Recently, there has been an increasing number of reports on promising treatment paradigms based on reactivation of the immune system against cancer cells. Cancer immunotherapies aim to counteract the tactics employed by tumors that deactivate the immune system. Nevertheless, solely reactivating the immune system is not enough for the thorough elimination of tumors. It is essential that the reactivated immune system can distinguish malignant cells from their healthy counterparts.

The immune recognition of tumor tissues primarily relies on tumor antigens. Short antigenic peptides derived from tumor antigens are presented on the surface of the tumor cell by major histocompatibility complex (MHC) molecules serving as targets for the antitumor immune response. In humans, the MHC-I and MHC-II proteins are encoded by Human Leukocyte Antigen (HLA) genes, which are polymorphic in the human population. Given that the tumor antigens are the major target for antitumor T cells, they play a pivotal role in effective tumor elimination. Tumor antigens are typically categorized as tumor-associated antigens (TAA) and tumor-specific antigens (TSA). TAAs include antigens derived from genes overexpressed in cancer cells due to their malignant transformation, and comprise a class of normal self-proteins that are minimally expressed by healthy tissues. TAAs are generally weakly immunogenic due to central immune tolerance mechanisms. In contrast, TSAs are expressed exclusively on tumor cells. Most TSAs are neoantigens resulting from somatic mutations, such as insertion or deletions (INDELs), single nucleotide variants (SNVs), frameshifts and gene fusions (1). Since these neoantigens are products of tumor-specific irregularities, they are less susceptible to central immune tolerance, making them suitable candidates for therapeutic targeting.

Neoantigen cancer vaccines have emerged as a novel clinical approach to treat cancer (2). The purpose of a personalized anticancer vaccine is to direct T cells towards tumor eradication by leveraging neoantigens while preserving healthy tissue. There are two broad categories of immunotherapy treatments. Vaccinating against cancer induces long-lasting de novo antitumor immunity and is termed active immunotherapy (3, 4). Adoptive cell therapy (ACT) approaches, such as adoptive transfer of tumor-infiltrating lymphocytes (TILs), transgenic T cells, or chimeric antigen receptor T cells are based on the in vitro generation of tumor-specific T cells with subsequent infusion to the patient (passive immunotherapy). Currently, there is a variety of clinical trials, testing neoantigen-based anticancer vaccines either independently or in conjunction with other immunotherapies, checkpoint inhibitors or novel drugs under investigation. Numerous articles comprehensively review the field of mutation-derived neoantigen cancer vaccines. For detailed insights into preclinical and clinical studies, we recommend the review of Aurisicchio et al. (5). The review paper of Shemesh et al. (6) presents the clinical trial landscape of personalized therapeutic cancer vaccines, highlighting their opportunities and emerging challenges. Further insights into the challenges associated with targeting cancer neoantigens are outlined in the work of Chen et al. (7). Designing neoantigen cancer vaccines, trials, and trial outcomes are described in Biswas et al.’s work (8).

Detection of neoantigens is crucial for developing personalized cancer immunotherapies. Currently artificial intelligence (AI) is widely used to assess the factors that shape tumor immunogenicity. The use of AI for neoantigen prediction enhances the accuracy, efficiency, and personalized nature of cancer immunotherapy development by effectively analyzing and interpreting complex genomic data. However, the identification of putative neoantigens from genomic data still remains a challenge. To address this, specialized software tools have been developed for specific sub-tasks such as HLA typing and in silico prediction of peptide binding affinity to MHC molecules. Complex pipelines that encompass multiple analytical tasks have also been created. Current strategies for the identification of neoantigens are extensively reviewed in multiple articles (9–11).

For the successful implementation of AI vast amount of data is required. Genomic data comes in various forms, such as DNA sequences, RNA expression profiles. AI models can be trained to handle diverse data types, allowing for a more comprehensive, fast analysis of the factors influencing neoantigen formation. Significant amounts of high-throughput biomedical data, including omics and immunological data, have been accumulated in public databases, and can be transformed into novel insights. These data can be used for training a model with AI - based computational algorithm to properly interpret the data and learn from it in order to make accurate decisions based on the input information provided ( Figure 1 ). Additionally, AI models can help to identify novel neoantigens by recognizing patterns and associations in the molecular and cellular profiling data that may be challenging with the traditional methods.

Most state-of-the-art computational approaches for ranking and selecting candidate neoantigens predominantly rely on prediction methods, rooted in conventional machine learning (ML) algorithms, including artificial neural networks (ANNs), and modern AI architectures, trained on large experimental datasets.

Artificial Neural Networks are computational models inspired by biological neural networks. They learn the relationship between the inputs and outputs using samples from the training dataset (e.g., peptide sequences) and make predictions for the new samples. ANN’s are optimized by adjusting their parameters (weights and biases) based on the difference between the predicted values and actual values, utilizing the error-correction-learning rule known as back propagation.

Deep Learning (DL), a subset of machine learning and artificial intelligence stemming from ANNs, has gained increasing attention over the past years. The most commonly applied architectures include deep neural networks (DNNs) and convolutional neural networks (CNNs). DNNs consist of an input layer, multiple hidden layers, and an output layer with nodes in adjacent layers fully interconnected. CNNs primarily feature convolutional and pooling layers, often followed by fully connected layers. For an in-depth understanding of deep learning principles and concepts, we recommend the book of Goodfellow et al. (12). For definitions of AI and DL-related terms, please refer to our AI glossary ( Table 1 ).

Notable applications of deep learning in biomedicine, including medical imaging and drug discovery, are comprehensively covered in Wainberg et al. (13), while Wen et al. (14) delve into DL methods in proteomics.

Deep learning requires all input and output variables to be numeric. One important aspect of DL is data preprocessing or input encoding, which transforms raw data, such as peptide or protein sequences, into a suitable format for learning. Designing novel representation methods for protein sequence data is an active research direction. For example, the DeepLigand (15) approach treats each peptide sequence as a sentence, and each amino acid as a word, using the deep language model ELMo (16) to embed peptides into vector representations for tasks like peptide-MHC binding affinity prediction.

In addition to DNN and CNN, other DL architectures, such as gated recurrent unit (GRU) and long short-term memory (LSTM) neural networks, have proven effective for the peptide sequence-based prediction tasks. These methods can model dependences between amino acid residues within peptides of varying lengths without artificial lengthening or shortening, and they tend to be substantially faster than standard neural networks.

Recent advances in Natural Language Processing (NLP) have demonstrated the effectiveness of complex models, such as Transformers, including BERT (Bidirectional Encoder Representations from Transformers) (17), and GPT (Generative Pretrained Transformer) (Radford et al., 2018) ¹ , in learning rich contextual word representations. They can be trained to understand semantics from text without labels (self-supervised learning) (18). Similar techniques have also been applied to learn features from a large corpus of protein sequence data from public datasets (19, 20).

Another important characteristic of DL is transfer learning, which involves initializing training with representations learned from a previous task. Instead of training a new network from scratch, pretrained models can be downloaded and further trained for a new task by adding additional layers or fine-tuned using the new data. Examples include BERTMHC (21), MHCRoBERTa (22) which use transformers and transfer learning for peptide-MHC binding prediction. The authors found that leveraging self-supervised pretraining on large protein sequence corpora can lead to improved performance, particularly when training data is limited.

Achieving optimal prediction accuracy requires the tuning of model settings, or hyperparameters, e.g. determining how fast the weights of NN should be adjusted during training. Hyperparameter search techniques use validation examples that are held out from training. We provide the reader with a helpful background for understanding approaches assessing the performance of AI systems and establishing the trust in it.

Numerous publications have explored the application of AI in cancer research, precision medicine (23), cancer immunotherapy (24), and neoantigen identification (25). To gauge the potential of AI-driven software solutions, several benchmarking studies have been conducted. Evaluating and comparing tools is an essential part for their future application in the medical field and everyday clinical practice, as no single approach is universally applicable and having a dependable predictor or genotyper is vital. Despite the continually improving performance, critical questions regarding the application of AI technology in cancer immunotherapy remain. In this review, we summarize the core neoantigen calling pipeline, the recent research progress, and discuss the potential of artificial intelligence-enabled neoantigen identification, along with its current limitations and challenges.

The core computational pipeline established for the process of identification and selection of genomically encoded antigens that are of immunological significance includes the following steps (25):

The effective pattern recognition by AI allows for the development of personalized cancer treatments by considering the unique genomic profile of each patient’s tumor. As standard practice, neoantigens are predicted from the mutated peptides by assessing their ability to trigger an immune response. The development of AI-based prediction tools allows immunologists to streamline the search for neoantigen candidates that require experimental validation ( Figure 2 ).

In the following, we provide an overview of the most common computational methods used in the neoantigen identification pipeline and outline the challenges associated with the process.

HLA typing of the individual patient samples, specifically the accurate identification of the individual set of HLA alleles (HLA allotypes), is essential. Peptide-MHC affinity strongly depends on HLA alleles, resulting in distinct immune responses among individuals (45). Genotyping the class I genes HLA-A, -B and -C, as well as the class II genes HLA-DRB1, -DQB1, and -DPB1 presents a non-trivial task.

Sequence-based typing (SBT) based on Sanger sequencing can be used for HLA typing. However, due to certain limitations, such as the need for additional sequencing to identify cis/trans polymorphism, the concordance rate of Sanger sequencing-based HLA genotyping is approximately 84% among different laboratories (46). Commercial software, such as uTYPE (Life Technologies. Brown Deer, WI), Assign-SBT (Conexio, San Francisco, CA) (47), and SBTEngine (GenDx, Utrecht, Netherlands) (48), along with some open-source tools, e.g. SOAPTyping (49) are capable of producing predictions from Sanger sequencing data. However, they are increasingly being replaced by NGS-based methods. High-throughput WES and RNA-seq sequencing data also serve as a foundation for HLA typing. Most HLA genotyping tools take NGS sequencing data as the input and output HLA types. The algorithms employed by the tools primarily differ in how they map sequencing reads to a panel of reference HLA allele sequences and the strategy they use to subsequently score candidate alleles (50).

OptiType (51) is a HLA genotyping algorithm based on integer linear programming, capable of producing accurate 4-digit HLA genotyping predictions (for example, A01:01) from NGS data. To maximize the number of explained reads by simultaneously considering all major and minor HLA-I loci when predicting 4-digit HLA genotypes, this process involves aligning sequences from whole exome/genome/transcriptome sequencing data with a known MHC class I allele reference. Many tools for HLA typing are freely available for academic use, such as seq2HLA, ATHLATES, HLAminer, SOAP-HLA-2.2. A comprehensive list is provided in Table 2 . Figure 3 depicts a generalised workflow for NGS-based HLA genotyping.

T cells recognize peptides presented on MHC molecules of tumor cell. These molecules come in two main classes: peptide-MHC class I complexes, found on nucleated cells and recognized by CD8 + T cells, and peptide-MHC class II complexes, displayed on antigen-presenting cells like dendritic cells, activating CD4 + T cells. The diverse peptide repertoire is influenced by allele-specific amino acid preferences of MHC molecules. Due to individual variations in MHC alleles, the presented repertoire varies across people, with certain alleles being more common. The peptide-MHC interaction determines neoepitope presentation, impacting the level and type of T cell responses generated. While experimental MHC binding assays involve synthesizing and testing peptides, this is laborious and expensive on a large scale. Consequently, various computational algorithms and tools have been developed to predict peptide-MHC binding or assess binding affinity between mutated peptides and the patient’s MHC alleles (70).

It is important to note that other biologic processes can impact antigen presentation and immunogenicity of a particular neoantigen beyond MHC binding. Other factors, such as delivery of antigen to antigen presenting cells, antigen cleavage and processing by immunoproteasomes, peptide-MHC complex stability, are also important determinants of immunogenicity (7).

Early prediction tools relying on techniques as position-specific scoring matrices (PSSM) or sequence-scoring functions, such as SYFPEITHI (71), RANKPEP (72), PickPocket-1.1 (73), MixMHCpred (74), encountered difficulties in recognizing correlated effects. These effects manifest when an amino acid’s binding is influenced by the other amino acids in the peptide. The limitations of earlier tools in recognizing such correlated effects emphasize the suitability of neural networks as methods adept at considering these complex interactions (75).

Over the last decade, MS-based MHC peptidomics has become the dominant source of information about MHC binding specificities, with the ability to analyze ligands at greater depths than in vitro binding assays. Compilation of MHC ligandome data – the entirety of HLA presented peptides has been advanced by mass spectrometry (MS) based immunopeptidomics, in which the whole immunopeptidome of the cell is harvested and then eluted ligands (EL) are identified using MS. First application of direct neoepitope candidate identification using MS in native human tumors was presented in the paper of Bassani-Sternberg et al. (76). The authors assembled the ligandomes from human melanomas to a depth of 95,500 ligands. Eleven ligands were derived from candidate neoantigens, and four were proven to be immunogenic in T cell validation assays. MS profiling of HLA-associated peptidomes in mono-allelic cells enabled more accurate MHC-I epitope prediction in the study of Abelin et al. (77). MS immunopeptidomics is also able to identify protein hotspots, or regions within a protein prone to proteasomal cleavage and ligand production (78). Freudenmann et al. (79) constructed their own dataset and identified thousands of peptides bound to 16 different HLA class-I alleles to assess critical factors needed to epitope presentation.

However, in EL MS workflows typically pan- or locus-specific antibodies are used for immunoprecipitation (IP) during the purification of peptide–MHC complexes. This results in inherently poly-specific or Multi Allelic (MA) data, which comprises peptides that align with multiple cognate MHC binding motifs (80). For example, any of the six different MHC-I proteins present in a cell might be responsible for a peptide observation. These data need to be deconvoluted, i.e. transformed to Single Allelic (SA) or single peptide-MHC annotations, to be employed for the training of MHC-specific binding predictors. The method NNAlign_MA (81) resolved this limitation by incorporating into the prediction algorithm training procedure a strategy called pseudolabeling, which clustered EL sequences with ambiguous cognate MHCs into single MHC specificities.

Various AI-based tools have been developed to predict peptide-MHC binding using a range of neural network architectures and strategies in an attempt to improve predictive performance and generalizability of their models. They work on multiple data types including peptide sequences and mass spectrometry profiles.

One major issue impeding the generalizability of ML models is the lack of binding affinity data for rare MHC alleles. This can be addressed using various approaches such as using the sequence homology of rare MHC alleles with common MHC alleles to infer potential ligand preferences as NetMHCpan (82, 83) does. Also, NNAlign_MA was deployed in NetMHCpan to deconvolute ligandomes from MS datasets (80).

Another way is to use transfer learning by pre-training models on more common MHC classes and fine-tuning the models on the data for rare MHC classes. This approach is used by tools such as MHCnuggets (84), ImmunoBERT (85) and MHCRoBERTa (22). ImmunoBERT used transfer learning from the Tasks Assessing Protein Embeddings (TAPE) (86). The TAPE model was trained on a dataset of over 31 million protein sequences from the Pfam database. The authors of MHCRoBERTa used self-supervised training with label-agnostic protein sequences from UniProtKB (87) and Swiss-prot databases, and then fine-tuned the training with data from the Immune Epitope Database and Analysis Resource (IEDB) (88).

Many tools use approaches adopted from other domains. From the image processing domain comes the convolutional neural network which can learn multiple intrinsic features of the peptide sequence that can be used to predict binding affinity. Examples of these tools include ConvMHC (89), HLA-CNN (90) and DeepMHC (91). MHCSeqNet (92) uses techniques from the natural language processing domain by treating epitope peptide sequences as sentences composed from amino acids as individual words.

Some tools use ensemble learning, a technique that combines the output of several models using a weighted or uniform consensus. The concept behind the consensus methods is that prediction performance can be further improved by integrating the outputs from several individual tools using a weighted scheme. This includes tools such as MHCflurry (93) and NetMHCcons (94). MHCflurry is supporting only a fixed set of alleles (95).

Others tools provide or require additional data. Tools such as HABIT (96) provides an interpretation of the impact of amino acid variants alongside the binding affinity prediction. EDGE (97) and MARIA (98) require transcript abundances and flanking sequence in addition to the peptide sequence and MHC allele.

A class of tool use mass spectrometry and immunopeptidomics data as input data instead of peptide sequence data. This class of tool includes HLAthena (99) which shows 1.5-fold enhanced accuracy compared to sequence based tools and SHERPA (100).

An overview of tools used for MHC binding prediction is shown in Table 3 .

Other tools focus on visualizing and comparing different MHC molecule binding specificities to aid the understanding of main binding properties An example of such as tool is MHC Motif Atlas (128, 129) which contains 1,013,733 ligands interacting with 135 MHC-I and 88 MHC-II molecules, including information about binding motifs, peptide length distributions, motifs of phosphorylated ligands, multiple specificities and enables users to download curated datasets of MHC ligands, MHC sequences and MHC X-ray crystallography structures.

T cell receptors (TCRs) play a pivotal role in surveillance and response to disease by recognizing peptide-MHC (pMHC) complexes. However, not all neoantigen candidates elicit an immune response from T cells even though they are expressed and presented on the cell surface (11). Understanding the rules governing how T cells recognize cognate antigen-MHC complexes remains a challenge in systems immunology.

The TCR is a heterodimeric protein comprising an α- and β-chain. Peptide specificity is primarily defined by the complementarity-determining region 3 (CDR3) loops. The diversity of the CDR3s results from genomic recombination of the variable (V), diversity (D), and joining (J) genes (142). The majority of previous studies have focused on the β-chain alone due to its higher diversity, resulting from the V-, D-, J genes together (142). In contrast, the α-chain results from V- and J recombination which leads to lower diversity and less interest. However recent research has highlighted the importance of both α- and β-chain CDR3s in TCR specificity (143, 144).

T cell receptor sequencing (TCR-Seq) is an NGS approach allowing scientists to study clonal expansion by selectively amplifying and sequencing antigen-specific CDR3 regions of the T cell receptor. However, TCR-Seq data analytics is challenging as tumor-specific T cell responses constitute a small proportion of the overall pool of in vivo T cell responses with irrelevant specificities (145). New analytical tools have been developed to parse and draw meaningful sequence concepts or motifs from the TCR-Seq data (146). The TCRdb database contains more than 277 million TCR sequences from over 8265 TCR-Seq samples across hundreds of tissues, clinical conditions and cell types (147).

Assessing the interactions between neoepitopes and TCRs is essential for designing immunotherapies. For instance, identifying compatible TCRs in the patient’s circulation can help inform the selection of neoantigen vaccine candidates. Various experimental approaches, such as tetramer analysis (148), TetTCR-seq (149) and T-scan (150), have been developed to detect pairing of TCR–pMHC complexes. However, in vitro experiments associated with the testing of a large number of putative candidates demand experimental time and costs.

TCRdist (143) represents an unsupervised distance-based method exploiting the similarity between TCRs to produce clusters of TCR sequences that likely recognize the same antigen, and predicting binding for a given epitope sequence. The methods like TCRex (151) and DeepTCR (152) trained antigen-specific TCR models and would have problems to generalize to unseen peptides. In response, the scientific community has turned to ML and AI-based approaches to develop computational solutions for TCRs and peptide–MHC and TCR–peptide interaction prediction.

NetTCR (153) facilitates sequence-based prediction of TCR binding to pMHC complexes using CNNs. CNN is an appropriate model to handle unaligned peptide and TCR sequences differing in length. The model was trained on the IEDB data, containing TCR β-chain CDR3 sequences and corresponding peptide targets presented by most common MHC-I HLA-A*02:01 allele. Negative data examples were supplied for the learning by generating wrong combinations of TCRs and peptides, and additional negatives constructed from the TCRs of healthy donors. For NetTCR-2.0 (142) is a “shallow” CNN model, similar to NetTCR (153), it was exploited, but trained on paired TCR α and β chain sequence data. Nonbinding peptide-CDR3β pairs derived from 10X Genomics (154) Chromium Single Cell Immune Profiling of four donors were used as negative data set. The model has the potential to infer not only which TCRs are specific for a given peptide, but also which peptide is specific for a given TCR. This work also underlined the need for technologies for high-throughput paired sequencing of TCRs with known pMHC targets. The current optimal way to pair TCR α- and β- chain is through single-cell TCR sequencing (scTCR-Seq) (155). The authors of NetTCR-2.1 (156) provide lessons and guidance on how to develop models for TCR specificity predictions, how to best define negative data, and why it is recommended to apply similarity-based modeling, and include a performance evaluation as a function of “distance” to the training data when validating predictive power of ML-based approaches.

Most of the peptides in the published databases originate from viruses but not from tumor-associated antigens and there are only a few CDR3α sequences in databases available. Therefore, AI-driven approaches with improved generalization ability are needed, which do not show significant performance drop when evaluated on peptide sequences not used during model training. This challenge can be addressed by approaches based on transfer learning and NLP, capable to benefit from unsupervised pre-training.

As an example for the application of a newly emerging DL approach, Lu et al. (157) used transfer learning to develop pMTnet, a model predicting the TCR binding specificity of class I pMHCs. Utilizing the “Atchley factor” (158) they encoded TCR CDR3β sequences with five numeric values per amino acid, providing comprehensive biochemical characterization. These “Atchley matrices” were input into a stacked auto-encoder, an effective unsupervised learning algorithm. During training, the auto-encoder reconstructed input data, generating a 30-neuron numeric vector that encapsulates the inherent structure of the original CDR3s. The embedding of pMHCs closely followed the NetMHCpan algorithm. Fixed numeric encodings of TCRs and pMHCs were integrated into a DL network with a single neuron as the final layer for pairing prediction. To train this model, Lu et al. (157) employed a differential learning schema, using known interactions as positive data and introducing true and mismatched pairs for negative data, resulting in ten times more negative data by randomly mismatching TCRs and pMHCs. This approach allowed them to capitalize on a substantial volume of related TCR and pMHC data without explicit pairing information, showcasing the effectiveness of transfer learning.

For their NLP-based approach BERTrand (159) the authors constructed a hypothetical human TCR-peptide repertoire pre-training set comprising peptides from MHC-I MS peptide presentation experiments and TCRs from healthy donors, and this hypothetical TCR-peptide repertoire was used to perform masked language modeling (MLM), pre-training of the BERT model. Then the pre-trained BERT model was fine-tuned to predict TCR-peptide binding using the dataset of known TCR binders with their cognate epitopes and negative decoy examples generated by random pairing of reference TCRs with peptides. ERGO (pEptide tcR matchinG predictiOn) (160) and ERGO-II (161) utilize unsupervised TCR pre-training and use a pre-trained LSTM neural network architecture.

Further published tools for TCR-pMHC binding prediction are shown in our Table 4 .

Only a small fraction of predicted neoepitopes can be experimentally validated in vitro as true neoepitopes (180). Several general criteria are currently employed in the field to narrow down and prioritize the candidate epitopes. These criteria guide the selection of epitopes to induce specific “on target” immunogenic response while overcoming self-tolerance.

Several integrated software and comprehensive pipelines have been developed for tumor-specific neoantigen detection. The purpose of these tools is to make the prediction and prioritization of neoantigen candidates accessible. Here, we describe some of the notable tools and frameworks and their approaches.

For seamless vaccine design there have been several end-to-end pipelines developed. One of the frequently used end-to-end pipelines is FRED2 (FRamework for Epitope Detection), a Python-based immunoinformatic framework (204). Among the included tools there are several HLA genotyping tools (e.g.: OptiType), as well as peptide-MHC binding predictors (e.g.: NetMHCpan, NetMHCIIpan), and also the proteasomal cleavage predictor NetChop (205) is integrated. FRED2 ensures straightforward workflow and provides analysis tools to epitope detection and vaccine design (204). Another end-to-end pipeline is pVACtools, which produces an end-to-end solution for neoantigen characterization (206). To aid the vaccine design, pVACtools supports the identification of altered peptides and prioritizes them by incorporating various data sources, such as clonality of the mutation, mutant allele expression and peptide binding affinities. Among the tools integrated inside pVACtools there are binding predictors (e.g.: MHCflurry), databases (e.g.: IEDB), and a proteasomal cleavage predictor (NetChop). To extract neoepitopes from tumor sequencing data such as VCF files and expression files generated from RNA-seq, MuPeXI (Mutant peptide extractor and informer) provides a prioritization suggestion based on a combined score named priority score (207). It generates an output file with the list of mutated peptides and all the information needed (expression level, similarities to self-peptides, mutant allele frequency) to select the peptides for vaccine design (207). For HLA binding prediction NetMHCpan is integrated. It is a web-based tool, and also available as a command-line tool. TIminer is also a computational framework that provides complex immunogenomic analysis including HLA typing (Optitype), neoantigen prediction (NetMHCpan), characterization of immune infiltrates and quantification of tumor immunogenicity (208).

Another solution for peptide design includes prioritization algorithms. One such predictor is PRIME (predictor of immunogenic epitopes) (209). It captures molecular properties of both antigen presentation and TCR recognition. PRIME reveals experimentally validated biophysical determinants of TCR recognition and also establishes correlations with T cell potency. MixMHCpred is integrated for predictions of antigen presentation and TCR recognition. Beside the above-mentioned features, it improves the overall prioritization of neoepitopes. Another notable prioritization algorithm is DeepImmuno (210), a CNN based tool that predicts the epitope immunogenicity for CD8+ cells of 9-10-mer peptides. The prediction can run from the command line or from their web interface. The easy-to-use web interface has MHCflurry integrated to not only predict the immunogenicity of the specific HLA-peptide pairs, but the binding affinity score as well. DeepImmuno includes an independent generative adversarial network model, which can generate immunogenic peptide with the possibility of training your own model.

Most of the tools can predict neoepitopes from SNVs, some also incorporate INDELs (pVACseq (211), MuPeXI (207), TSNAD (212), CloudNeo (213), Epidisco (214), pTuneos (215), antigen.garnish (195), NeoPredPipe (216), NeoEpiScope (217), OpenVax (218)). A few focus solely on INDELs (ScanNeo (219)) or gene fusions (NeoFuse (220), INTEGRATE-neo (221)), while others allow users to input the variants as peptides (EDGE (97), DeepHLApan (103)).

A summary of various integrated pipelines and software tools for neoantigen discovery is provided in Table 5 .

The training of novel and improved algorithms requires continuous accumulation of verified tumor neoantigen data. Several studies have curated cancer antigen data, and constructed publicly available cancer antigen resources. These databases support the community in understanding the landscape of antigen presentation and provide necessary information for the development of neoantigen prediction tools. In addition to the well-curated data sets, several so-called in silico neoantigen databases that omit the experimental validation step have been built by taking advantage of existing neoantigen prediction software.

There are several well-curated datasets. One of the widely used, well-known resource is the Immune Epitope Database and Analysis Resource (IEDB) (88). It is a freely available comprehensive repository for diverse immunological data. This database contains experimental data from various host organisms about peptidic and non-peptidic epitopes, MHC ligand (Class I and II), T cell and B cell assays with a chance to gain insight into the possible disease context such as allergy, autoimmune or infectious diseases (234, 235). The database exists since 2003 and due to its enormous data content with over 1,600,000 epitopes and availability, this database is integrated in many other databases we have mentioned. However, IEDB’s data sets of verified T cell epitopes primarily consists of epitopes from bacteria or viruses and were not obtained by standardized experimental methodologies in the context of cancer. Furthermore, CEDAR (236) is the cancer epitope focused companion site of IEDB. This freely available database is similarily built to its companion and houses over 1,290,000 epitopes. Here, B cell, T cell and MHC ligand assay results are available in various hosts focusing on cancer types and stages.

Further curated databases include NeoPeptide (237), dbPepNeo (238), dbPepNeo 2.0 (239), TANTIGEN (240) and NEPdb (241). NeoPeptide focuses on cataloguing neoantigens from somatic mutations across different cancer types from clinical trials and in vitro experiments. At the time of its creation in 2019 it already contained 36,000 antigens and over 180,000 epitopes which has been expanded since (10). It provides details on various neoantigen characteristic such as mutation site, sequence and MHC restriction. The dbPepNeo databases include curated information about neoantigen data validated by mass spectrometry or immunoassays in human tumors. While version 1 focuses on validated MHC-I antigens in various tumor types, in version 2 the included neoepitope candidates increased to over 840,000 while also adding MHC-II data. Both versions help the user by categorizing all neoantigen’s confidence based on the strength of the experimental validation. TANTIGEN focuses on cancer antigens whose HLA binding is experimentally validated from tumor tissues. Over 1,000 tumor peptides from close to 300 proteins are catalogued based on which the T cell epitopes and HLA ligands are easy-to-list. However, it does not include peptides shown to be ineffective and lacks any association with clinical data. NEPdb was constructed via curating published literature with a semi-automatic pipeline by parsing and filtering abstracts with NLP toolkit. It includes curated data of 173 MHC-I and MHC-II neoepitopes and over 17,000 non-immunogenic peptides from 23 tumor types. The validation focuses both on in vitro and in vivo T cell assays.

Also, there are databases on verified binding and presentation. This category includes caAtlas (242), SPENCER (243), IEAtlas (244), HLA Ligand Atlas (245) and CARMEN (246). caAtlas is a database that contains information about mass spectrometry results of 9 cancer types and non-tumor samples. The data focuses both on MHC-I and MHC-II molecules and comprises around 140,000 modified peptides. SPENCER focuses on small peptides in cancer patients that are encoded by non-coding RNAs. The database contains mass spectrometry data of 15 cancer types from over 1,700 patients resulting in the identification of near 30,000 small peptides encoded by non-coding RNA in tumors. IEAtlas collects the immunopeptidome data of mass spectrometry datasets to find epitopes that bind MHC-I/II from non-coding regions. Currently over 245,000 such epitopes are identified from 15 tumor types and 30 non-tumor tissues. the database HLA Ligand Atlas provides a collection of natural HLA ligands presented on benign tissues. Natural HLA ligand information could be important for further tool development.

Besides the experimentally verified databases there are also a number of in silico predicted neoantigen databases with an enormous variety of potential neoantigens. TSNAdb v1 (247) collected information about millions of potential neoantigens from somatic mutation data. The predictions of version 1.0 are based on the HLA data of 16 tumor types collected from TCGA (68) and TCIA (248) and are generated by NetMHCpan. TSNAdb v2.0 (249) upgrades its toolkit to use DeepHLApan, MHCflurry and NetMHCpan and predicted neoantigens not only from SNVs but from INDELs and fusions. The altered criteria in v2.0 decreased the false-positive predictions resulting in almost 400,000 SNV-derived, around 140,000 INDEL derived and over 11,000 fusion-derived predicted neoantigens. TSNAdb includes HLA binding info for both mutant and wild-type peptides thus, facilitating the assessment of the DAI (247). TRON Cell Line Portal (TCLP) (250) catalogues MHC types and predicted neoepitopes amongst other publicly available data of 1,082 cancer cell lines. The data focuses on both MHC-I/II neoantigens in a cell-line-specific manner.

The set of verified neo-epitopes is still limited, and we envisage that larger neo-epitope datasets will lead to additional refinements in immunogenicity predictions. For a summarized overview of the above-mentioned neoantigen databases, see Table 6 , for a summary on immunology related databases and datasets see, Table 7 .

In 2016, the Tumor Neoantigen Selection Alliance (TESLA) was established as a collaborative effort to identify the most effective predictive algorithms for targeting neoantigens through large scale validation. Supported by the Parker Institute for Cancer Immunotherapy and the Cancer Research Institute (CRI) (189, 258), TESLA involved 35 public and private research teams worldwide. Each team employed its own unique neoantigen prediction algorithms to identify and prioritize neoantigens. The initial focus was on advanced melanoma, colorectal cancer and non-small cell lung cancer (NSCLC). Genomic data from the same six patient samples (3 melanoma, 3 NSCLC) was provided by the Alliance. The immunogenicity of candidate neoantigens was validated through MHC-restricted T cells in subject-matched peripheral blood mononuclear cells (PBMC). This study highlighted the significant differences in the prediction methodologies among the groups. No single methodology identified every neoantigen, nor a large majority of neoantigens, indicating the need for a standardized approach.

Besides testing the already existing predicting algorithms, the other goal of the TESLA was to identify key parameters shaping tumor epitope immunogenicity. The Alliance determined that approximately 50% of immunogenic epitopes are characterized by strong MHC binding affinity, prolonged half-life, high expression, and either low agretopicity or high foreignness. A model based on these five peptide features associated with presentation and recognition was developed and tested against independent cohorts of cancer samples. TESLA data is available (259) to qualified investigators and provides opportunities to benchmark the performance of neoantigen workflows.

Using the TESLA dataset, Buckley et al. (260) evaluated performance of seven publicly available methods - IEDB model (261), NetTepi (262), iPred (263), Repitope (264), PRIME (209), DeepImmuno (210) and Gao (265) - predicting whether an MHC-presented peptide might invoke a T cell response (i.e. whether a peptide is immunogenic). Filtering the TESLA dataset, originally comprising cancer peptides from 13 class I alleles, to retain alleles for which all models are applicable, and excluding peptides observed in any model’s training data, resulted in 27 immunogenic and 372 non-immunogenic peptides (lengths 9 or 10 aminoacids) that were experimentally tested against seven HLAs. They observed high numbers of false positives for all model. In this benchmark, PRIME identified 26 neoantigen from the total 27, successfully reaching the highest number of identified TESLA neoantigens.

Learning from a large set of data and identifying patterns of interest is the greatest strength of AI. The integration of AI applications in cancer immunotherapy and personalized medicine holds great promise, however, also comes with various technical and implementation challenges. Figure 4 summarizes the introduced bottlenecks of AI-based neoantigens discovery along with their potential solutions.

AI has already proven to be useful in everyday life from refining the text of manuscripts to troubleshooting codes (282). However, the risks are higher when applying AI to human health. The implementation of AI in general clinical practice can be a sensitive topic. Medical professionals spend decades learning, practicing, improving and the gained experience along the way is extremely valuable. Comparing AI that has unknown or unexplainable processes to the medical professional when it comes to diagnosis and decision making related to possible therapy or necessary surgery, is a rather delicate topic for discussion (283).

Nonetheless, it is undeniable that AI technology is currently needed in the medical field. One such field where AI´s involvement is certainly required is cancer immunotherapies. In the past decades, immunotherapy has become increasingly important as a new form of cancer therapy. For the development of cancer vaccines, quick and efficient processing of large data is required. One challenge is to identify tumor-specific antigens, the majority of which are unique for individual patients. Combining tumor sequencing data with the use of predictive algorithms based on machine learning and artificial intelligence allows clinical investigators to accelerate identification of therapeutically relevant neoantigens.

We reviewed multiple tools and a broad selection of prediction servers for neoantigen detection based on advanced AI methodologies. These tools are still far from widespread use in clinical practice as it can be difficult for users to choose the best server. There is a lack of reference data that should serve as an open benchmark to compare the approaches and validate the concordance of predictions among different tools. We encourage the standardization of techniques and harmonized protocols for sequencing, mutation detection, immunogenicity testing, and neoantigen candidate prioritization.

Our work highlights the barriers of applicability and clinical adoption of AI approaches. The insufficiency of experimental data for training and associated with it the lack of generalizability of AI-based models represents the major challenge. Novel approaches capable to overcome the critical role of data limitations are required for further development of in silico methods. Transfer learning has become increasingly relevant in this regard. AI models that can efficiently use all of the limited available data and transfer knowledge from other sources are extremely valuable.

Carefulness must be applied to the issue of performance guarantees both for training the model and for assessing how it will perform when deployed. Standard statistical and ML methods should be employed, such as bootstrap or a Bayesian method to assess prediction confidence intervals, to quantify the uncertainty of AI model in the output, and analyzing the sensitivity of the model’s output to certain parameters. Often the target and loss function used for training may not match the target and loss function important for the users. Bridging this training-application gap can be addressed by grounding methods, i.e. supplementing the model’s training with context-specific information, improving its ability to function effectively in disparate real-life situations.

A mechanistic explanation of the relationship between the peptide sequence, HLA allele and binding affinity remains an open topic of investigation. AI-based tools provide a potential solution in two ways: 1) Deep learning approaches can learn features automatically from unstructured data, bypassing the need to discover a mechanistic explanation. 2) Explainable AI techniques, such as attention mechanism, may be able to provide clues about aspects of the relationship that require further investigation. The two possibilities are not mutually exclusive and if early efforts focus on producing accurate and generalizable black-box models, then later efforts should attempt to use explainable AI techniques to understand the reasoning the model uses to make its predictions. As we navigate the path forward in personalized cancer immunotherapy, several questions remain. How can we expand the collection of well-curated neoantigen data, particularly for rare cancer types? What additional factors beyond peptide properties, such as protein structure and post-translational modifications, should be considered for neoantigen prediction? How can we enhance the interpretability of AI models, making them more transparent and accountable? These questions, among others, represent exciting avenues for future research and innovation.

By depositing the results of experiments and clinical trials in public databases, investigators will assist in making neoantigen prediction models more generalizable. Companies should agree to mutually exchange information beneficial to all parties in a benchmarking group and share the results within the group. As clinical studies will continually evolve to become more inclusive, harmonized and easily accessible, the aforementioned challenges of clinical integration of AI will also be bridged.

This review focuses specifically on AI and neoantigens, however, the use of AI approaches to predict cancer immunotherapy efficacy (284) and patient’s response to immunotherapy (285) is also worth mentioning. AI can utilize complex images such as histopathological slides and follow-up CT scans, extract information from multi-omics data (genomics, transcriptomics, epigenomics, proteomics, radiomics), integrating it with clinical data (medical history, laboratory tests, demographic information) to distinguish immunotherapy responders from non-responders. One of the major challenges in immunotherapy is to determine which patients are likely to benefit from the therapy. Tumor mutational burden (TMB) was proposed as biomarker and approved by the FDA to select patients eligible to receive pembrolizumab. The review of Addala et al. (285) discusses cancer-intrinsic and cancer-extrinsic features that can be analysed. Besides TMB, genomic intratumor heterogeneity (ITH) can also be used as cancer-intrinsic feature for outcome prediction, as it was linked to treatment resistance, recurrence and reduced patient survival. Advances in single-cell analysis technologies enable further insights into genomic ITH, neoantigen formation and presentation at single-cell level. Cancer-extrinsic features encompass the cellular composition of the tumor microenvironment (TME). AI deconvolution tools, e.g. CIBERSORTx (286), provide estimates of the immune cell proportions in the TME. The complex model capable to integrate multiple factors including tumor purity, TME composition, tumor evolution, genomic ITH and immunogenic neoantigen load would be of great importance. The parameters that govern the immunogenicity still remain largely unknown. The review of Xie et al. (287) outlines further barriers that must be overcome to enable effective anti-cancer immunotherapies. Tumors can escape from immunological surveillance through a number of mechanisms, including the loss of neoantigens induced e.g. by transcriptional repression or epigenetic silencing, disruption of neoantigen peptides presentation, and immunosuppressive TME. To compensate for the loss of targetable neoantigens, personalized neoantigen-specific immunotherapy should target multiple neoantigens (288). In the work of Xie et al. (287) additional compensatory strategies to address the issue of immune evasion of tumor cells are discussed.

The recent publication of Donisi et al. (289) also considers the mechanisms behind the resistance to immune therapeutic agents, in particular, the tumor immune microenvironment (TIME), a part of the TME, or microbiome influencing immune cells in the TME etc., and reviews multi-omics and AI approaches, e.g. those for dissecting the TME or inferring novel microbiome-linked biomarkers (289).

In conclusion, the field of neoantigen prediction is at the forefront of personalized cancer immunotherapy. The collaborative efforts of researchers, computational biologists, and immunologists have brought us closer to harnessing the full potential of neoantigens for precision medicine. With continued advancements in software, databases, and AI, we are on the cusp of a new era in cancer treatment, one that holds the promise of tailored immunotherapies that target the unique molecular signatures of each patient’s tumor. As both academic and industrial endeavors keep on to tackle the challenges outlined in this article, the future of personalized cancer immunotherapy appears brighter than ever.

AB: Conceptualization, Investigation, Writing – original draft, Writing – review & editing. ZN: Conceptualization, Investigation, Visualization, Writing – original draft, Writing – review & editing. FL: Writing – review & editing. MB: Writing – review & editing. MM: Writing – review & editing. MD: Funding acquisition, Writing – review & editing. LC: Writing – review & editing. RK: Conceptualization, Funding acquisition, Supervision, Writing – original draft, Writing – review & editing.