Bridging AI, molecular dynamics, and experiments for mechanism-aware TCR engineering in cancer immunotherapy

1 Introduction

T cell receptors (TCRs) play a central role in adaptive immunity by recognizing antigenic peptides presented by major histocompatibility complex (MHC) molecules enabling the immune system to discriminate between self and non-self antigens and to eliminate infected or malignant cells van der Merwe and Davis (2003). This highly specific interaction, mediated by the complementarity-determining regions (CDRs) of the TCR engaging the peptide-MHC (pMHC) complex Garboczi et al. (1996) (see also Figure 1A; Figure 1B), forms the molecular basis of T cell activation and has become a cornerstone for the development of cancer immunotherapies. Specifically, adoptive T cell therapies based on engineered TCRs have shown great promise for targeting tumor-associated and neoantigen-derived epitopes, especially in solid tumors Waldman et al. (2020); Zhao et al. (2021). Despite these advances, the identification and optimization of TCRs with high affinity, specificity, and functional efficacy remain major challenges. Naturally occurring TCRs often exhibit suboptimal affinity for tumor antigens, while affinity enhancement strategies can lead to reduced specificity and increased risk of off-target toxicity. These limitations highlight the need for rational design approaches capable of balancing binding strength and selectivity while preserving functional signaling.

Recent developments in artificial intelligence (AI) have enabled the large-scale exploration of TCR sequence and structural space, providing new opportunities for the generation and optimization of receptor variants Lu et al. (2021); Bujak et al. (2023); Asadi Sarabi et al. (2024). However, these strategies remain largely descriptive or data-driven and lack the ability to provide mechanistic insight into the molecular determinants of TCR-pMHC recognition. Moreover, AI-based approaches often rely on static structural representations and scoring functions that do not fully capture the dynamic nature of protein-protein interactions Ribeiro-Filho et al. (2024). In this context, physics-based methods such as molecular dynamics (MD) simulations and force field-based interaction energy calculations have played an increasingly important role in the study of both TCR–pMHC complexes and isolated TCR and pMHC systems Narzi et al. (2008); Reboul et al. (2012); Nurzia et al. (2012); Zhang et al. (2016); Alba and D’Abramo, 2022; Tomasiak et al. (2022), providing atomistic insight into conformational flexibility, residue-specific energetic contributions, and the dynamic behavior of the protein–protein interface. Previous computational studies have shown that TCR recognition is not a rigid lock-and-key process but involves significant conformational plasticity of both the receptor and the peptide-MHC complex, with fluctuations in the complementarity-determining regions contributing to binding adaptability and specificity Karch et al. (2019); Dominguez and Knapp (2019); Natarajan et al. (2018). Importantly, MD simulations have also been used to investigate the energetic landscape of TCR-pMHC interactions, enabling the identification of key residues that contribute to binding through favorable energetic contributions and cooperative effects Bingöl et al. (2021); Frezzini and Narzi (2026), as illustrated in Figure 1C.

In a recent study, the application of residue-specific energy decomposition to a pair of closely related TCRs (i.e., the engineered variant c796 and its parental counterpart c728) provided a clear example of how subtle sequence changes can reshape the energetic landscape of the interaction Frezzini and Narzi (2026). The two receptors differ by a single conservative mutation in the β-CDR1 loop (E31D), which experimentally results in an approximately six-fold increase in binding affinity toward the same peptide-MHC complex Border et al. (2019); Simister et al. (2022). Notably, detailed energetic analysis revealed that this affinity enhancement cannot be directly attributed to the interaction of the mutated residue with the pMHC complex. Instead, the increase in binding strength arises from a redistribution of interaction energies across multiple residues, as shown in Figure 1C, including positions located in distal CDR loops, which contribute more favorably to the stabilization of the interface in the engineered variant. These findings highlight that binding affinity in TCR-pMHC systems emerges from a cooperative network of interactions rather than from isolated residue contributions. Beyond local energetic effects, increasing evidence points to the existence of long-range communication within the TCR-pMHC complexes Hawse et al. (2012); Rangarajan et al. (2018); Bingöl and Ozbek (2026). Network-based analyses derived from MD trajectories have revealed the presence of allosteric communication pathways connecting the binding interface to distal regions of the receptor, suggesting that local perturbations can propagate through the structure and modulate the behavior of residues located tens of angstroms away, as shown in Figure 1D. In this framework, residues positioned in different CDR loops can become dynamically and energetically coupled, contributing collectively to affinity and specificity. Taken together, the identification of energetic hotspots and allosteric communication pathways provides a powerful strategy for guiding mutagenesis, as modifications at both interfacial and distal positions can modulate binding affinity and functional responses.

Here, we propose an integrated framework for TCR engineering that combines biologically grounded receptor selection, AI-driven sequence design, molecular dynamics simulations, and experimental validation. Within this framework, MD serves not only as a filtering tool but also as a source of mechanistic insight into conformational dynamics, energetic determinants, and allosteric communication, enabling a more rational and efficient approach to TCR engineering for cancer immunotherapy.

2 Selection of TCR candidates and generation of engineered variants

The identification of suitable T cell receptors for therapeutic engineering represents a critical and non-trivial step in the development of TCR-based immunotherapies. Current strategies largely rely on the analysis of TCR repertoires derived from tumor-infiltrating lymphocytes, peripheral blood, or antigen-specific T cell populations. High-throughput sequencing combined with bioinformatic analysis enables the identification of candidate clonotypes based on features such as clonal expansion, persistence, and sequence similarity within the complementarity-determining region 3 (CDR3), as extensively reviewed elsewhere Demaël et al. (2025). These approaches have been successfully used to link TCR sequence patterns with antigen specificity and functional responses Zhao et al. (2025); Zhang et al. (2024). Integrative approaches combining repertoire data with transcriptomic and phenotypic information have enabled a more comprehensive characterization and prioritization of TCRs associated with effective anti-tumor responses Valkiers et al. (2022). High-throughput experimental platforms have also been developed to directly identify antigen-reactive TCRs from large and diverse repertoires. For instance, microfluidics-based approaches enabling the generation and screening of libraries of natively paired TCRα β chains have been successfully used to isolate rare tumor- and virus-specific TCRs from primary human samples Spindler et al. (2020). Together, these methods provide a biologically grounded framework for the selection of initial TCR candidates.

From a translational perspective, however, the identification of clinically relevant TCRs remains challenging. Experimental screening strategies are required to validate antigen specificity and functional activity, while simultaneously minimizing the risk of off-target toxicity. As highlighted in recent reviews, the balance between affinity, specificity, and safety represents a major bottleneck in TCR-based therapies, particularly in the context of solid tumors Lin et al. (2025). Even TCRs with high binding affinity may exhibit cross-reactivity toward unrelated peptides, underscoring the need for robust validation pipelines and careful candidate selection.

Recent developments in artificial intelligence have enabled the large-scale generation and evaluation of protein variants, offering new opportunities for TCR engineering. Deep learning–based protein design methods can efficiently explore vast regions of sequence space and generate sequences compatible with a desired structural framework, substantially expanding the range of candidates that can be evaluated in silico Dauparas et al. (2022). In the framework proposed here, however, AI-based sequence generation is primarily intended to explore the local sequence space surrounding biologically validated TCR scaffolds, rather than to produce highly divergent receptor variants. This design strategy is consistent with the subsequent application of MD simulations, whose mechanistic and energetic comparisons are expected to be most informative when evaluating structurally related variants differing by a limited number of mutations. More recently, deep learning–based protein design approaches have been applied to TCR engineering, where structure-based deep learning models have shown promise in proposing receptor variants with improved binding properties while highlighting current challenges associated with TCR specificity and cross-reactivity Ribeiro-Filho et al. (2024). A representative example is provided by recent AI-driven pipelines that combine generative design, structural prediction, and computational filtering to produce large libraries of de novo protein binders targeting peptide–MHC complexes Johansen et al. (2025). In these approaches, thousands of variants can be generated and screened in silico, often followed by additional refinement steps such as molecular dynamics simulations to reduce false positives and improve the selection of promising candidates. However, such strategies are primarily designed for de novo binder generation and do not directly address the challenges associated with engineering naturally occurring TCRs. In particular, they typically rely on large-scale generation and filtering workflows and do not explicitly incorporate mechanistic information on TCR–pMHC interactions, such as conformational dynamics, energetic contributions, and allosteric communication. As a result, a gap remains between the identification of biologically relevant TCRs and the rational design of optimized variants.

In this perspective, we argue that an effective TCR engineering strategy should combine biologically informed selection of initial receptors with AI-driven generation of sequence variants, followed by mechanistic refinement using physics-based approaches and experimental validation. Such an integrated framework would enable both the exploration of sequence space and the incorporation of molecular-level insight, providing a foundation for a more rational and efficient design of therapeutic TCRs.

3 Molecular dynamics simulations for mechanistic refinement and design

While the combination of repertoire analysis and AI-driven approaches enables the identification and generation of large numbers of candidate TCR variants, the application of physics-based methods such as MD simulations requires a fundamentally different strategy. In practice, reliable MD-based characterization would require all-atom simulations under NPT conditions with physiological ionic strength and unrestrained production trajectories, together with multiple independent replicas (e.g., approximately ten per system) Dominguez and Knapp (2019) and extensive sampling times reaching at least hundreds of nanoseconds per replica to achieve statistical convergence Frezzini and Narzi (2026); Bingöl et al. (2021). Given these computational requirements, this strategy is expected to be applicable to the analysis of a few tens of closely related variants per TCR scaffold, rather than to large-scale screening of hundreds of candidates. For this reason, MD simulations should be positioned as a low-throughput, high-information step within the design pipeline. Rather than being used for large-scale screening, MD is more effectively applied to a limited number of carefully selected candidates, for example, focusing on a single TCR scaffold identified through biological selection and a restricted set of mutations prioritized through preliminary filtering. Such pre-screening can be performed using fast computational approaches based on static structural models, including empirical scoring functions, rapid energetic estimates, or machine learning-based predictors, to reduce the candidate space to a manageable size prior to MD refinement.

A critical prerequisite for MD simulations is the availability of a reliable structural model of the TCR–pMHC complex. Although experimentally determined structures obtained by X-ray crystallography or cryo-electron microscopy provide ideal starting points, such information is often unavailable for newly identified or engineered TCRs. Recent advances in AI-based structural prediction, exemplified by AlphaFold Jumper et al. (2021); Abramson et al. (2024), have enabled the generation of accurate models for previously uncharacterized proteins and complexes Wu et al. (2024). Combined with emerging structure-based deep learning design approaches, these developments are opening new opportunities for TCR engineering Ribeiro-Filho et al. (2024). Still, obtaining accurate structural models of TCR–pMHC complexes remains challenging because of the conformational flexibility of the CDR loops and the intrinsic complexity of protein–protein recognition Bennett et al. (2026). Although AI-based methods have substantially improved structural prediction, their performance for receptor–ligand complexes remains limited in the absence of experimental structural information Hitawala and Gray (2025).

Once a reliable structural model is obtained, MD simulations can provide detailed insight into the molecular determinants of TCR–pMHC recognition. In this context, MD should not be viewed merely as a tool to estimate binding affinity, but rather as a framework for dissecting the interplay between structure, dynamics, and energetics. Analyses of MD trajectories enable the characterization of conformational flexibility, particularly in the complementarity-determining regions, as well as the identification of interaction patterns that stabilize the complex over time Crean et al. (2020), Crean et al. (2022). Importantly, MD-based approaches also allow for the quantitative evaluation of residue-specific energetic contributions, for example, through energy decomposition schemes, which can be used to identify key residues that contribute to binding Bingöl et al. (2021).

A particularly illustrative example is provided by recent work on engineered TCR variants, where MD simulations were combined with residue-specific energy decomposition and network-based analyses to elucidate the molecular origin of affinity enhancement Frezzini and Narzi (2026). In this case, comparison between an engineered TCR (c796) and its parental counterpart (c728), differing by a single conservative mutation in the β-CDR1 loop, revealed that the experimentally observed increase in binding affinity does not arise from the direct contribution of the mutated residue. Instead, the mutation induces a redistribution of interaction energies across multiple residues, including positions located in distal complementarity-determining regions, which contribute more favorably to binding in the engineered variant. In parallel, the analysis of MD-derived interaction networks highlighted the presence of allosteric communication pathways connecting the mutation site to distal regions of the receptor Naganathan (2019). These pathways enable the propagation of local perturbations over long distances, leading to coordinated changes in the interaction network and ultimately affecting binding affinity and specificity Ma et al. (2025); Bingöl and Ozbek (2026). Together, these findings demonstrate that TCR–pMHC recognition is governed by a complex interplay of local and long-range effects, and that even minimal sequence modifications can lead to substantial functional consequences through dynamic and allosteric mechanisms. Taken together, these considerations support the integration of MD simulations into TCR engineering pipelines as a mechanism-aware refinement step rather than a brute-force screening approach. By identifying energetic hotspots, characterizing conformational dynamics, and mapping allosteric communication pathways, MD generates testable mechanistic hypotheses that complement AI-driven design and guide targeted experimental validation.

These opportunities come with an important caveat: the mechanistic insight provided by MD simulations is only as reliable as the structural model on which the simulations are based. Structural inaccuracies may propagate into the predicted conformational dynamics and residue-specific energetics, thereby affecting the interpretation of allosteric communication and mutation prioritization. Consequently, mechanistic hypotheses derived from MD should be interpreted in light of the confidence of the starting model and validated experimentally before guiding subsequent rounds of sequence optimization.

Looking ahead, emerging generative approaches such as AlphaFlow and BioEmu may substantially accelerate conformational sampling and could serve as upstream ensemble generators or pre-screening tools Jing et al. (2024); Lewis et al. (2025). However, their current implementations have not been validated for multichain TCR–pMHC complexes, mutation-induced binding-energy changes, or residue-resolved allosteric-network analyses. Because the mechanistic observables considered here depend on interchain geometry, side-chain packing, explicit solvation, and physically consistent trajectories, we view these approaches as complementary accelerators rather than replacements for all-atom MD. A promising hybrid strategy would use AI-generated ensembles to initialize diverse conformations, followed by explicit-solvent MD for physical refinement, energetic analysis, and validation of mechanistic robustness.

4 Experimental assessment of engineered TCR candidates

The experimental validation of engineered TCRs represents a critical step in the development pipeline, providing the necessary link between computational predictions and functional outcomes. In this context, binding affinity and kinetic parameters are typically assessed using established biophysical techniques, among which surface plasmon resonance (SPR) remains the gold standard for the quantitative characterization of TCR–pMHC interactions Hebeisen et al. (2015). SPR enables the direct measurement of association and dissociation rates, providing detailed insight into binding kinetics and affinity. This level of characterization is particularly valuable for comparing engineered variants and identifying improvements in binding strength Johnson et al. (2006); Zhao et al. (2007). However, despite its accuracy and robustness, SPR provides only a partial description of TCR–pMHC recognition, as binding affinity and kinetics alone do not fully capture the conformational and energetic complexity underlying T cell activation Armstrong et al. (2008). Furthermore, the requirement for recombinant protein expression and purification, combined with relatively low throughput, makes SPR impractical for screening large numbers of candidates. As a result, SPR is more appropriately positioned as a downstream validation tool for a limited set of highly promising variants.

To address the need for higher-throughput functional evaluation, cellular assays have been developed to assess TCR activity in a more physiologically relevant context. Among these, reporter-based platforms such as RAPTR Dobson et al. (2022) have emerged as powerful tools for the functional screening of engineered TCRs. These systems enable the evaluation of TCR activation in response to specific peptide–MHC complexes within a cellular environment, thereby capturing not only binding but also downstream signaling and functional output. Compared to purely biophysical measurements, RAPTR-like approaches offer increased scalability and can be integrated with high-throughput TCR discovery and single-cell profiling workflows, enabling the functional screening of large numbers of variants while simultaneously providing information on T cell activation Gaglione et al. (2026). This is particularly important given that binding affinity alone does not necessarily correlate with functional efficacy, as TCR signaling depends on a complex interplay of factors beyond the stability of the TCR–pMHC complex. At the same time, however, cellular assays typically provide less direct mechanistic information at the molecular level, making them complementary to, rather than a replacement for, detailed structural and computational analyses.

Taken together, these considerations support an iterative TCR engineering strategy in which functional assays are used for efficient candidate prioritization, followed by detailed biophysical characterization and computational refinement of selected variants. The integration of experimental and computational approaches enables successive design cycles that progressively guide the optimization of TCR affinity, specificity, and functional activity.

5 An integrated iterative workflow for TCR engineering

Current TCR engineering strategies are often fragmented, combining repertoire analysis, experimental screening, computational modeling, and AI-driven design without a fully unified framework. To address this limitation, we propose an integrated, iterative workflow that merges these complementary approaches into a coherent pipeline for rational TCR engineering (Figure 2).

The process starts from biologically grounded TCR selection, using repertoire sequencing, clustering, and high-throughput screening to identify receptors with known or putative antigen specificity. This ensures that subsequent design efforts remain anchored to functional, physiologically relevant scaffolds. AI-based approaches are then employed to explore the local sequence space around selected scaffolds, generating libraries of candidate variants. Given the resulting combinatorial complexity, fast structural or machine learning-based filters are used to prioritize a reduced set of promising mutations.

At the core of the workflow lies molecular dynamics simulations, which serve not only as a refinement step but as an integrative mechanistic engine of the entire pipeline. MD is used to interrogate TCR-pMHC recognition at multiple levels, combining residue-wise energy decomposition with the analysis of dynamical allosteric communication pathways. This dual description enables the identification of mutations that are not only locally favorable in terms of interaction energetics, but also capable of reshaping long-range coupling networks that control binding affinity and specificity. In this way, MD provides a mechanistic criterion for prioritizing and rationalizing candidate variants beyond static structural or sequence-based metrics. Importantly, MD is also embedded in an internal iterative loop with AI-based design. Insights from energetic fingerprints and allosteric pathways are fed back to refine the sequence space explored by AI models, enabling the design of improved variants guided by mechanistic understanding rather than purely statistical sampling. In this sense, MD acts as a bidirectional bridge between data-driven design and physical principles, continuously improving the quality of the proposed TCR repertoire.

The refined candidates are then subjected to experimental characterization, combining biophysical techniques such as SPR with functional assays including reporter-based systems such as RAPTR. SPR provides quantitative measurement of binding kinetics and affinity for a limited set of high-confidence variants, enabling detailed biophysical validation of TCR–pMHC interactions. In parallel, RAPTR offers a complementary, higher-throughput assessment of TCR functionality in a cellular context, capturing antigen-dependent activation and downstream signaling outputs. Together, these approaches provide orthogonal layers of validation, linking molecular binding properties to functional T cell responses, while also generating data that can be used to benchmark and improve computational predictions. As illustrated in Figure 2, the most promising TCR candidates emerging from this workflow can be further assessed in vivo and advanced toward preclinical and clinical testing.

This integrated strategy enables efficient exploration of sequence space while preserving biological relevance, incorporates mechanistic insight into design decisions, and unifies computational and experimental approaches within a scalable adaptive framework. Recent studies have demonstrated the feasibility of closed-loop, AI-guided optimization frameworks integrating computational design with iterative experimental validation for therapeutic protein engineering Frey et al. (2025). Our proposed workflow builds upon this emerging paradigm by emphasizing the integration of molecular dynamics simulations as a source of mechanistic insight, enabling residue-resolved energetic analyses and allosteric communication pathways to guide the rational engineering of TCRs. In conclusion, the convergence of AI-guided design, physics-based modeling, and experimental validation provides a powerful foundation for next-generation TCR engineering, with the potential to accelerate the development of high-affinity, specific, and safe receptors for cancer immunotherapy.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. ChatGPT (OpenAI) was used exclusively to improve grammar, language clarity, and readability. All scientific content, interpretations, and conclusions were developed and verified by the authors.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Annarita Fiorillo, Sapienza University of Rome, Italy

Reviewed by: Francis Gaudreault, National Research Council Canada (NRC), Canada