Computational prediction of TCR cross-reactivity: principles, challenges and translational opportunities Authors: Wenzhen Li, Yu Pan, Xu Wu, Yuting Song, Yin Wang, Lu Xie Published: 2026 ABSTRACT AbstractT-cell receptor (TCR) cross-reactivity, whereby a single TCR recognizes multiple peptide–MHC (pMHC) ligands, is essential for immune surveillance but also represents a major safety challenge for TCR-based therapeutics because of severe off-target toxicities. Rapid advances in immune-repertoire sequencing, structural biology, and immunopeptidomics have accelerated computational efforts to characterize cross-reactive recognition at scale. However, most existing methods are adapted from conventional TCR–pMHC specificity prediction and remain insufficiently aligned with the one-to-many, structurally plastic, and context-dependent nature of cross-reactivity. Here, we review the main data resources supporting TCR cross-reactivity modeling, including sequence- and structure-centric databases, and highlight persistent limitations, such as scarce explicitly annotated cross-reactive TCRs, strong biases toward viral epitopes and common HLA alleles, limited paired αβ-chain information, and a lack of rigorously validated negative examples. We then compare the major computational paradigms—sequence-based, structure-based, machine-learning, and multimodal approaches—with emphasis on their respective strengths and limitations in interpretability, generalization, and scalability. Finally, we discuss emerging translational applications and outline key priorities for the field, including dedicated datasets, rigorous benchmarking, and biologically grounded multimodal models to enable safer and more clinically actionable TCR-based immunotherapies.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12967-026-08476-5. Introduction T cell receptors (TCRs) are heterodimeric glycoproteins, most commonly composed of αβ chains in conventional T cells, although γδ TCRs constitute another major TCR lineage with distinct recognition properties. In the context of peptide–MHC (pMHC)-restricted immunity, αβ TCRs recognize pMHC complexes through their complementarity-determining regions (CDRs), with CDR3 playing a central role in antigen specificity [1, 2]. Although αβ TCR recognition has traditionally been viewed as highly specific, it is now well established that individual αβ TCRs can be cross-reactive and recognize multiple distinct pMHC ligands [3]. This property enables a finite αβ TCR repertoire to survey an immense antigenic landscape and supports protective antiviral and antitumor immunity. However, cross-reactive recognition does not by itself imply autoimmune pathology or off-target toxicity; its pathological consequences depend on the surrounding tolerance and activation context, including thymic selection, peripheral tolerance mechanisms, inflammatory cues, antigen abundance, and, in therapeutic settings, engineered changes in TCR affinity or signaling. Recent advances in immune tolerance biology further support this context-dependent view, highlighting peripheral immune tolerance and regulatory T cells as central mechanisms that prevent immune cells from attacking the body’s own tissues [4, 5]. Mechanistic studies over the past decades have revealed the molecular basis of TCR cross-reactivity. As summarized in Fig. 1, three major but non-mutually exclusive mechanisms can contribute to degenerate TCR recognition: conformational plasticity of the CDR3 loop during engagement with different pMHC ligands (Fig. 1a) [6], shared physicochemical motifs among distinct peptides that enable molecular mimicry (Fig. 1b) [7], and peptide sequence degeneracy within MHC-restricted TCR recognition, in which selected residue substitutions can be accommodated without abolishing recognition (Fig. 1c) [8]. Given the vast size of the pMHC universe relative to the limited responsive TCR repertoire, some degree of cross-reactivity is considered a biological necessity. Together, these mechanisms indicate that cross-reactivity cannot be explained by sequence identity alone, but instead reflects structural adaptability, motif-level similarity, and context-dependent recognition. They therefore motivate computational approaches that attempt to approximate flexible TCR recognition, although current models can capture CDR3 loop plasticity, conformational rearrangement, and induced-fit effects only in part. Molecular and structural determinants underlying TCR cross-reactivity. (a) CDR3 structural plasticity enables the same TCR-anchored CDR3 loop to adopt alternative conformations over different pMHC ligands, without implying physical circularization or loop closure. (b) Shared peptide motifs allow distinct peptides with similar sequence or physicochemical features to be recognized by the same TCR through molecular mimicry. (c) Peptide sequence degeneracy within MHC-restricted TCR recognition allows selected residue substitutions to be tolerated at the TCR–pMHC interface without abolishing recognition A wide range of experimental strategies has been used to characterize TCR cross-reactivity (Fig. 2). Figure 2 summarizes these approaches across three complementary layers, from traditional screening assays to high-throughput ligand discovery and downstream computational analysis. Traditional screening approaches can be further divided into pMHC binding-based assays, peptide-stimulation functional assays, and APC/target-cell co-culture assays. pMHC tetramer or multimer staining uses peptide-loaded recombinant MHC molecules to detect TCR–pMHC binding, whereas peptide-stimulation assays, including ELISA, CBA/Luminex, ELISpot, FluoroSpot, intracellular cytokine staining, and activation-marker assays, measure T-cell activation outputs induced by defined peptides [9, 10]. APC or target-cell co-culture assays further evaluate antigen presentation or naturally processed antigen recognition in a more cellular context, but they remain relatively low-throughput and limited in the scale of antigen testing [11, 12]. Recent high-throughput platforms—including combinatorial peptide libraries [13–17], multiplexed tetramer technologies [18], single-cell sequencing [19], and mass-spectrometry-based immunopeptidomics [20–22]—have substantially expanded the accessible ligand space, providing systematic views of potential cross-reactive interactions. These experimental datasets also provide the foundation for structural modeling [23–25] and deep learning-based prediction frameworks [26], which have further enriched mechanistic understanding of TCR–pMHC recognition. Experimental and computational platforms for profiling TCR cross-reactivity. Traditional screening is divided into pMHC binding-based assays, peptide-stimulation functional assays, and APC/target-cell co-culture assays. High-throughput platforms expand ligand discovery, while systematic analysis and prediction methods use these datasets for structural modeling and machine-learning-based cross-reactivity assessment These insights provide the basis for linking molecular recognition to functional outcomes [27]. The approval of tebentafusp, a soluble TCR-based therapeutic product, marks a milestone in the development of TCR-targeted drugs [28, 29]. High-profile clinical observations further underscore the need to evaluate off-target recognition during therapeutic TCR development. A prominent example is the MAGE-A3/Titin toxicity case, in which an affinity-enhanced TCR recognized a Titin-derived peptide and caused severe cardiovascular toxicity [30]. This event highlights that cross-reactive or off-target recognition in therapeutic settings may be introduced or amplified by TCR engineering, rather than reflecting natural physiological TCR degeneracy alone [31, 32]. Recent therapy-oriented reviews have summarized TCR-T development in solid tumors, including target selection, TCR engineering, toxicity, and resistance mechanisms [33]. However, these discussions have generally treated cross-reactivity primarily as a safety issue within TCR-T development rather than as an explicit computational prediction problem. Direct characterization of cross-reactivity remains low-throughput, labor-intensive, and difficult to scale to the vast peptide universe relevant to immune recognition [34, 35]. This limitation has made computational modeling increasingly important for extending cross-reactivity analysis into high-dimensional antigen spaces. In this context, it is important to distinguish cross-reactivity-specific prediction from conventional TCR specificity prediction: whereas specificity prediction is typically framed as pairwise classification of a TCR against a single peptide–MHC complex, cross-reactivity-specific prediction requires models to identify, rank, or recover multiple ligands that may be recognized by the same TCR within a defined candidate space, often under a fixed or controlled HLA context. Accordingly, the central challenge is not merely to enumerate multiple putative epitope associations for a given TCR, but to assign biologically meaningful confidence to candidate cross-reactive ligands while controlling for database-driven sequence similarity, assay heterogeneity, incomplete HLA information, negative-set uncertainty, and annotation bias. Therefore, rather than reviewing TCR-T development in general [33, 36], the present Review focuses specifically on TCR cross-reactivity as a distinct computational task, with particular emphasis on task definition, data resources, benchmark design, negative-set construction, and translational safety assessment. In this review, we summarize the data resources that support computational modeling of TCR cross-reactivity and systematically compare prediction paradigms spanning sequence-, structure-, machine-learning–, and multimodal-based approaches. We further outline current limitations and highlight future directions for developing interpretable, generalizable, and clinically deployable models. Data resources for TCR cross-reactivity Advances in high-throughput immune-receptor sequencing (AIRR-seq) [37], structural biology, and immunopeptidomics have led to specialized databases that support training and benchmarking of computational models for TCR cross-reactivity. These resources fall into two complementary categories: sequence-centric databases, which provide TCR–peptide–MHC annotations for pattern discovery and machine learning; and structure-centric databases, which offer atomic-resolution insights for mechanistic modeling (Table S1). Sequence-centric databases Sequence-centric resources form the earliest and most widely used data foundation for modeling TCR recognition. Early curated resources, such as McPAS-TCR [38] and IEDB (updated to 2025) [39], established catalogs of TCR–pMHC interactions but were limited by incomplete αβ pairing, heterogeneous annotation, and polyclonal or bulk-sequencing noise. Expanded repertoire databases, including PIRD [40] and UcTCRdb [41], broadened the available sequence space by integrating large-scale repertoires and unconventional T cell subsets, but continued to suffer from inconsistent metadata and sparse functional annotation. More recent curated resources—notably TCRdb2.0 [42] and NeoTCR [43]—have improved the standardization and contextual annotation of TCR sequence resources. TCRdb2.0 provides large-scale curated TCR sequence records with standardized processing pipelines, focusing on β, γ, and δ chains while excluding the α chain. NeoTCR, in contrast, is a neoantigen-oriented resource that includes TCR–pMHC triplets and paired αβ-chain information for a subset of records, but it is not exclusively composed of paired-chain entries. Despite these advances, key gaps remain, including scarce negative examples, limited explicit cross-reactivity labels, incomplete HLA and paired-chain coverage, and incomplete harmonization across sources. Structure-centric databases Structure-centric databases provide mechanistic grounding by describing the TCR–pMHC interface at atomic resolution. CrossTope [44] pioneered structural labeling of cross-reactivity cases, but its limited size and restricted HLA coverage constrain modern applications. Subsequent databases such as ATLAS [45] and STCRDab [46] expanded structural coverage and enabled analyses of conformational plasticity, yet lacked kinetic and functional annotations needed to connect structure with recognition dynamics. More recent integration efforts, including TCR3d [47] and VDJdb [48], standardized structural annotations and began bridging sequence–structure representations. However, we estimate that the direct overlap between currently available structural and sequence-centric resources remains very limited (likely below 5%), depending on how overlap is defined—for example, by shared TCR identifiers, cognate peptide–MHC annotations, or explicit mapping to experimentally resolved PDB complexes. This limited overlap creates a major bottleneck for multimodal modeling and restricts the ability of deep learning frameworks to learn mechanistically grounded representations. Thus, while structural resources offer critical mechanistic insights, their limited scale and poor interoperability with sequence datasets prevent their systematic integration into predictive pipelines. Conceptual integration and emerging multi-dimensional resources Sequence- and structure-centric resources provide complementary but fragmented views of TCR cross-reactivity, spanning population-level repertoire patterns and atomic-scale recognition mechanisms. Emerging multi-dimensional datasets such as TRAIT [49] integrate sequence, structural, and experimental evidence, representing an important step toward unified modeling frameworks. Importantly, most existing repositories reflect experimental and clinical convenience rather than an unbiased representation of the immunological recognition landscape. Current datasets are heavily enriched for viral epitopes, common HLA alleles, and well-characterized therapeutic targets, while self-antigens, rare HLAs, and subdominant peptide repertoires remain systematically under-sampled. This selection bias not only limits biological inference but also propagates into computational models, which may learn dataset-specific shortcuts rather than generalizable principles of TCR cross-reactivity. More fundamentally, existing databases were constructed around single-specificity paradigms and do not encode cross-reactivity as a first-class biological entity. As a result, annotations are often descriptive rather than mechanistic, lacking standardized ontologies for degeneracy, polyreactivity, or molecular mimicry. Furthermore, the presence of multiple database-associated epitopes for the same or similar CDR3 sequence should not be interpreted as validated cross-reactivity without considering HLA restriction, paired-chain information, assay context, clonal origin, and functional evidence. Consequently, the field lacks dedicated cross-reactivity–oriented datasets with harmonized annotation, validated negative controls, mechanistic metadata, and standardized benchmarking protocols—representing a central unmet need for developing interpretable and clinically deployable predictive models. Computational models for predicting TCR cross-reactivity Computational prediction of TCR cross-reactivity has evolved alongside advances in repertoire sequencing, structural biology, and machine learning [50, 51]. Importantly, most currently available computational tools were originally developed for pairwise TCR specificity prediction, in which a model evaluates a single TCR–peptide–MHC interaction. In contrast, cross-reactivity prediction aims to identify or prioritize multiple ligands that may be recognized by the same TCR within a candidate peptide space, and many existing approaches therefore adapt specificity predictors as retrieval, ranking, or safety-screening tools rather than being explicitly designed for cross-reactivity. Early models focused on scalable sequence-based representations, whereas later frameworks incorporated structural and multimodal information to better approximate the multi-scale determinants of one-to-many recognition. These developments have led to four major modeling paradigms reviewed below. We therefore evaluate each paradigm not only by its ability to model pairwise TCR–pMHC associations, but also by whether it supports one-to-many ligand retrieval, confidence ranking, bias control, and downstream experimental validation. Sequence similarity-based approaches Sequence similarity–based approaches represent the earliest computational attempts to infer TCR cross-reactivity (Fig. 3; Table S2). As illustrated in Fig. 3, these methods typically quantify similarities among TCR CDR3 sequences, among candidate peptide sequences, or between TCR–peptide pairs through motif comparison, substitution-matrix–based scoring, or sequence embedding, followed by clustering, ranking, or classification to identify potentially cross-reactive ligand sets. Figure 3 is intended to illustrate a representative CDR3β-based sequence analysis workflow, because many current motif- and distance-based methods primarily rely on CDR3β sequences, reflecting the greater availability of β-chain data than paired αβ-chain information in public TCR databases. These methods assume that TCRs with similar CDR3 motifs may recognize overlapping peptide sets, enabling repertoire-scale clustering and rapid exploration of antigenic landscapes. However, β-chain-only analyses provide an incomplete approximation of TCR specificity and cross-reactivity, because the α chain and paired αβ context can also contribute to peptide and MHC contacts. Therefore, CDR3β-based motif prediction should be interpreted as a candidate-prioritization strategy rather than a complete representation of αβ TCR–pMHC recognition. Sequence similarity–based workflow for TCR cross-reactivity-oriented candidate prioritization. This representative CDR3β-based workflow compares TCR and peptide sequences through motif matching, BLOSUM62-based similarity scoring, sequence embedding, clustering, and candidate cross-reactivity annotation Early distance- and alignment-based metrics (e.g., Levenshtein and substitution-matrix–weighted alignment) improved motif-level comparison by accounting for conservative substitutions and positional constraints, but remained incapable of modeling non-additive or high-dimensional interactions. In the illustrative example shown in Fig. 3, amino acid similarity is scored using BLOSUM62 as a representative substitution matrix. The choice of substitution matrix is an important methodological parameter in CDR3 distance calculations, because different matrices assign different scores to conservative and non-conservative substitutions and may therefore influence motif grouping, nearest-neighbor inference, and downstream cross-reactivity annotation. More recent representation-learning approaches embed CDR3 sequences into latent spaces to capture subtle sequence patterns [52], while region-weighted models prioritize solvent-exposed or TCR-facing residues to approximate physical interaction relevance [53]. Context-aware frameworks further incorporate biological metadata, such as tissue expression, to prioritize potential off-target risks [54]. Despite these refinements, sequence similarity–based methods remain intrinsically limited because key drivers of cross-reactivity—conformational plasticity, alternative docking geometries, transient hydrogen-bond networks, and shallow binding-energy landscapes—cannot be inferred from linear sequence information alone. These limitations are particularly relevant for β-chain-only workflows, which may miss α-chain contributions, paired-chain effects, and HLA-restricted constraints on peptide presentation and TCR–pMHC recognition. Consequently, these approaches are best suited for motif-level clustering, candidate prioritization, and exploratory repertoire analysis rather than mechanistically grounded prediction across diverse antigenic contexts. These limitations motivated the development of structure-based models that explicitly represent three-dimensional interface geometry and energetics. Structure-based approaches Structure-based approaches extend beyond sequence-level inference by explicitly modeling the three-dimensional TCR–pMHC interface (Fig. 4; Table S3). As illustrated in Fig. 4, these frameworks typically involve structural modeling of TCR–pMHC complexes, extraction of interface-level features such as residue contacts and energetic scores, and comparison of structural similarity to infer potential cross-reactive binding modes. By capturing geometric complementarity, electrostatic features, and interfacial energetics, structure-based approaches provide a more mechanistic view of cross-reactive recognition, particularly when sequence-divergent ligands generate similar TCR-facing surfaces or related interaction landscapes [24]. Structure-based framework for TCR cross-reactivity-oriented candidate refinement. Modeled or experimentally resolved TCR–pMHC interfaces are used to extract structural and energetic features, compare interface similarity, and generate clustered heatmaps or cross-reactivity network graphs Early structure-based studies relied largely on template-driven modeling. Homology modeling tools such as TCRmodel [55] and Modeller [56] enabled rapid construction of TCR–pMHC complexes from known structures, but their performance was limited by sparse templates and substantial CDR-loop variability. Because cross-reactivity often depends on subtle interface rearrangements rather than conserved global folds, these constraints reduced the reliability of template-based inference in many settings. Physics-based docking frameworks provided a more explicit treatment of binding geometry and energetics. Methods such as RACER [57] sampled alternative binding orientations and scored affinity-related landscapes, offering deeper mechanistic insight into how a TCR might accommodate distinct peptide ligands. However, docking remains computationally expensive, sensitive to starting conformations, and difficult to scale across large peptide–TCR search spaces, particularly when local loop flexibility and peptide-induced rearrangements must be considered. Recent advances in deep-learning–based structure prediction have substantially expanded the scope of structural inference. AlphaFold-derived frameworks [58] and TCRmodel2 [59] enable end-to-end prediction of TCR–pMHC complexes directly from sequence, improving structural coverage, geometric plausibility, and, in many cases, CDR-loop modeling. These models provide useful starting points for downstream interface analysis. However, they do not fully resolve the key structural challenges of TCR cross-reactivity. Most current predictors still generate a single dominant conformation and therefore cannot adequately represent conformational ensembles, induced-fit effects, or alternative docking geometries that often underlie cross-reactive recognition. A further concern is the risk of false confidence: model confidence scores do not necessarily guarantee that a predicted TCR–pMHC interface is biologically correct or unique, and interface-level inaccuracies may substantially alter inferred contact patterns or electrostatic complementarity [58, 60]. Deep-learning–based structural models should therefore be viewed as hypothesis-generating tools rather than definitive representations of cross-reactive binding mechanisms. Recent structure-based methods such as TCRen further extend this framework by combining modeled TCR–pMHC complexes with residue-level statistical potentials to rank candidate unseen epitopes for a given TCR. This illustrates how structure-based approaches can support not only complex reconstruction but also interface-aware epitope prioritization [61]. Beyond modeling individual complexes, another branch of structure-based approaches infers cross-reactivity by comparing structural or physicochemical features across interfaces. These methods are based on the idea that distinct peptides may still be cross-recognized if they generate similar TCR-facing landscapes in terms of shape, charge distribution, or local interaction potential. For example, MatchTope [62] quantifies electrostatic similarity across pMHC surfaces to identify ligands with comparable TCR-exposed features. TCRpcDist [63] further integrates geometric and physicochemical descriptors to detect structurally analogous binding modes among sequence-divergent complexes. Such approaches can reveal convergent recognition patterns that motif-based methods might miss. At the same time, structure similarity–based prediction inherits many limitations of the underlying structural models. Most methods rely on static structural snapshots and cannot represent dynamic binding pathways, transient interaction networks, or switching between alternative recognition modes [64, 65]. Similarity scores may also depend on the modeling protocol, alignment strategy, and surface representation. Structural similarity is therefore best interpreted as a mechanistically informative proxy rather than direct evidence of functional cross-reactivity. Overall, structure-based approaches provide an important bridge between molecular recognition and computational prediction. They are particularly valuable for explaining how sequence-divergent ligands may converge onto similar binding surfaces and for generating interface-level hypotheses for downstream validation. However, their broader application remains limited by computational cost, incomplete structural coverage, sensitivity to modeling assumptions, and the difficulty of representing conformational dynamics at scale. Thus, in cross-reactivity pipelines, structure-based approaches are currently most useful for mechanistic interpretation and refinement of prioritized candidates, rather than for exhaustive ligand-space screening. Machine learning approaches Machine-learning (ML) approaches extend heuristic sequence matching by learning nonlinear and high-dimensional patterns from repertoire-scale TCR–peptide–MHC data (Fig. 5). In the context of TCR cross-reactivity, however, an important distinction should be made between pairwise TCR–pMHC binding prediction and cross-reactivity-specific prediction. Many existing ML models are still trained and evaluated primarily as pairwise association or binding predictors, asking whether a given TCR is likely to recognize a given peptide–MHC complex. By contrast, cross-reactivity prediction requires identifying, ranking, and ultimately validating multiple potential pMHC ligands that may be recognized by the same TCR within a defined candidate space and HLA context. Therefore, current ML models are best interpreted as candidate-prioritization tools that can narrow the search space for experimental validation, rather than as definitive mechanistic predictors of one-to-many TCR recognition. Machine-learning framework for TCR cross-reactivity-oriented candidate prioritization. TCR and peptide sequences are preprocessed and encoded into engineered or learned representations, which are then analyzed by classical machine-learning or deep-learning models to generate interaction scores, cross-reactivity maps, and candidate TCR–peptide networks As illustrated in Fig. 5, typical ML pipelines include sequence preprocessing, feature encoding of TCR, peptide, and sometimes HLA or structural inputs, followed by algorithmic layers that generate interaction scores, ranked candidate ligands, or TCR–pMHC interaction networks. Recent tools that can be used or repurposed for cross-reactivity-oriented candidate prioritization are summarized in Table 1, whereas a broader catalog of related ML/DL models and resources is provided in Table S4. Conceptually, existing methods span feature-engineered models based on sequence similarity, physicochemical encodings, or structural energy proxies, as well as end-to-end deep representation learning models that embed TCR and peptide sequences directly. Representative examples include HLA-aware classical frameworks such as TCRGP [67], scalable deep-learning baselines such as NetTCR [72, 76], and geometry-informed formulations such as dGraph [77] that encode residue contacts and interface topology. Table 1Representative machine-learning methods that can be used or repurposed for TCR cross-reactivity prioritizationNameYearMethodInputData SourceCross-reactivity useRefERGO2020RNNCDR3β; peptideMcPAS-TCR, VDJdbAdapted[66]TCRGP2021Gaussian processCDR1/2/2.5/3(α/β)VDJdb, IEDB, McPAS-TCRAdapted[67]DLpTCR2021CNNCDR3α/β; peptideVDJdb, IEDBAdapted[68]TABR-BERT2023BERTCDR3β; MHC; peptideIEDB, VDJdb, McPAS-TCRAdapted[69]TEIM-Res2023CNN + few-shot pretrainingCDR3β; pMHCVDJdb, McPAS-TCR, COVID-19 datasets, IEDBAdapted[70]Pan-pep2023Meta-learning + neural Turing machine + attention + CNNCDR3β; peptideIEDB, VDJdb, PIRD, McPAS-TCRExplicit[71]NetTCR-2.22024CNNCDR1/2/3 (α/β);peptideIEDB, VDJdbAdapted[72]TouCAN2024CNN + contrastive learningCDR1/2/3 (α/β)VDJdb, IEDB, McPAS-TCRAdapted[73]Tspred2024CNN + attentionCDR1/2/3 (α/β);peptideVDJdb, IEDB, 10x, in-house datasetsAdapted[74]LightCTL2025CNN + contrastive learningCDR3β; peptide; MHCMcPAS-TCR, YFV, PIRD, VDJdb, IEDB, 10x, COVID-19, in-house datasetsAdapted[75]“Cross-reactivity use” indicates whether a method was explicitly designed or evaluated for cross-reactivity-related tasks (Explicit) or adapted from pairwise TCR–peptide/pMHC specificity prediction for candidate ranking or screening (Adapted). Model outputs should therefore be interpreted as prioritization hypotheses rather than definitive evidence of cross-reactive recognition. Representative machine-learning methods that can be used or repurposed for TCR cross-reactivity prioritization CDR1/2/3 (α/β); peptide CDR1/2/3 (α/β); peptide “Cross-reactivity use” indicates whether a method was explicitly designed or evaluated for cross-reactivity-related tasks (Explicit) or adapted from pairwise TCR–peptide/pMHC specificity prediction for candidate ranking or screening (Adapted). Model outputs should therefore be interpreted as prioritization hypotheses rather than definitive evidence of cross-reactive recognition. Different ML architectures provide complementary but incomplete views of TCR–pMHC recognition. Convolutional neural networks efficiently capture local motif-like patterns and are computationally scalable, making them useful for detecting shared motif-like signals associated with cross-reactive recognition [72]. Recurrent models can encode variable-length inputs and positional context, but their sequential inductive bias limits scalability and makes global interaction modeling more difficult in large, heterogeneous datasets [78, 79]. Transformer-based architectures improve representational capacity by modeling global dependencies and enabling more flexible integration of TCR, peptide, and contextual features [80, 81]. Attention mechanisms may highlight sequence regions associated with model outputs, but such signals should not be interpreted as causal biophysical determinants without external validation, because they may also reflect dataset artifacts or shortcut learning [74, 82]. Graph-based and geometry-aware networks move closer to interface-level modeling by encoding residue-level contacts, topology, or spatial proximity [77]. Recent structure-aware models such as SageTCR further extend this direction by integrating residue- and atom-level structural representations within a bi-level graph framework for TCR–pMHC binding prediction, illustrating how multi-scale structural encoding can strengthen interface-aware learning even when the task is not explicitly formulated as cross-reactivity prediction [83]. Nevertheless, geometry-aware models still depend on structural availability, docking quality, or reliable structural proxies, and their performance can be sensitive to how TCR–pMHC complexes are generated [84–86]. Despite their increasing predictive performance, current ML approaches remain limited in several respects. First, many models learn correlations from incomplete and heterogeneous datasets rather than explicit mechanisms of cross-reactive recognition. Key determinants such as CDR3 loop conformational plasticity, alternative docking modes, induced-fit effects, transient hydrogen-bond networks, and shallow binding-energy landscapes are rarely observed directly and are therefore captured only indirectly through sequence, static structural, or proxy features [87]. Second, apparent model generalization can be strongly influenced by dataset composition, including limited paired αβ-chain coverage, imbalanced peptide and epitope representation, heterogeneous assay definitions, and scarce experimentally validated negative examples. These factors can encourage shortcut learning and reduce robustness under distribution shift [88]. HLA diversity represents an additional and particularly important challenge for ML-based cross-reactivity prediction [88]. Public datasets are often dominated by a limited number of well-studied HLA alleles, whereas less common or population-specific HLA types remain sparsely represented [89]. As a result, models trained on overrepresented HLA contexts may not generalize reliably to underrepresented alleles, even when TCR or peptide sequence patterns appear similar. Future ML frameworks therefore require HLA-aware input representations, allele-stratified benchmarking, external validation across underrepresented HLA types, and calibrated uncertainty reporting when predictions are made outside well-sampled HLA contexts. Overall, ML models provide valuable tools for prioritizing candidate cross-reactive ligands, off-targets, and validation experiments, but their predictions should be treated as confidence-ranked hypotheses requiring functional, biophysical, or structural confirmation rather than direct evidence of true TCR cross-reactivity. Multimodal approaches for TCR cross-reactivity prediction Single-modality approaches are often insufficient for modeling TCR cross-reactivity, which emerges from the combined effects of sequence patterns, three-dimensional complementarity, antigen presentation, and cellular context. Multimodal frameworks therefore seek to integrate sequence, structural, and functional information to better represent the multi-scale nature of TCR recognition (Fig. 6). As illustrated in Fig. 6, these frameworks typically combine modality-specific encoders for sequence, structural, and functional data with cross-modal fusion strategies that integrate heterogeneous representations before generating interaction scores, cross-reactivity predictions, or candidate risk rankings. In addition to model architecture, Fig. 6 highlights a TCR–ligand interaction network as a useful representation of cross-reactive recognition. In such networks, TCRs and peptides or pMHC complexes can be represented as nodes, while experimentally observed or computationally predicted TCR–ligand associations are represented as edges. This network view is particularly useful for visualizing one-to-many relationships, in which a single TCR recognizes multiple peptide ligands, as well as many-to-one relationships, in which multiple TCRs converge on the same antigenic peptide. Sequence-only models cannot fully capture conformational plasticity or interfacial energetics, whereas structure-based and functional approaches each provide only partial views of cross-reactive recognition. These limitations have driven increasing interest in multimodal strategies that bridge molecular and cellular determinants of TCR recognition. Multimodal fusion framework for TCR cross-reactivity-oriented candidate prioritization. Sequence, structural, and functional or multi-omics features are integrated through cross-modal encoding or contrastive learning to generate cross-reactivity maps and TCR–peptide interaction networks A major branch of multimodal modeling combines sequence-derived representations with structural descriptors and repertoire-level information. Frameworks such as DeepAIR [90] integrate V(D)J usage, CDR3 embeddings, and structural cues to improve detection of cross-reactive signatures, whereas Multimodal-AIR-BERT [91] further highlights the contribution of paired αβ-chain information. Structural context has also been incorporated through template-aware models such as pMTnet-omni [92], which aims to improve generalization across less represented HLA contexts and species. In parallel, joint representation models such as TranspMHC [93] suggest that certain structural constraints may be captured implicitly from sequence-derived inputs, although the biological interpretability of these representations remains limited. These models illustrate how multimodal learning can move beyond isolated sequence similarity by incorporating paired-chain information, structural priors, and repertoire-level context, but they still depend heavily on the availability and quality of modality-specific annotations. Beyond sequence and structure, recent multimodal approaches have begun to incorporate higher-order interaction context and cellular state. Models such as TCR-Engine [94] explicitly encode the tripartite TCR–peptide–MHC relationship to better approximate physiological recognition. Other methods integrate T-cell phenotypic information, including single-cell transcriptomic features associated with activation, differentiation, or exhaustion states [95, 96]. These cellular-state features should be interpreted as modulators of the functional manifestation and detectability of cross-reactive recognition rather than determinants of intrinsic TCR specificity. Activation and differentiation states may influence signaling competence, cytokine production, effector output, and activation thresholds, thereby affecting whether a cross-reactive molecular interaction produces a measurable cellular response. By contrast, exhaustion is generally associated with impaired effector function, reduced cytokine production, altered transcriptional programs, increased inhibitory receptor expression, and decreased sensitivity to TCR-generated signals; it may therefore suppress, attenuate, or mask cross-reactive responses rather than create new recognition specificity or increase intrinsic TCR cross-reactivity [97–99]. These extensions move prediction beyond molecular binding alone toward functional immune readouts that may shape the observable consequences of cross-reactive recognition. In translational applications, multimodal integration has also expanded to multi-omics risk modeling. Tools such as iCrossR [100], sCRAP [101], and CrossDome [102] combine peptide features, HLA binding predictions, immunopeptidomics, and tissue-expression data to prioritize potential off-target liabilities in TCR-T design. Structural–biochemical similarity frameworks such as PepSim [103] further quantify convergent recognition patterns beyond direct sequence comparison. Collectively, these studies illustrate how multimodal integration extends cross-reactivity analysis across molecular, cellular, and translational levels (Table 2). In this setting, multimodal models are especially valuable for triaging candidate off-targets, linking molecular similarity to tissue-specific risk, and guiding experimental validation. Table 2Representative multimodal integration–based methods that can be used for TCR cross-reactivity or off-target risk prioritizationNameYearMethodInputData SourceCross-reactivity useRefiCrossR2017CR indices and tissue profilesPeptide; MHC allelePaxDB; NCBI RefSeq; NetMHCExplicit[100]sCRAP2021Rule-based scoring + conventional prediction toolsPeptide; MHC alleleImmunopeptidome databases of healthy tissuesExplicit[101]CrossDome2023Multi-omics integration + statistical modelingPeptide; HLA alleleHLA Ligand Atlas; HLAthena; IEDB; in-house datasetsExplicit[102]PepSim2023Structural and biochemical similarity analysisPeptide; pMHC complex structuresIEDB; CrossTope; in-house datasetsExplicit[103]“Cross-reactivity use” indicates whether a method was explicitly designed or evaluated for cross-reactivity-related tasks, off-target risk assessment, or cross-reactive ligand discovery (Explicit), or adapted from related peptide/pMHC similarity or specificity-prediction frameworks for candidate prioritization (Adapted). Model outputs should be interpreted as prioritization hypotheses requiring experimental validation rather than definitive evidence of functional cross-reactivity Representative multimodal integration–based methods that can be used for TCR cross-reactivity or off-target risk prioritization “Cross-reactivity use” indicates whether a method was explicitly designed or evaluated for cross-reactivity-related tasks, off-target risk assessment, or cross-reactive ligand discovery (Explicit), or adapted from related peptide/pMHC similarity or specificity-prediction frameworks for candidate prioritization (Adapted). Model outputs should be interpreted as prioritization hypotheses requiring experimental validation rather than definitive evidence of functional cross-reactivity Despite their promise, current multimodal models remain constrained by limited structural coverage, heterogeneous annotation depth, modality imbalance, and incomplete representation of conformational dynamics. Improved predictive performance therefore does not necessarily translate into mechanistic interpretability or robust generalization across datasets and biological settings. Moreover, it remains unclear whether different multimodal models converge on the same cross-reactive candidates or off-target risks, because current approaches differ substantially in their input modalities, feature weighting schemes, training datasets, label definitions, evaluation metrics, and benchmarking strategies [88]. This lack of standardization makes it difficult to homogenize, compare, or rank models across studies. Future work will require standardized benchmark datasets, harmonized evaluation metrics, shared negative-control definitions, calibrated uncertainty reporting, and comparative validation studies to determine whether different multimodal frameworks identify consistent and biologically meaningful cross-reactivity patterns. Multimodal integration should therefore be viewed as an important emerging direction in TCR cross-reactivity modeling, but not yet as a mature, standardized, or directly comparable solution. Taken together, existing modeling paradigms can be compared along several practical dimensions, including alignment with the one-to-many nature of cross-reactivity, mechanistic interpretability, confidence calibration, robustness under distribution shift, and suitability for translational safety screening. Across these dimensions, current methods reveal a consistent trade-off between scalability and biological realism: sequence-based and many deep learning approaches remain effective for retrieval and pattern recognition, whereas structure-based, geometry-aware, and multimodal models offer stronger biological grounding but remain constrained by static representations, limited structural coverage, uneven modality availability, and heterogeneous benchmarking. More broadly, the main bottlenecks arise not only from model architecture, but also from simplifying assumptions in task formulation, including pairwise classification, weakly defined negative sets, incomplete HLA and paired-chain annotation, and the absence of dynamic or functional supervision. Potential applications Computational modeling of TCR cross-reactivity has enabled applications spanning immunological discovery, immune-network analysis, neoantigen interpretation, and safety assessment of TCR-based therapeutics (Fig. 7). By integrating sequence, structural, and multimodal information, these frameworks have begun to support practical workflows for candidate prioritization, hypothesis generation, and safety triage. Potential applications of computational modeling for TCR cross-reactivity. Computational models support candidate epitope prioritization, immune-network mapping, neoantigen interpretation, and therapeutic safety assessment, with predicted cross-reactive candidates requiring experimental validation At the level of immunological discovery, cross-reactivity–aware sequence modeling supports epitope identification and repertoire exploration. Tools such as TCRMatch, originally developed for TCR specificity inference, have been used as filtering or plausibility-scoring modules in adversarial TCR sequence–generation pipelines. However, TCRMatch is HLA-agnostic and relies primarily on CDR3 sequence similarity, which limits its ability to represent the full TCR–peptide–MHC recognition context. Moreover, the immunological annotations used for discovery largely derive from the underlying IEDB records on which TCRMatch builds, rather than from TCRMatch itself. By constraining generated receptors to remain within biologically plausible similarity space, TCRMatch can help prioritize candidate epitope associations and support systematic exploration of antigenic landscapes (Fig. 7, upper left) [104]. Therefore, TCRMatch-based outputs should be interpreted as database-informed candidate associations rather than independently validated cross-reactive epitopes. Beyond individual epitopes, cross-reactivity modeling facilitates immune network mapping and comparative immunology. The dGraph framework embeds peptide and protein descriptors into a shared physicochemical latent space, enabling visualization of cross-reactive recognition patterns across viral families. By capturing structural and biochemical convergence among antigenic variants, dGraph supports reconstruction of immune recognition networks and tracking of antigenic drift (Fig. 7, upper right) [77]. Cross-reactivity prediction also contributes to the mechanistic interpretation of neoantigen immunogenicity. Structure-guided pipelines such as HLA-Arena evaluate differences between mutant and wild-type peptide presentation across relevant HLA alleles, quantifying changes in binding energetics, interface geometry, and TCR accessibility. Applications to KIF5B–RET fusion–derived epitopes and recurrent mutations in TP53, EGFR, and IDH1 in glioblastoma illustrate how these approaches can reveal mutation-driven shifts in antigen presentation that shape T-cell recognition (Fig. 7, lower left) [105, 106]. Despite their mechanistic value, structure-based methods remain difficult to deploy routinely in high-throughput cross-reactivity pipelines [58]. Accurate prediction of TCR–pMHC complexes remains challenging, particularly when cross-reactive recognition involves CDR3 loop flexibility, alternative docking geometries, peptide-induced conformational changes, or induced-fit effects [107]. In addition, practical deployment requires reliable ranking of alternative docking poses, benchmarking against experimentally validated cross-reactive and non-cross-reactive pairs, manageable computational cost across large candidate peptide spaces, and integration with sequence- or ML-based prioritization workflows [108]. Thus, structure-based modeling is currently most useful as a mechanistic refinement or validation layer after candidate prioritization, rather than as a standalone large-scale screening strategy. In therapeutic development, multimodal cross-reactivity prediction has become increasingly central to safety assessment and off-target risk evaluation. The Pan-pep dataset, a publicly released benchmark comprising thousands of TCR–peptide interactions, provides a standardized reference for peptide-level similarity assessment and baseline binding prediction. By enabling systematic comparison of on-target versus off-target recognition profiles, Pan-pep supports early-stage screening of cross-reactive liabilities during TCR engineering and benchmarking of predictive models used in TCR-T workflows (Fig. 7, lower right) [71]. In practice, these models are best viewed as prioritization and triage tools, with high-risk predictions requiring targeted experimental validation before clinical translation. Collectively, these applications show that TCR cross-reactivity modeling already functions as an enabling analytical layer linking molecular recognition, immune-system–level organization, and translational safety evaluation. Among these directions, safety screening and off-target prioritization appear most immediately actionable, whereas discovery-oriented applications remain more exploratory. Across all applications, predicted cross-reactive candidates should therefore be interpreted as hypotheses for prioritization and validation rather than definitive evidence of functional recognition. Challenges and future directions: a staged roadmap for TCR cross-reactivity prediction Although multimodal integration has emerged as an important direction in TCR cross-reactivity modeling, the field remains in transition from proof-of-concept prediction toward biologically grounded, robust, and clinically useful systems. Compared with conventional TCR–pMHC specificity prediction, cross-reactivity poses additional challenges because it involves one-to-many ligand recognition and determinants spanning molecular, cellular, and tissue contexts [109, 110]. In our view, the most urgent near-term need is not more complex architectures, but stronger data and evaluation foundations; longer-term progress will depend on biologically grounded multimodal frameworks built on rigorous benchmarks. Near-term priorities: data quality, negative controls, and benchmarking A fundamental limitation of current approaches lies in the scarcity and design of available datasets. Public repositories such as VDJdb, IEDB, and McPAS-TCR contain large numbers of TCR–pMHC records, yet they overwhelmingly represent single-specificity interactions rather than the multiple cognate ligands required to characterize cross-reactivity. Experimentally validated cross-reactive TCRs remain rare, high-throughput peptide screens cover only a small fraction of the antigenic landscape, and rigorously verified negative examples are largely absent [111]. Existing datasets are also strongly biased toward viral epitopes and common HLA alleles, while tumor neoantigens, self peptides, and rare HLAs remain underrepresented [109]. Even CrossTope—the only structural resource explicitly designed for cross-reactivity—remains too small and infrequently updated to support modern ML or multimodal frameworks. A further practical challenge is data attrition during multimodal integration [112]. As additional modalities are required—paired αβ-chain information, HLA restriction, peptide identity, structural models, assay metadata, tissue-expression profiles, immunopeptidomic evidence, and functional readouts—the number of usable entries can decrease sharply [113, 114]. This attrition is already evident in simpler annotation tasks: CDR3β-only matching against current antigen-annotated databases often leaves many repertoire entries unannotated, and coverage decreases further when HLA matching, paired-chain availability, or functional validation is required [38]. Thus, increasing biological realism through multimodal integration can paradoxically reduce annotation coverage and statistical power, creating a trade-off between predictive confidence and data retention. Given these constraints, including both limited cross-reactivity-specific data and modality-dependent data attrition, the most actionable short-term goal is to establish datasets and benchmarks explicitly designed for cross-reactivity analysis. Priority should be placed not only on increasing dataset size, but also on improving annotation quality. Key elements include explicit representation of cross-reactive ligand sets, paired αβ-chain information where available, and standardized metadata describing experimental assays and readout criteria. Equally important is the development of reliable negative-control schemes. Instead of relying mainly on randomly paired decoys, future benchmarks should distinguish among experimentally validated non-binders, assay-negative peptides, and hard negatives derived from matched HLA backgrounds or biochemically similar peptide families. Clearer negative definitions would improve biological interpretability and confidence in reported performance. A pragmatic way to balance confidence and retention is to define tiered evidence levels rather than requiring all modalities for every entry [34]. For exploratory repertoire annotation, a minimal modality set may include CDR3β sequence, peptide identity, and HLA information when available, while explicitly reporting uncertainty and incomplete coverage. For more reliable cross-reactivity prioritization, paired αβ-chain information, assay metadata, and well-defined negative controls become increasingly important. For therapeutic safety assessment or clinically actionable prioritization, higher-confidence predictions should ideally incorporate paired αβ chains, HLA restriction, functional validation, tissue-expression context, and, where possible, structural or immunopeptidomic evidence [115]. Acceptable missingness should therefore be defined by the intended use case: exploratory screening can tolerate lower coverage and higher uncertainty, whereas clinical prioritization requires stronger evidence support, calibrated uncertainty estimates, and explicit flags for missing modalities. Standardized evaluation represents an equally urgent priority. Many studies still rely on heterogeneous in-house datasets with inconsistent annotation depth and variable negative controls, making cross-study comparison difficult [35, 116]. Moreover, common metrics such as AUC or accuracy focus primarily on binary binding classification and do not capture essential features of cross-reactivity. For one-to-many recognition, evaluation should therefore move beyond binary classification and emphasize ranking-based metrics, such as recall@k or precision@k, together with measures of recognition breadth across candidate peptide sets. For off-target safety assessment, a minimally acceptable benchmark should require models to recover known cross-reactive liabilities from large candidate peptide pools and provide calibrated confidence estimates that flag uncertain predictions for further validation. Community benchmarks should therefore incorporate out-of-distribution evaluation strategies—such as splits by HLA allele, antigen family, source study, or assay type—together with metrics that assess ranking performance, recognition breadth, calibration, and predictive uncertainty. At this stage, clearer benchmarks may be more valuable than increasingly complex models because they define what meaningful progress actually looks like. Intermediate opportunities: biologically grounded multimodal modeling With stronger data and benchmarking foundations in place, the next major opportunity is to develop models that are both biologically grounded and robust under distribution shift. Generalization remains a persistent challenge: many existing models perform poorly on unseen epitopes, rare HLAs, self antigens, or cross-species repertoires, reflecting a mismatch between current training strategies and the biology of cross-reactive recognition [88, 116]. Apparent predictive success may still reflect dataset-specific shortcuts driven by HLA imbalance, epitope redundancy, assay-dependent labels, uneven modality coverage, or biased representation of TCR sequence features. In addition to HLA and epitope imbalance, CDR3 length should also be treated as an explicit biological and computational variable in cross-reactivity prediction [117]. Many sequence-based and ML pipelines may therefore be implicitly biased toward common-length, well-annotated CDR3β sequences, because these dominate public datasets and benchmark collections. Longer or atypical CDR3 loops may increase the complexity of structural modeling, docking, and conformational sampling, and they may be more difficult to annotate using current sequence-similarity resources. At the same time, such CDR3s may be biologically important for recognizing neoepitopes, unusual peptide conformations, or non-canonical pMHC surfaces. If training data and benchmarks are biased toward common-length CDR3β sequences, both ML and structure-based models may underperform on clinically relevant but less represented receptors. Future benchmarks should therefore report CDR3 length distributions, evaluate performance across length strata, and include longer or atypical CDR3s when assessing model generalization. A key intermediate direction is the development of biologically grounded multimodal models. Rather than simply concatenating sequence, structural, repertoire, and functional features, future frameworks should adopt fusion strategies that better reflect the mechanisms of TCR recognition. Important directions include incorporating structural dynamics, modeling docking variability, and using biologically motivated constraints such as contact-map supervision, interface-geometry regularization, energetic priors, or structure-guided re-ranking. Such designs may help multimodal systems move from correlation-heavy pattern recognition toward more mechanistically meaningful representations of cross-reactive binding. Progress in multimodal modeling will also depend on addressing modality imbalance and inconsistent annotation depth across datasets. Sequence data are far more abundant than high-quality structural, immunopeptidomic, or single-cell functional data, making representation alignment difficult and increasing the risk that one modality dominates model behavior [118–120]. Better strategies will therefore be needed to handle missing modalities, unequal evidence quality, and partial supervision, especially across heterogeneous studies. Complementary interpretability analyses—such as perturbation-based tests, in silico alanine scanning, residue-swap experiments, and counterfactual comparisons across matched peptide variants—may further help determine whether models capture causal determinants of recognition rather than superficial correlations. Uncertainty-aware prediction should also become a core design principle, particularly for therapeutic safety screening. Models used for off-target risk assessment should not only rank potential liabilities, but also indicate when predictions are unreliable and require structural or experimental validation. Calibration-aware and abstention-aware frameworks may therefore help translate multimodal predictions into high-confidence decision support. Long-term opportunities: dynamic, context-aware, and clinically actionable prediction Looking ahead, the field will need to move beyond static binding prediction toward integrated models that link molecular recognition with functional and clinical outcomes. Cross-reactivity is shaped not only by sequence similarity or interface geometry, but also by conformational dynamics, antigen processing and presentation, T-cell state, tissue expression patterns, and activation thresholds. Capturing these processes will likely require ensemble-based structural modeling, physics-informed learning, and broader integration of immunopeptidomics, single-cell phenotypes, and tissue-contextual datasets [121–123]. Such frameworks would enable prediction not only of whether a TCR can recognize an alternative ligand, but also whether that interaction is likely to trigger meaningful activation, tissue-specific toxicity, or therapeutically relevant immune responses. Clinically actionable cross-reactivity prediction will likely depend on hybrid pipelines that combine data-driven scoring, structure-guided refinement, calibrated uncertainty estimation, and iterative experimental validation, with computational models serving as prioritization engines for TCR design, off-target screening, and mechanistic discovery. From today’s perspective, some expectations raised in earlier TCR-T–focused reviews have begun to materialize, particularly with the 2024 accelerated FDA approval of afamitresgene autoleucel and early clinical progress of PRAME-targeted autologous TCR-T therapies in solid tumors [124]. Yet these advances mainly validate the therapeutic platform itself, whereas cross-reactivity-specific task definition, benchmark design, negative construction, and AI-assisted off-target assessment remain comparatively underdeveloped. This contrast reinforces the need to treat TCR cross-reactivity prediction as a distinct computational and translational problem. In summary, future progress in TCR cross-reactivity prediction will benefit from a staged research roadmap. Near-term efforts should prioritize dedicated, analysis-ready datasets, reliable negative controls, standardized benchmarks tailored to one-to-many recognition, and pragmatic evidence tiers that balance predictive confidence with data retention [93]. Intermediate advances will depend on biologically grounded multimodal modeling that explicitly addresses structural dynamics, modality imbalance, CDR3 length bias, interpretability, and predictive uncertainty. Over the longer term, the goal is to develop dynamic, context-aware systems that integrate structural, functional, and experimental evidence to support clinically meaningful prediction. Achieving this vision will require sustained collaboration across data generation, computational modeling, and experimental validation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Author Contributions Wenzhen Li conceived the review, performed the literature search, and drafted the manuscript. Yu Pan collected data and prepared the figures and tables. Xu Wu and Yuting Song collected and organized the literature. Yin Wang revised the manuscript. Lu Xie revised the manuscript, supervised the study, and acquired funding. All authors approved the final manuscript. Funding This work was supported by the STCSM “Science and Technology Innovation Action Plan” Computational Biology Program [No. 24JS2840300; 25JS2850400]; Shanghai Municipal Health Commission Collaborative Innovation Cluster Project under Grant [2019CXJQ02]; and the National Natural Science Foundation of China under Grant [No.31870829]. Data availability Data sharing is not applicable to this article as no new data were created or analyzed in this study. Declarations Ethics approval and consent to participate Not applicable. Consent for publication All authors have approved the submission. Competing interests All authors declare no competing interests related to this review. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Wenzhen Li and Yu Pan contributed equally to this work. References Associated Data Data Availability Statement Data sharing is not applicable to this article as no new data were created or analyzed in this study.