Neoantigens arise from five primary sources: genomic variations, transcriptomic variations, proteomic variations, virus-derived tumor antigens, and other sources (11). Genomic variations include SNVs, INDELs, and gene fusions. (Next-Generation Sequencing) NGS technologies, such as Whole Genome Sequencing (WGS) and Whole Exome Sequencing (WES), have enabled the rapid identification of genomic alterations in tumors. By comparing tumor and normal tissue sequencing data, researchers can pinpoint somatic mutations that may lead to neoantigen formation. Complementary RNA sequencing (RNA-Seq) further helps in assessing the expression levels of these mutant genes and detecting alternative splicing events (26). By analyzing sequencing data from individual tumors, neoantigens derived from SNVs and INDELs can be identified as potential target epitopes. Conceptually, predicting epitopes based on NGS data involves three steps: (a) converting genomic mutations into mutated protein sequences, (b) predicting MHC binding, and (c) evaluating immunogenicity based on predicted binding, expression levels, and sequence similarity to non-mutated self-proteins (27) ( Figure 1 ). These steps reflect the core principle that an immune response requires the transcription, translation, processing, and MHC presentation of abnormal gene products. The integration of genomic and transcriptomic data is critical for filtering potential neoantigen candidates. Consequently, verifying expression and assessing MHC-binding affinity are pivotal in current computational pipelines (9). For example, the automated workflow pVAC-Seq (28), developed by J. Hundal et al., evaluates the binding capacity of candidate neoantigens to a patient’s HLA molecules, integrating tumor gene expression and mutation coverage data. Tools such as BWA (Burrows-Wheeler Aligner) and GMS (Genome Modeling System) align WES/WGS and RNA-seq data, while HLAminer and Athlate perform HLA typing. Mutations are annotated using the Ensembl database and Variant Effect Predictor (VEP), and NetMHC (29) predicts binding affinities, followed by filters to select optimal mutated peptide candidates. Similar pipelines include TSNAD (30), CloudNeo (31), and TIminer (32). Although these methods consider MHC-peptide affinity, they often omit other crucial factors like TCR recognition and sequence similarity to highly immunogenic epitopes (33, 34). In contrast, pipelines such as MixMHCpred2.2 and PRIME2.0 (35), address these limitations, improving predictions of antigen presentation and TCR recognition.

Immunopeptidomics uses mass spectrometry (MS) to directly identify peptides presented by MHC molecules in tumor cells. This approach involves two steps: (1) isolating MHC-bound peptides from tumor cells and (2) analyzing these peptides via MS to pinpoint tumor-specific neoantigens (36) ( Figure 1 ). MS-based immunopeptidomics has emerged as a powerful tool to directly profile the peptides presented on MHC molecules. This method complements NGS by verifying the actual presence of neoantigens on tumor cells (37). The direct detection of MHC-bound peptides addresses some limitations of bioinformatic predictions, particularly for neoantigens that are rare or prone to immune escape (38). To improve the positive predictive value of neoantigen prediction, advanced models integrate deep learning with MS data. The EDGE model (39) is trained directly using MS data rather than HLA-peptide binding affinity measurements in vitro. SHERPA (40) systematically combines 128 monoallelic and 384 multiallelic samples, employing an HLA-null K562 parental cell line with stably transfected HLA alleles to better emulate native antigen presentation. However, proteomics-based approaches have their own drawbacks, including limited sensitivity in detecting low-abundance peptides, technical variability in MS data, and the inability to capture all potential neoantigens due to sample processing constraints ( Table 1 ).

Bioinformatics tools play a pivotal role in predicting neoantigen immunogenicity. Several algorithms, including RPEMHC (41), UniPMT (42) and TEIM-Res (43) have been developed to assess MHC-peptide binding affinity, stability, and the likelihood of TCR recognition, respectively. Machine learning and deep learning further improved prediction accuracy by integrating large-scale multi-omics datasets, such as proteomic and immunopeptidomic profiles (44). However, the variability in HLA alleles and the complexity of antigen processing pathway continue to pose challenges for these computational models (45).

The inherent heterogeneity of tumors, both spatially and temporally, complicates the accurate prediction of neoantigens. Variability in mutation profiles within a single tumor and across different tumor sites may lead to inconsistent neoantigen expression and immune recognition. The analysis of melanoma samples from two distinct metastatic sites—the gastrointestinal tract and the pelvic cavity—in the same patient revealed significant heterogeneity in the functionality of the antigen presentation machinery (APPM) (44). Some tumor cells lose their ability of antigen presentation due to the deletion or mutation of β2-microglobulin (β2M) gene (46). This highlights the importance of considering tumor spatial heterogeneity when predicting neoantigens. By integrating multi-omics data, NeoDisc (44) can identify the defects in the mechanism of antigen presentation, shaping the heterogeneous landscape of tumor antigens. Over time, certain neoantigens may no longer be effectively presented due to the tumor evolution (47), or their immunogenicity may diminish as a results of alterations in the tumor microenvironment (48). To better understand the temporal heterogeneity of tumors, analyzing tumor samples at different time points provides valuable insights into their evolution.

The immunogenicity of neoantigens is shaped by both their intrinsic physicochemical properties, such as MHC-binding affinity and T cell receptor recognition and the dynamic interplay within the tumor microenvironment (TME). Neoantigens from mutant proteins at the cell membrane are more accessible to immune cells, enhancing immunogenicity, whereas those confined to the nucleus or other organelles are less likely to be processed and presented effectively (49).

Within the TME, immunosuppressive cells play a critical role. Regulatory T cells (Tregs) express high levels of CTLA-4 and PD-1, suppressing effector T cell activity (50). Myeloid-derived suppressor cells (MDSCs) secrete inhibitory factors such as IL-10, TGF-β, and arginase, further attenuating T cell responses (51). Tumor-associated macrophages (TAMs) also contribute by releasing cytokines like IL-10, which impair dendritic cell maturation and antigen presentation (52). TME exhibits considerable heterogeneity across different tumor types and individuals, which significantly influences the immunogenicity of neoantigens (53). Moreover, the composition of the TME varies significantly among tumors, with some dominated by Tregs and others by MDSCs, underscoring the need for personalized neoantigen prediction models that account for these differences. Combining TME remodeling strategies with immunotherapies—such as pairing ICIs with chemotherapy or targeted therapies—has improved clinical outcomes (52, 54). To achieve this, it is critical to incorporate TME remodeling strategies into neoantigen prediction frameworks, enabling more precise and tailored immunotherapies. This necessitates a holistic understanding of TME complexity, leveraging multi-omics data from genomics, transcriptomics, proteomics, and metabolomics. Furthermore, advanced simulations of the TME during neoantigen screening are vital to assess the immunogenicity of neoantigens within their specific heterogeneous contexts. For example, integrating the functional profiles of immunosuppressive cell populations, such as Tregs and MDSCs, into predictive models can refine neoantigen selection. These evolving requirements underscore the growing challenges and opportunities in developing robust neoantigen prediction pipelines.

The high degree of HLA polymorphism among individuals adds another layer of complexity. The HLA gene is situated on the short arm of human chromosome 6 and represents one of the most complex gene systems in the human genome (55). Due to the polymorphism of HLA gene, there are significant differences in genetic background among different races and regions (56). This polymorphism is primarily manifested in the amino acid positions of the antigen-binding groove, where variations at these sites determine the specific antigenic peptides that HLA molecules can bind, as well as the peptide binding affinity, stability and TCR recognition ability (57). Differences in HLA binding preferences and antigen presentation can significantly affect the immunogenicity of predicted neoantigens, necessitating personalized approaches in prediction and validation.

The performance of current bioinformatics models is inherently constrained by the quality and size of available datasets, which directly influence their predictive accuracy and generalizability. Incomplete datasets with lacking essential genomic, transcriptomic, or proteomic information, pose significant challenges to comprehensive characterization of neoantigens (58). Furthermore, many models are trained on datasets that lack sufficient representation across diverse cancer types, potentially introducing biases and reducing their applicability to underrepresented malignancies (33). These limitations often result in overfitting, where models exhibit high performance on training data but fail to generalize to independent datasets, or result in misclassification of neoantigen immunogenicity, producing false-positive or false-negative predictions that may impact therapeutic decision-making (59). To overcome these challenges, efforts are required to enhance dataset curation, including the integration of multi-omics data, the establishment of standardized data collection protocols, and the inclusion of diverse patient cohorts (60). Additionally, the complexity of algorithmic models presents difficulties in interpretability, which may compromise their reliability in both research and clinical settings. Improving model transparency and developing explainable AI frameworks could mitigate these issues, ultimately fostering greater confidence in computational predictions (61).