identifying immunotherapy-related AS derivates is the comprehensive detection of AS events. Numerous computational tools have been developed to detect AS events, yet selecting the optimal algorithm remains challenging due to the large inconsistencies between the outputs of different software (Wang et al., 2015; Mehmood et al., 2020). Given the different designs and sensitivity of algorithms in detecting different types of AS events, one naive way is to integrate the power of multiple algorithms by manually selecting the results, which will induce artificial bias (Mehmood et al., 2020). There is an urgent need for the customized design of more reliable algorithms, considering the specific characteristics of AS events and research requirements. For example, SF3B1 mutation leads to the cryptical 3`splcing sites (3′ss), many of which are not reported in the latest annotation reference. As there is no method specific to detect these 3′ss events, an effective method is needed to expand the reference by creating a dataset-specific annotation file (DeBoever et al., 2015; Liu and Rabadan, 2021).

In addition to algorithm limitations, sequencing technology also encounters obstacles in identifying AS events. Currently, most studies identify and quantify splicing isoforms starting from bulk short-read RNA-seq data (Cieślik and Chinnaiyan, 2018; Ferragut Cardoso et al., 2022; Toffali et al., 2023). These studies typically map the short-read RNA sequences to a reference genome using software such as MISO (Katz et al., 2010) or rMATS (Shen et al., 2014), or assemble de novo using tools including StringTie (Pertea et al., 2015) or Trinity (Grabherr et al., 2011). These methods enable the identification of splicing junctions and estimation of isoform abundance based on read counts, providing a global picture of alternative splicing events over the bulk tissue level. However, the short sequencing reads are limited to detect complex and full-length novel isoforms (De Paoli-Iseppi et al., 2021). Additionally, Bulk RNA-seq data is unable to depict intratumor heterogeneity or identify AS events specifically and commonly expressed in all cancer clones, which are the important feature of neoantigen. Single-cell long-read RNA-seq is a powerful tool for studying splicing heterogeneity, while it still has limitations such as low throughput and technical noise. Future efforts should focus on developing more efficient single-cell long-read RNA-seq technology (Singh et al., 2019) and customized algorithms. Finally, the best practices may involve coupled analysis using both long reads and short reads sequencing techniques (Au et al., 2013).

After identifying and quantifying AS events, researchers need to pinpoint AS event candidates contributing to cancer immunotherapy. Many in silico tools have been employed to predict immunological activity against new antigens (Rammensee et al., 1993; Lin et al., 2008; Andreatta and Nielsen, 2016; Jurtz et al., 2017) and analyze the function of cancer-related AS events (Kahles et al., 2018; Liu B. et al., 2021; Liu and Rabadan, 2021; Qi et al., 2022). However, experimental validation of the immunogenicity of such computationally predicted neoantigens will need to be seriously assessed. Understanding the function of AS events is often challenging because it is difficult to introduce abnormal AS isoforms into the experimental model. Therefore, many experimental studies are only limited to AS events that lead to NMD, because the loss-of-function consequences are easier to manipulate and interpret (Lareau et al., 2007; Thomas et al., 2020; Bradley and Anczuków, 2023). Excitingly, there are new approaches of DNA/RNA-targeted CAS with CRISPR-based screening to carry out more unbiased analysis of splicing events and their impacts on tumor (Mou et al., 2017). It is worth noting that manipulating DNA sequence may induce off-target effects, such as creating unexpected splice sites and disrupting the chromatin structure.

Moreover, current experiment models in this field are based on cell lines or mouse models, but they harbor limitations as discussed above. Organoids offer an alternative model for investigating AS derivatives and neoantigen screening; however, their application in AS-based immunotherapy is still in its infancy. There is an urgent need to build one-stop organoid platforms for studying splicing-based immunotherapy (Chandrakesan et al., 2020; Yang et al., 2022). It should be noted that all organoids, cell lines, and mouse models possess distinct strengths and limitations. The selection of an appropriate model should be guided by the specific research need.

AS-derived neoantigen may also display a high tumor heterogeneity with a varying distribution in different tumor clones. Immunotherapies based on one single target could only eliminate a part of tumor cells, which may accelerate tumor evolution and disease relapse. Thus, a better option is using multiple-target immunotherapies (such as polyclonal antibodies) or combining multiple therapy strategies, which may cover all cancer clones. Furthermore, some AS products show dramatically higher expression in tumors, which does not mean these AS events are totally absent in normal cells. For example, the reported AS-derived antigens CD44v isoform, CD20 isoform D393-CD20 (Vauchy et al., 2015) and CLDN18 isoform CLDN18.2 (Sahin et al., 2008), are also detected in normal cells. Targeting these neoantigens would also influence normal cells, which may lead to serious side effects. Therefore, it is necessary to control dosing by individualized assessment before treatments.

Instead of targeting AS-derived neoantigens directly, combining immunotherapy with splicing modulations is also a promising treatment solution. Targeting the splicing regulators can induce global splicing changes to expand peptide diversity, enhance immunogenicity, and increase ICB therapy’s efficacy. However, the increased complexity of tumor transcriptomics may enable the rapid evolution of tumors to develop new carcinogenic characteristics, such as immune evasion and treatment resistance (Jayasinghe et al., 2018; Kahles et al., 2018). Thus, it is crucial to balance the benefits and risks when combining SF modulation with immunotherapy, either by adjusting the dose or the timing of the therapy.

It is difficult to directly target harmful AS derivatives due to the design or delivery of antisense oligonucleotides or small molecule drugs. Only a few compounds that target specific RNA isoforms have shown clinical utility to date (Sheridan, 2021). ASOs are designed to correct splicing errors by binding a reverse complementary sequence in a target pre-mRNA, thereby preventing its interaction with the splicing machinery. However, it is still a challenge to deliver ASOs to tumor lesions. For example, Spinraza, an FDA-approved treatment that corrects the splicing of SMN2 for spinal muscular atrophy, must be administered by direct injection into the spinal column. Besides, there are also small-molecule compounds that can induce targeted RNA degradation to prevent harmful mis-splicings (Umuhire Juru and Hargrove, 2021). The first small-molecule drug is Evrysdi, which also targets and corrects the splicing of SMN2 (Sivaramakrishnan et al., 2017; Sheridan, 2021). It allows oral administration but is relatively more complicated to design than ASOs. Notably, currently these two types of drugs are more commonly used in genetic diseases rather than tumors, which means it still demands extensive efforts to apply these methods in tumor immunotherapy.

In conclusion, we emphasize that alternative splicing presents a promising avenue for immunotherapy. While the current study in this field is still in its early stage, breakthroughs in both bioinformatics algorithms and biological technologies are critical to accelerate the development of AS-based immunotherapies.