Chemical Strategy and AI-Enabled Peptide Vaccine Development

Abstract

Peptide vaccines have emerged as a versatile platform complementing traditional vaccines by offering high safety, precise epitope targeting, and ease of manufacturing; however, they suffer from intrinsically weak immunogenicity, human leukocyte antigen (HLA) restriction, and poor in vivo stability. Recent progress in immunoinformatics and artificial intelligence (AI) has transformed epitope discovery from empirical trial-and-error screening toward rational, systems-level design. Machine learning models trained on binding assays, liquid chromatography–tandem mass spectrometry–derived ligandomes, and T-cell response data now enable increasingly accurate prediction of peptide–major histocompatibility complex binding, T-cell receptor recognition, and even the integrated antigen-presentation cascade, spanning sequence-based, structure-informed, and pan-allelic frameworks. In parallel, medicinal chemistry strategiesincluding cyclization, lipidation, glycosylation, stapling, incorporation of noncanonical amino acids, and construction of multicomponent or tolerogenic conjugatesaddress key biopharmaceutical bottlenecks by enhancing immunogenicity, extending half-life, and enabling organ- and cell-type–specific delivery. Together, these computational and chemical advances are beginning to bridge the gap between in silico performance and clinical efficacy, supporting the development of peptide vaccines that can be tailored to diverse HLA backgrounds, disease settings, and therapeutic goals. This review summarizes current progress at the interface of AI-driven epitope prediction and chemical modification with a focus on how their integration can yield next-generation peptide vaccines with improved potency, durability, and safety.

1. Introduction

Vaccines comprise weakened microorganisms, subunits, or genetic material and mimic a first infection to establish a durable immunological memory. Upon future encounters with the real pathogen, memory B and T cells quickly trigger a strong immune response, acting as a preventive or therapeutic means. Based on this principle, modern vaccinology provides protection against a wide range of infectious diseases and even noncommunicable diseases such as cancer. Traditional vaccines typically elicit strong immune responses, preventing millions of deaths each year. − Nevertheless, many diseases still require effective vaccines, spurring the development of novel vaccine platforms and design strategies.

mRNA vaccines and peptide vaccines have become new strategies for disease treatment and prevention. mRNA vaccines utilize translational machinery to produce antigens in host cells, whereas peptide vaccines deliver defined short epitopes derived from pathogen or tumor proteins. Both platforms avoid the use of live pathogen components, elicit a strong immune response, and generally exhibit more favorable safety profiles than traditional vaccines. mRNA vaccines played a pivotal role during the COVID-19 pandemic by enabling rapid vaccine development. , Beyond infectious disease, mRNA vaccines are also used in cancer therapy: a recent randomized trial showed that an individualized mRNA neoantigen vaccine combined with checkpoint blockade markedly prolonged recurrence-free survival in high-risk melanoma patients compared to immunotherapy alone. Despite the success of mRNA vaccines, current mRNA vaccines still have challenges. The mRNA molecules are inherently unstable and easily degrade, which complicates storage and delivery and necessitates formulation in protective carriers (e.g., lipid nanoparticles). , Given these limits of mRNA platforms, there has been a growing interest in peptide-based therapeutics.

Peptide vaccines offer a chemically definable platform for precision epitope targeting and controllable immune engineering, providing a compelling way to address limitations that are not optimally served by mRNA approaches. A multiepitope peptide vaccine for melanoma (incorporating CD8⁺ and CD4⁺ T-cell targets) was shown to durably improve long-term survival in a phase II trial. Similarly, a personalized neoantigen peptide vaccine in glioblastoma elicited robust T-cell responses and was associated with extended median survival in patients who developed multiepitope immune responses. These advancements highlight the growing potential of peptide vaccines in combating a diverse range of diseases. Peptide vaccines offer several attractive features that address some limitations of other vaccine types. Peptide formulations can be lyophilized and stored in solid form at room temperature, a notable logistical advantage over nucleic-acid vaccines. Moreover, peptide immunogens consist of specific antigenic epitopes with no infectious potential, and they can be engineered or chemically modified to minimize off-target effects and improve the safety profile. , Importantly, peptide-based vaccines have already shown promise in settings such as cancer immunotherapy and emerging infectious diseases. For example, a multipeptide vaccine targeting melanoma antigens (12 CD8⁺ T-cell epitopes plus 6 helper T-cell epitopes) improved overall survival rates in male patients with a high risk of melanoma, demonstrating the value of including cognate helper peptides. In another trial, a multiepitope peptide COVID-19 vaccine (CoVac-1) safely generated robust, variant-proof T-cell immunity in immunocompromised patients, outperforming the effects of even standard vaccines in that population. Together, such examples highlight how peptide vaccines can achieve potent immune activation with minimal toxicity and can be rationally designed for a wide range of applications.

Despite their notable advantages, peptide vaccines face several constraints. Immunogenicity is relatively weak. Without optimized delivery systems and adjuvants, peptide vaccines typically elicit low or short-lived antibody titers and comparatively modest T-cell responses. , In addition, peptide vaccines are also subject to HLA restriction. T-cell recognition requires peptide presentation by the host MHC molecules. Single-epitope formulations may perform well in individuals carrying compatible alleles, yet they show limited efficacy in others. Peptide vaccines often exhibit poor in vivo stability and are rapidly cleared. Unmodified peptides are quickly degraded by proteases and filtered out by the kidneys, shortening systemic half-life and narrowing the window for dendritic-cell uptake, cross-presentation, and germinal-center maturation. ,

Recent advances in immunoinformatics and AI have begun to solve the problems related to immunogenicity and HLA coverage. Machine learning models trained on vast immunological data sets can predict which peptide fragments are likely to be antigenic and bind strongly to a wide range of common HLA alleles. Modern deep learning approaches have greatly increased the accuracy of B-cell and T-cell epitope prediction, outperforming older sequence-based tools. For example, using large pathogen sequence databases and machine learning algorithms, scientists can screen for peptide candidates that are conserved and antigenic and predicted to bind with high affinity to multiple HLA alleles. These AI models can rank candidate vaccines based on predicted HLA binding ability and integrate factors such as antigen processing (AP) and T-cell receptor (TCR) recognition to predict results that can translate into actual immune responses. AI-driven design enables a “rational vaccine design” approach. Researchers no longer rely on trial and error or single antigens but instead use algorithms to select combinations of peptide epitopes that provide broad immune coverage.

Complementing design advances, chemical modification strategies improve stability, presentation, and innate immune activation. Cyclization constrains conformational flexibility, shields protease-susceptible termini, extends the serum half-life, and stabilizes conformational epitopes in bioactive geometries. Lipidation converts peptides into self-assembling lipopeptides that enhance lymphatic trafficking and multivalent presentation while intrinsically engaging pattern-recognition receptors, thereby integrating delivery and adjuvant functions. Glycosylation enhances solubility and protease resistance and enables targeting of dendritic-cell lectin receptors to facilitate uptake and cross-presentation. These chemistries are compatible with multiepitope and synthetic long-peptide formats, as well as modern adjuvants and particulate carriers, promoting robust CD4⁺ and CD8⁺ T-cell responses. Peptide vaccines constitute a safe, precise, and readily manufacturable platform. Continued integration of computational prediction, immunopeptidomics, and medicinal-chemistry toolkits promises durable, population-tailored, and clinically translatable peptide-based immunization strategies. This review focuses on recent advances in how AI and chemical modifications are empowering peptide vaccine design (Figure ).

1.

AI epitope prediction and chemical modification refine AP to engineer effective peptide vaccines.

2. AI Empowers Peptide Vaccine Design

T-cell activation requires recognition of peptide-loaded MHC complexes presented on the surface of antigen-presenting cells (APCs). , The peptide fragment processed from antigens is a T-cell epitope, which is presented by an MHC complex and recognizable by CD4 and CD8 T cells. , Upon epitope recognition, T cells perform effector functions. MHC class I ligands typically comprise 8–12 amino acids, and in individuals expressing up to 6 distinct MHC class I alleles, the theoretical immunopeptidome diversity exceeds 1 billion peptides. MHC class II binds longer peptides (12–25 amino acids), making prediction more complex. − Therefore, developing epitope prediction methods with greater generalizability and accuracy is a crucial component in the advancement of peptide-based vaccines and immunotherapies. This section summarizes AI-driven approaches for peptide vaccine design, focusing on antigenic peptide prediction and the role of chemistry in epitope prediction (Figure ).

2.

Evolution of the AI-aided epitope prediction. Early motif-based and matrix-based scoring approaches have progressively evolved into data-driven, sequence-based deep learning models trained on large-scale antigen presentation data sets. While sequence-based methods remain predominant, structure-based strategies that incorporate molecular and conformational features are receiving increasing attention and are expected to further enhance the predictive accuracy and generalizability (PDB code: 5F9J).

2.1. Prediction of Peptide–MHC Binding

MHC–peptide binding is a prerequisite for determining T-cell epitopes. Therefore, MHC–peptide binding affinity is a breakthrough point for T-cell epitope prediction. ,

Early epitope predictors were built on experimentally derived binding motifs and position-specific scoring matrices, capturing key anchor residues in MHC grooves. By the 2000s, neural network models (e.g., NetMHC) and allele-independent methods extended coverage across the polygenic MHC landscape and established a robust foundation for contemporary pMHC prediction. − The advent of LC–MS/MS has been a decisive advancement, as the availability of eluted ligand (EL) data has enabled predictive models to learn from physiological peptide libraries, thereby better reflecting AP and presentation. Crucially, LC–MS/MS has facilitated simpler identification of naturally presented ligands and markedly improved the accuracy of extracting peptide–HLA complexes from cellular and tissue samples. Building on this, Jurtz integrated MHC binding affinity data with EL data to develop NetMHCpan 4.0.

NetMHCpan 4.0 demonstrates excellent predictive performance in identifying naturally processed ligands, cancer neoantigens, and T-cell epitopes and achieves a higher area under the curve (AUC) on ligand data sets. Similarly, Keskin’s team profiled 186,464 ELs across 95 HLA alleles and leveraged these data sets to develop the HLAthena prediction model. In summary, integrating EL data to construct a model that more closely mimics real physiological conditions can better predict the information that T cells actually recognize on the cell surface.

Another way to improve prediction accuracy is to add descriptors such as AP to make computer simulation predictions match those of the in vivo antigen presentation process. Modern algorithms increasingly account for factors beyond the static MHC–peptide binding affinity, such as protease cleavage preferences, peptide transporter biases, and MHC allele expression, to predict whether a peptide will actually appear on cell surfaces. Xu et al. developed ImmuneApp by integrating embedding information from HLA-associated peptide repertoires, enabling it to perform antigen presentation prediction, immunogenicity assessment, and immunopeptidomic cohort data set analysis. Combining MHC class I binding affinity (BA) and AP predictors within a logistic regression model, O’Donnell et al. achieved at least a 40% increase in positive predictive value (PPV) across all comparisons. Wohlwend et al. likewise incorporated antigen processing information into their predictive model to enhance its performance. These results support a key point: integration of processing features can meaningfully improve the precision of epitope prioritization.

Rational selection of common MHC alleles is more cost-effective and efficient; whereas for individual applications, pan-allelic prediction is typically required, particularly in tumor immunotherapy. Subsequently to predict all MHC alleles, Nielsen et al. developed the pan-specific prediction computational method NetMHCpan, which can predict even for previously untested MHC molecules. However, the predictive performance of sequence-based (SeqB) methods heavily relies on large data sets, making accurate prediction for certain MHC alleles with limited data particularly challenging. The foundation of structure-based (StrB) prediction relies on the biochemical properties of the amino acids involved in peptide–MHC interactions and the structural information on the relevant pMHC complexes, thus endowing these methods with strong generalization capabilities. For example, compared with former pMHC prediction methods that relied on basic sequence encodings, OEDICHO embeds amino acid physicochemical properties to capture the physical drivers of peptide binding, which greatly enhanced prediction accuracy. With the innovations in protein structure prediction, such as AlphaFold, and advances in geometric deep learning, StrB approaches are increasingly emerging. , The inputs for machine learning algorithms typically comprise energy terms, statistical potentials, and structural descriptors. The 3pHLA-score designs input features for machine learning models by decomposing the energy contributions of individual residue positions in peptides, resulting in a per-residue protocol for predicting peptide binding affinity to HLA receptors. HLA-Inception predicts peptide-binding motifs through their correlation with the 3D electrostatic potential distribution of the MHC-I binding pocket. Motmaen et al. showed that a classifier placed on top of the AlphaFold network, based on binder/nonbinder classification, along with fine-tuning the combined network parameters, can construct a model with strong generalization capability, enabling it to generalize to novel biological systems. Marzella et al. developed an end-to-end geometric deep learning (GDL) approach that predicted peptide–MHC binding by directly analyzing structural features at the atomic or residue level. This method also demonstrated remarkable potential for data efficiency without relying on any binding affinity data. Epitope prediction based on sequence data (SeqB) is relatively mature and widely applicable owing to the easier accessibility and processing of such data compared to three-dimensional structural data. Compared with SeqB predictors, structure-informed models explicitly encode the 3D geometry and energetics of peptide–MHC complexes. Recent studies using large-scale modeled pMHC structures, geometric deep learning, and fine-tuned AlphaFold pipelines have demonstrated improved generalizability to unseen alleles, higher data efficiency, and enhanced interpretability for rational peptide vaccine design.

MHC class II ligands are typically 12–25 residues long, and the binding motifs of MHC class I and class II can be represented by a 9-amino-acid core. Using systematic single–amino-acid substitution scans, Hammer et al. identified the binding preferences of HLA-DRB1*0401–restricted peptides. They then encoded these preferences into a computational scoring program, establishing an early approach for predicting MHC class II–binding peptides. For the generally longer MHC class II peptides, accurate identification and alignment of the binding core are essential. Because the peptide-binding groove of MHC class II is open, precise identification of the binding core and accurate estimation of the binding affinity are particularly critical. Accordingly, peptide–MHC class II interactions have been predicted using in silico algorithms, including hidden Markov models (HMMs) and artificial neural networks (ANNs). Nielsen et al. proposed that prespecifying “anchor” positions within class II motifs can improve the convergence and predictive accuracy of Gibbs-sampling–based models. The motifs were inferred from peptide sequences known to bind HLA-DRB1*0401 in public databases such as SYFPEITHI and MHCPEP.

2.2. TCR–pMHC Binding Affinity Prediction

Only pMHC complexes that were recognized by TCRs with binding affinities within an optimal range can trigger effective T-cell activation. Thus, predicting immunogenic peptides requires the consideration of strong MHC class I binding and favorable TCR recognition. Predicting TCR–pMHC interactions enables the rational prioritization of high-affinity candidates from large-scale pMHC screens, considerably improving hit rates and reducing the cost of blind experimental verification. In addition, the prediction of TCR binding profiles can identify peptides with potential cross-reactivity to self-antigens, supporting the exclusion of unsafe candidates and enhancing vaccine safety. The diversity of the TCR repertoire, generated through random rearrangement of TCR α/β gene segments and refined through positive and negative selection in the thymus, poses a major obstacle. Current strategies, similar to pMHC prediction, rely on SeqB or StrB computational approaches to predict the TCR–epitope affinity. Tong et al. developed SETE, a method that leverages CDR3 sequences and their corresponding binding epitopes to train a Gradient Boosting Decision Tree model for the prediction of TCR epitopes. Montemurro et al. showed that training NetTCR-2.0 on paired TCR α- and β-chain sequences produced more accurate predictions than the original NetTCR models trained using only peptide sequences and CDR3β information. , The authors introduce TCRen, a StrB computational method that uses experimentally determined crystal structures or homology models to rank candidate epitopes for a given TCR. This approach relies on a statistical potential learned from residue-level contact frequencies by scoring TCR–peptide interfaces with this potential; TCRen improves the identification of “unseen” epitopes and mitigates limitations inherent to purely SeqB predictors. To develop a generalizable peptide–TCR binding predictor, Gao et al. combined meta-learning with a Neural–Turing–Machine-style external memory and formulated peptide–TCR recognition as a set of peptide-specific tasks. This design allowed the model to learn “how to learn peptide-specific recognition” strategies, enabling rapid adaptation in few-shot scenarios and true zero-shot generalization to previously unseen peptides. ,

2.3. Integrated Antigen Presentation Cascade Prediction

Effective T-cell activation requires engagement of the TCR with short antigen-derived peptides (P) presented by MHC molecules on the surface of APC. Elucidating the binding rules that govern interactions among peptides, MHC, and TCR is central to autoantigen discovery and rational vaccine design. Although numerous computational methods predict peptide–MHC (P–M) or peptide–TCR (P–T) interactions, recent work increasingly focuses on the full ternary peptide–MHC–TCR (P–M–T) complex. Lu et al. introduced pMTnet, a model for pairing TCRs with pMHCs that leverages transfer learning to integrate information from TCR sequences, antigenic peptides, and MHC alleles. Building on rapid advances in prediction methods, Zhao et al. introduced UniPMT, a unified deep learning framework that in combination learns three interrelated taskspeptide–MHC (P–M) binding, peptide–TCR (P–T) binding, and the ternary P–M–T interaction within a single model. UniPMT encodes peptides, MHC molecules, and TCRs as distinct node types in a heterogeneous graph; initializes peptide and TCR features with evolutionary scale modeling (ESM) embeddings; derives MHC pseudosequences using TEIM; and applies GraphSAGE to learn node representations. A deep matrix factorization module then combines the P–M and P–T representations to estimate P–M–T binding probabilities, achieving excellent performance across all three tasks. Similarly, Li et al. introduced Neo-intline, which integrates gene expression, proteasomal cleavage, TAP transport, MHC class I/II binding, and TCR recognition into an interpretable scoring framework. By modeling the antigen presentation pathway end-to-end, the method prioritizes neoantigens likely to traverse all presentation steps, thereby helping to bridge the gap between strong in silico performance and modest clinical efficacy.

3. Modification for Enhanced Vaccination Efficiency

Using AI-assisted prediction, optimized antigen epitopes are discovered, leading to improved vaccination precision and efficiency. The overall efficiency can be further enhanced by chemical modifications. Using various modification strategies, antigen peptides may achieve regulated immunogenicity to induce expected immune responses, enhanced stability for longer in vivo circulation time, and the ability for targeted delivery (Figure ).

3.

Three key modification strategies to enhance vaccine efficacy include immunogenic modulation, stability enhancement, and targeted delivery.

3.1. Regulation of Immunogenicity

Traditional peptide vaccines may have unsatisfactory immunogenicity to elicit a robust immune response. − To improve the immunogenicity, mainly two strategies are proposed, namely, modifications on amino acids or linking antigen peptide to another functional molecules. Amino acid modifications help regulate the intrinsic properties of peptides such as binding affinity and efficiency. Introducing molecules can easily achieve multifunctionality that regulates the in vivo biological process.

Bernardes and Corzana et al. used non-natural MUC1 glycopeptide to achieve enhanced immunogenicity. They introduced the (4S)-4-fluoro-l-proline (fPro) residue and S-glycosidic linkage into the traditional MUC1 glycopeptide sequence, eliciting robust humoral immune responses with the release of IgG2 and inflammatory cytokines. The substitution of O with S at the glycosidic linkage increased the distance between the sugar and peptide, inducing the formation of an optimized conformation for antibody recognition and binding. Replacement of Pro with fPro also contributes to enhanced interactions with the antibodies. Similarly, Gibadullin, Anguita, Fiammengo, and Corzana et al. used β-amino acids to regulate the property of MUC1 glycopeptide. Although the replacement of the immunodominant sequence APDTRP leads to limited binding affinity, using β-amino acids in the external region of the MUC1 glycopeptide showed enhanced immunogenicity and stability. However, β-amino acid substitution in the OVA peptide (SIINFEKL) fails to induce sufficient T-cell responses, where T cells that are specific to β-amino acid–substituted antigens exhibited no reactivity to native antigens. This further indicates that replacement should not occur on key sequences of the peptide. The post-translationally modified peptide can also serve as an immunotherapeutic target. Hunt and Malaker identified several antigen peptides with post-translational modifications, including phosphorylation, O-GlcNAcylation, methylation, and kynurenine. Various modifications of the antigen peptide can greatly affect the MHC-binding affinity, thus regulating immunogenicity. Wang et al. introduced unnatural amino acids into the target antigen to select antibodies with high affinity. Through chemical cross-linking with an introduced unnatural amino acid, antibodies were collected and identified. This strategy may be further applied to target antigens with poor immunogenicity.

Constructing vaccines with multiple components has shown promise in peptide-based vaccination therapies. Antigen linked with other epitopes or adjuvants can markedly increase the therapeutic efficiency. Li and Chen et al. synthesized a peptide with CD8 and CD4 epitopes. This bisepitope vaccine enhances the interactions between DCs, CD4⁺ T cells, and CD8⁺ T cells. The efficiency was also validated in prophylactic and therapeutic tumor models. Collier et al. further designed nanofiber peptide vaccines containing CD4, CD8, and B epitopes to achieve comprehensive immune activation. They linked the antigen peptide to Coil29 (QARILEADAEILRAYARILEAHAEILRAD), which has the ability to self-assemble to α-helical nanofibers. As a result, nanoparticles with various epitopes on their surfaces were prepared. Antigens with B epitope promote the secretion of antibodies, thus facilitating antibody-dependent cellular cytotoxicity and phagocytosis and enhancing their effects.

Except for traditional T-cell or B-cell epitopes, combination with antigens for other immune cells also showed promise. Painter and Hermans et al. used α-GalCer, an agonist for NKT cells, to induce acute NKT responses and further enhance T-cell activation. The coadministration of NKT agonists and T-cell antigens notably improves immune responses, owing to the cytokines secreted from NKT cells. Similarly, Painter and Compton developed an α-GalCer-peptide system for enhancing immunogenicity. They optimized the linking strategies with a succinct and effective synthetic route. Guo and Zhang et al. identified the potential role of γδT cells in treating cancer. They revealed the signaling process of how phosphoantigens activate γδT cells. Yin, Liu, and Chen et al. further used designed nanovaccines to activate γδT cells, resulting in reduced tumor growth.

Binding antigens to adjuvants is a more common strategy for the regulation of the immunogenicity of vaccines. Chen et al. used mannan with dodecyl chains to act as an adjuvant and carrier. The modified alkyl chain can assemble into nanoparticles, where the OVA peptide was encapsulated. Mannose and dodecyl chains serve as agonists for Dectin-2 and TLR4 respectively, resulting in the abundant release of pro-inflammatory cytokines. Liu and Sun et al. constructed a framework of nucleic acids for regulating the epitope spacing of linked antigens. Frameworks with various sizes and epitope spacings were screened to obtain the optimized immunogenicity. Moreover, the TLR9 agonist CpG was also bound on the framework as a stimulant. Since interacting with multiple epitopes facilitates B-cell activation, the authors further validated their design on COVID-19 peptide vaccines, and an obvious increase in IgG secretion could be found. Zhu, Yin, and Li et al. also prepared peptide–nucleic acid conjugate for enhanced vaccination. Propargyl sulfonium was used as the key linker of peptide and nucleic acid, which was reduced after being exposed to an intracellular reduction environment, leading to the release of antigens and adjuvants in APC. Sun and Ding et al. synthesized a series of fluorinated supramolecules with arginine and fluorinated diphenylalanine peptide to regulate the binding of antigen and adjuvant efficiency. Optimized supramolecules were screened for enhanced lysosomal escape of antigens.

In autoimmune or related diseases, inducing regulatory immune responses is required, which can also be achieved by peptide modifications. Using the clinical biomarkers of rheumatoid arthritis, anticitrullinated peptide antibodies, Wang, Hu, and Li et al. modified antigen peptide with citrulline. The generation of anticitrullinated peptide antibodies in patients with rheumatoid arthritis is caused by the peptidyl arginine deiminase–based conversion of arginine residues into citrulline, which triggers autoimmune responses. Thus, the authors selected rheumatoid arthritis–related autologous peptides with citrulline modification to induce antigen-specific immune tolerance. Zuo and Yang et al. also focused on rheumatoid arthritis and used autoantigen type II collagen peptide and the immunomodulator leflunomide to induce dendritic cell and regulatory T-cell tolerance. Kim et al. combined Aβ peptide with rapamycin, successfully inducing anti–Aβ antibodies and Aβ-specific regulatory T cells. Reduced neuroinflammation and Aβ plaques were observed, indicating their potential as a novel therapeutic strategy.

3.2. Enhancing Stability

The stability of peptide vaccines is one of the most important challenges faced in clinical applications. Thus, normally, multiple dosing or invasive administration is used. Effective peptide modifications can enhance the resistance to enzymatic degradation and circulation time, resulting in higher bioavailability.

Synthesizing stapled peptides is a common method for improving stability. The conformationally rigid structure and the introduced α-helix structure enable the high proteolytic stability. Various stapling strategies have been developed. However, the formation of a rigid structure may limit antigen peptide–receptor interactions, leading to “undruggable” vaccines. To solve this limitation, Hossain and Bathgate et al. designed a noncovalent stapling strategy. They used the unnatural amino acid α-methyl-l-phenylalanine to induce the formation of the α-helix structure. The α-methyl group exerts steric constraints on the backbone to form helix-favored dihedral angles, whereas π–π interactions between the two phenyl rings further lead to a noncovalently stapled helical peptide. The biological function of an unnatural amino acid–substituted protein can also be fully mimicked.

Utilizing a stapled peptide rather than stapling the antigen peptide may also help improve the stability. Li and Yin et al. used a sulfonium-based stapling peptide as a carrier for tumor immunotherapy. Neoantigen peptides were linked with agonist CpG by sulfonium and coassembled with a stapling peptide, forming stabilized nanoparticles. The interactions between the positive charge on the stapled peptide and the negative charge on the phosphoric acid skeleton in CpG contribute to assembly and stabilization. Gan, Li, and Yin et al. further simplified their stapling peptide system by validating the adjuvant properties of a positively charged sulfonium stapling peptide. The stapled peptides directly coassembled with the antigen peptide and contributed to the internalization of the antigen peptide as well.

Considering that the instability of antigen peptide is mainly caused by enzymatic degradation, introducing unnatural amino acids is another effective strategy. Gellman et al. replaced an α-amino acid residue with a β-amino acid residue and evaluated the influence on biological function. Although many of these substitutions limit antigen-triggered immune responses, several antigens with unnatural amino acids showed similar or enhanced activity. Gibadullin, Anguita, Fiammengo, and Corzana et al. also used the α-to-β substitution method in MUC1 glycopeptide, as described before. Their study suggested that introducing a β-amino acid to a less notable segment for receptor recognition can be another promising strategy. Miles and Sewell et al. developed a D–amino acid–based strategy. This synthesized peptide showed resistance to physical and enzymatic degradation, as well as great immunogenicity. Gellman et al. replaced several amino acids with thioamides, achieving enhanced resistance to proteinase. The MHC binding and T-cell–activating ability can be retained or enhanced by multiple modifications.

3.3. Targeted Delivery

Targeted delivery has been widely investigated in peptide vaccine-based therapy. Precise delivery of peptide vaccines can improve their accumulation in particular organs or APC and reduce their degradation in the circulatory system and thus can be considered an alternative method to improve the overall immunogenicity and stability of the peptide vaccine system.

Cell-penetrating peptides (CPPs) are widely used in peptide delivery strategies, which have the ability to promote cell internalization and antigen presentation. Irvine et al. tested the function of several CPP-linked tumor antigen peptides. Two effects were identified to promote the antigen presentation process: enhanced cell internalization based on the abundant positive charges of CPPs and antigen peptide accumulation in draining lymph nodes, which is related to binding to lymph-trafficking lipoproteins and preventing proteolytic degradation. Using antibodies or antibody derivatives such as nanobody or single-chain variable fragment (scFv) is another feasible method for delivery to APC. Irvine and Pentelute et al. utilized scFv that recognizes chemokine receptor 1 (XCR1), a unique receptor on cDC1 cells, to facilitate antigen presentation. Two components of the anthrax delivery system were then used: protective antigen and the N-terminus of lethal factor, which was linked to scFv and antigen peptides. Triple mutant protective antigen was linked to scFv, targeting to cDC1 cells and oligomerizing into a heptameric prepore. This complex then promoted the cytosolic delivery of the N-terminus of lethal factor-linked antigen peptide.

Nanomaterials also play a remarkable role in peptide vaccine delivery. , Collier et al. used coiled-coil self-assembling peptide nanofibers to achieve mucosal immune activation by peptide vaccines. A 29-amino acid peptide Coil29 was used, which can self-assemble into nanofibers with a helical structure, and antigen peptides were linked to Coil29. Further modification of the random sequence of proline, alanine, and serine can decrease mucin complexation and increase the penetration of epithelia, leading to enhanced immune responses. They further used their proline, alanine, and serine modification strategy to improve the oral delivery of Q11-based peptide vaccines, resulting in promising results. Direct binding of targeting molecules on the surface of the nanoparticles is also a useful strategy. Using the bio-orthogonal click reaction of azide and dibenzocyclooctyne, Liu and Hong et al. performed antigen peptide binding to β-1,3-glucan particles. Combining with agonists such as PolyI:C and CpG, the antitumor effects were validated in various tumor models. Modified peptides with aliphatic chains to incorporate them into lipid nanoparticles are also commonly used.

Other strategies are also developed for peptide vaccine delivery. Stephenson et al. identified a cyclic decapeptide incorporating lipoamino acids to deliver various peptides. Mixing with cyclic peptides and lipoamino acids showed increased lymphatic targeting and antigen internalization. Moreover, these peptides can act as adjuvants for TLRs to further facilitate immune responses. Hammond et al. fused an antigen peptide to the cytosolic domain of the stimulator of interferon genes (STING) protein for promoting targeted delivery and immunogenicity. This modified STING protein was then coassembled with cGAMP, an agonist of STING, forming a peptide-STING-cGAMP tetramer with the ability for lymphatic trafficking.

4. Conclusions

This review first outlines the conceptual advantages and intrinsic limitations of peptide vaccines relative to traditional and mRNA platforms, emphasizing their safety, manufacturability, and epitope precision alongside challenges of weak immunogenicity, HLA restriction, and rapid clearance. It then surveys AI-enabled prediction of peptide–MHC binding, TCR–pMHC recognition, and integrated antigen-presentation cascades, spanning early motif- and matrix-based tools, modern sequence-based deep learning models, and emerging structure-informed or geometric approaches that generalize to data-sparse alleles. Building on these design advances, this review summarizes how chemical modifications can tune vaccine performance, including modulation of intrinsic immunogenicity, coordination of T-cell, B-cell, and innate responses, improved peptide stability, and targeted delivery that directs antigens to APCs or specific tissues. Collectively, these developments illustrate a coherent design pipeline in which AI narrows the epitope space to high-value candidates, and chemistry further optimizes their presentation, persistence, and functional immune outcomes across cancer, infectious, and autoimmune indications.

5. Discussion

Despite rapid progress, several key gaps remain before AI- and chemistry-enabled peptide vaccines can fully realize their translational potential.

On the computational side, current pMHC and TCR–pMHC predictors are largely trained on a limited subset of HLA alleles, tumor types, and experimental settings, which constrains their ability to generalize to rare alleles, nonhuman MHCs, and complex clinical contexts. Additionally, a natural future direction for peptide-vaccine epitope prediction is to move beyond unmodified sequences and explicitly model post-translationally modified (PTM) ligands. Immunopeptidomics studies show that phosphorylated, glycosylated, and other PTM-bearing peptides are frequently presented by MHC molecules. − However, current predictors largely ignore PTMs and operate in the 20-amino acid space, leading to systematic blind spots in identifying PTM-derived epitopes. The recent development of NetMHCphosPan, a pan-specific model trained on phosphopeptide immunopeptidomics data to predict MHC-presented phospholigands with broad allele and length coverage, illustrates how integrating PTM-aware training data can bridge this gap for at least one modification type. Parallel chemical biology work that installs PTM-mimicking modifications on synthetic peptides and systematically quantifies their impact on MHC binding and TCR recognition provides precisely the kind of labeled data needed to generalize these approaches. Going forward, combining large-scale PTM immunopeptidomes with PTM-mimetic libraries and structure-informed AI models should enable prediction frameworks that treat PTMs as first-class design variables, allowing vaccine developers to deliberately exploit phosphorylated or otherwise modified epitopes as part of multiobjective peptide-vaccine design.

In addition to T-cell epitope prediction, machine learning–based approaches have been increasingly applied to B-cell epitope identification, leveraging either protein sequence information or structure-derived features. Because most antibody epitopes are conformational, B-cell epitope prediction is generally more challenging than peptide–MHC binding prediction. Sequence-based models can leverage machine learning and representation learning to improve feature encoding and make more effective use of antigen sequence information. For example, deep neural networks for linear epitope prediction (e.g., EpiDope) and frameworks that incorporate protein language model representations (e.g., BepiPred-3.0) generally show more robust performance in benchmark evaluations. − Nevertheless, sequence-only predictors face a key limitation because conformational epitopes are defined by spatially proximal residues on the antigen surface, which makes three-dimensional surface determinants difficult to infer solely from sequence information. To address this gap, structure-informed approaches use antigen three-dimensional structures as input and prioritize surface regions that are more likely to be bound by antibodies, which can help capture the geometric nature of conformational epitope recognition. These approaches can be broadly divided into two categories. One category consists of antibody-agnostic, structure-based predictors, such as DiscoTope and GraphBepi. , These tools require antigen structure and are well suited for scanning the epitope propensity when information about a specific antibody is unavailable. However, due to the absence of antibody-specific context, these predictors may yield relatively broad candidate regions in multiantibody settings, which can limit localization precision. The other category comprises structure-driven, antigen–antibody complex–based approaches that explicitly model antigen–antibody recognition, thereby aligning predictions with real-world conformational epitope mapping. For example, antibody-specific epitope mapping frameworks (e.g., AbEMap) combined with AlphaFold3-generated antigen–antibody complexes have demonstrated improved performance relative to alternative tested strategies in benchmark evaluations. In parallel, machine learning is increasingly being used to enhance the interpretability and practical applicability of epitope mapping, with the potential to address biological challenges such as antibody polyreactivity and the ambiguity of conformational epitope definitions. With the integration of structure prediction, complex modeling, and high-throughput epitope mapping data, model interpretability and translational utility are expected to further improve, enabling more direct applications in antibody discovery and vaccine immunogen design. ,

Many chemical modifications, such as stapling, noncanonical residues, multicomponent nanofibers, and tolerogenic conjugates, have demonstrated impressive activity in preclinical models, but their manufacturability, regulatory acceptability, and long-term safety in humans have yet to be systematically evaluated. Rational design will increasingly require closed-loop pipelines in which AI-guided epitope selection, high-throughput chemical diversification, and quantitative in vivo readouts iteratively refine vaccine candidates. In addition, expanding from purely immunogenic formulations to precision immunomodulators capable of inducing regulatory responses in autoimmunity or reshaping the tumor microenvironment will demand a closer coupling of epitope prediction with system-level immunology. Ultimately, the convergence of advanced prediction algorithms, immunopeptidomics, and medicinal chemistry should enable peptide vaccines that are population-tailored, disease-specific, and dynamically adaptable to viral evolution, tumor heterogeneity, and individual immune histories.

§.

X.-X.W. and J.-Y.S. contributed equally. The manuscript was written through contributions of all authors. CRediT: Xiao-Xue Wang visualization, writing - original draft, writing - review & editing; Jing-Yun Su visualization, writing - original draft, writing - review & editing; Yan-Mei Li funding acquisition, supervision, visualization, writing - original draft, writing - review & editing.

The authors declare no competing financial interest.