Un activador bispecífico de células T selectivo para p53(R273H): diseño computacional y validación funcional.

The p53 protein is a critical tumor suppressor that functions primarily as a transcription factor. In response to cellular stresses, including hypoxia, DNA damage, and nutrient deprivation, p53 initiates a cascade of events leading to cell-cycle arrest, DNA repair, and restoration of normal cellular homeostasis. When cellular damage exceeds the cell’s repair capacity, p53 instead drives senescence, apoptosis, and reprogramming of cellular energy metabolism and immune responses.^[1],^[2],^[3]
TP53, the gene encoding p53, is mutated in over 50% of human cancers, rendering it one of the most frequently altered genes in oncology.^[4],^[5],^[6],^[7] Hotspot mutations cluster within the DNA-binding domain (e.g., R175H, R248W, R273H, and Y220C) and typically alter protein conformation and impair DNA binding, resulting in loss of tumor-suppressive function together with gain-of-function oncogenic activities that drive proliferation, invasion, and therapeutic resistance.^[4],^[5],^[6],^[7],^[8],^[9]

Since the early 2000s, efforts to target p53-mutant cancers have fallen into three broad categories: (1) small molecules, (2) gene therapies, and (3) immune-based therapies. Small molecules represent the majority of p53-directed programs, largely by disrupting p53 interactions with negative regulators (e.g., MDM2 and MDMX) to restore wild-type activity, or by rescuing mutant p53 conformation directly.^[10],^[11],^[12] Gene-based approaches have since expanded to include delivery of plasmid DNA, RNA, or viral vectors encoding wild-type p53, alongside gene-editing technologies like CRISPR, base editors, prime editors, and meganucleases designed to correct mutations.^[13],^[14] Immune-based strategies, including p53 vaccines and checkpoint inhibitors, constitute a third therapeutic avenues.^[13],^[14]

Despite this breadth of activity, important gaps remain, particularly for the R273H, one of the most common and biologically aggressive TP53 hotspot variants. Together with R175H, R248W, and R282W, R273H accounts for a substantial share of all TP53 mutations and is enriched across high-incidence cancers, including colorectal, breast, and lung adenocarcinomas.^[15] Tumors bearing this mutation display aggressive biology, including stemness, metastasis potential, therapy resistance, and poor prognosis. Yet while agents exist for other hotspots like APR-246 (Eprentapopt) for conformational mutants like R175H, or PC14586 (Rezatapopt) for Y220C mutant, no mutation-specific precision therapy for R273H is approved or in late-stage development. Small molecules that stabilize wild-type p53 or restore the native mutant conformation must also contend with accelerated wild-type p53 degradation in heterozygous R273H-mutant tumors, and frequently lack variant specificity for certain variants.^[11],^[12],^[13],^[14],^[16] Gene therapies likewise require substantial optimization of delivery efficiency and therapeutic efficacy. Early p53 vaccines, while generally safe and immunogenic, have rarely produced objective tumor responses, durable remissions, or improvements in overall survival.^[11],^[12],^[13],^[14],^[16]

Advances in tumor immunology have since spurred immunotherapies designed for greater precision. Cancer cells frequently generate neoantigens in the form of mutant peptides, including those derived from p53, presented by major histocompatibility complex (MHC) class I molecules, enabling T cell recognition and offering a promising strategy for targeted immunotherapy.^[7],^[9] T cell-based approaches broadly fall into two classes: (1) those targeting immunosuppressive pathways, such as immune checkpoint inhibitors (ICIs), and (2) those promoting immunostimulatory pathways, including chimeric antigen receptor (CAR)-T cells, T cell receptor (TCR)-engineered T cells (TCR-T), and bispecific T cell engagers (TCEs).^[13],^[14],^[15] ICIs block inhibitory receptors, such as programmed cell death protein 1 (PD-1) or its ligand (PD-L1) to reinvigorate cytotoxic T cell activity and have achieved clinical success across solid tumors,^[16],^[17] though response rates remain limited by immune evasion and a suppressive tumor microenvironment. CAR-T and TCR-T therapies instead genetically engineer autologous T cells to recognize tumor antigens directly, via receptors that couple an extracellular antigen-recognition domain to intracellular T cell signaling motifs. Their autologous origin limits rejection and autoimmunity risk, but manufacturing is resource-intensive and slow,^[17],^[18] while clinical use remains constrained by limited persistence in solid tumors and antigen escape. Targeting p53 neoantigens specifically is further complicated by their intracellular origin, variable presentation across HLA alleles, and the risk of cross-reactivity with wild-type p53 in normal tissue.^[3],^[6],^[7],^[17],^[19],^[20] These constraints motivate a more precisely targeted, pHLA-restricted platform, particularly the development of the TCE platform described herein.

TCEs, a subset of bispecific antibodies (bsAbs), simultaneously engage tumor-associated surface molecules and T cell-activating receptors, thereby redirecting cytotoxic T cells to malignant cells. Antigen selection determines the specificity of this approach: tumor-associated antigens (TAAs) are self-proteins that are overexpressed or aberrantly expressed in tumor cells but retained at lower levels in normal tissues, conferring an inherent risk of on-target/off-tumor toxicity. In contrast, tumor-specific antigens (TSAs), exemplified by mutation-derived neoantigens such as p53R273H, arise exclusively from somatic alterations, are absent from the normal genome, and are presented by HLA class I solely on malignant cells. This tumor-exclusive presentation makes TSAs particularly attractive targets for TCE design, offering a wider therapeutic window and reduced risk to normal tissues.^[21],^[22],^[23] Bispecific T cell engagers (BiTEs), which bridge CD3 and TAA/TSA through simultaneous binding, are off-the-shelf, exhibit nanomolar potency, and have shown clinical efficacy in hematologic malignancies (e.g., blinatumomab).^[24] By directly linking T cells to tumor cells through TSA recognition, TCEs amplify T cell-mediated tumor lysis, positioning them as a compelling tool in anticancer immunotherapy. Their application to solid tumors, particularly TCEs targeting neoantigen-derived peptide-MHC complexes, remains underexplored, however, owing to persistent challenges in affinity, specificity, and stability.^[19],^[20],^[25]

Addressing these limitations and the unmet clinical need for R273H-specific therapies in particular requires new protein engineering strategies. We employed a stapled single-chain format, in which an engineered interchain disulfide bond, introduced via site-specific cysteine substitutions at the Vβ-Vα interface, covalently locks two variable chains together in place of relying on a flexible peptide linker alone. This disulfide “staple” prevents domain mispairing and dissociation, reduces molecular size, improves folding fidelity, and enhances pharmacokinetics without compromising antigen-binding function.^[26] Machine learning (ML)-driven protein design complements this by enabling rapid affinity maturation and specificity optimization through prediction and iterative structural modeling.^[27],^[28],^[29] A comparable strategy has been pursued for p53 R175H-targeting TCE, CLSP-1025, demonstrating the feasibility of this approach for hotspot mutations.^[30] Integrating these tools with patient-derived TCRs offer a path to next-generation TCEs development against recurrent neoantigens, such as p53 R273H.^[15]

Here, we leveraged a naturally occurring TCR (TCR01) isolated from a colorectal cancer patient harboring p53R273H and HLA-C∗01:02 to engineer and optimize a TCE platform against this target. We identified HLA-C∗01:02 as the primary presenting allele for the immunodominant 9-mer epitope (SFEVHVCAC) and engineered stapled TCE variants with reduced molecular weight and enhanced potency, further refined through ML-guided directed evolution for affinity and specificity. Notably, HLA-C∗01:02 is a prevalent allele occurring in ∼20% of East Asian populations (approximately 10%–15% in Vietnamese and other Southeast Asian populations) and ∼12% of European populations, supporting the broad clinical applicability. Comprehensive in vitro and 3D spheroid assays demonstrated robust, allele-restricted cytotoxicity against endogenous p53R273H-expressing tumors, with minimal off-target effects (Figure 1A). By delivering a high-affinity, mutation-specific TCE against this prevalent and clinically challenging neoantigen, the present study offers a promising therapeutic strategy for a large subset of p53-mutant cancers currently without effective targeted options, and highlights the synergy of structural and computational optimization in neoantigen-directed TCE design.

Results

Characterization and evaluation of p53-targeting TCR in context of T cell engaging receptor

In our previous study, in an attempt to establish a procedure for identifying a shared neoantigen-reactive T cell receptor, samples from patients diagnosed with various cancers were collected and processed.^[31] Among them was one sample with colorectal cancer harboring an endogenous hotspot mutation in TP53 gene, resulting in substitution of arginine at codon 273 with histidine (R273H). Peripheral blood mononuclear cells (PBMCs) from this patient were exposed to 25-mer peptides corresponding to the p53 R273 region carrying wild-type or R273H residues (SGNLLGRNSFEVRVCACPGRDRRTE or SGNLLGRNSFEVHVCACPGRDRRTE). IFN-γ assessment via ELISPOT was used to evaluate T cells reactivity.^[31] T cell populations exhibiting more than 2-fold increase when exposed to mutant peptide were subjected for single cell isolation and next-generation sequencing (NGS). Sequence analysis of this sample identified a novel T cell receptor, designated TCR01 (corresponding to TCR12.1 in patent application 12442046).^[31] Because clone selection required functional threshold (>2-fold increase in IFN-γ secretion against the mutant peptide), this approach enriches for T cell clones with higher relative avidity for their cognate pHLA compared with unselected or tetramer-sorted repertoires, consistent with the established phenomenon of avidity maturation during antigen-driven T cell clonal selection. TCR01 was used in its naturally isolated form, without subsequent affinity engineering. In addition, HLA class I typing of the patient revealed the following alleles: HLA-A∗02:06, HLA-A∗11:01, HLA-B∗40:01, HLA-B∗54:01, HLA-C∗01:02, and HLA-C∗07:02 (Figures 1B; Table S1). These data collectively suggest that p53R273H can serve as a potential neoepitope presented by one or more of the patient’s HLA-class-I alleles.^[31] In this study, the sequence of identified TCR01 was used to engineer p53R273H-targeting TCE. Although preliminary studies have proposed candidate TCRs capable of recognizing the p53R273H neoepitope, by the time of this study, only limited data on pHLA/p53R273H tetramer staining or on TCR01 itself has been reported, particularly in the context of its incorporation into TCE format. To further evaluate the therapeutic potential of this approach, we first characterized the p53 epitope presentation capacity of the identified patient-derived HLA alleles as well as the efficacy of the TCR01-based TCE, which we refer as TCE01 class.

Previously, we have established a T cell engaging receptor platform by incorporating TCR 1-2C, which targeted KRASG12V presented by HLA-A∗11:01.^[22] Using the similar approach, we generated TCE01A, in which a single-chain TCR01 (sTCR01) is fused to N-terminus of single-chain variable fragment anti-CD3 (sCD3) and followed by a single-chain Fc tail in knob-into-hole [KIH] format. The designated TCE01A label is used to distinguish from other TCE formats we will use later in this study (Figure S1A). In addition, we engineered SW620 colorectal cancer cell line to stably express relevant HLAs identified in patient’s sample. There were two different versions of the same HLA engineered into CRC SW620, including the original sequence, and DYKDDDDK (FLAG)-6xHis-DYKDDDDK tagged version. The epitope tags were inserted at the C-terminus of the alpha 3 (α3) domain in the extracellular region, immediately preceding the N-terminus of the transmembrane (TM) domain. This design ensures proper membrane localization of engineered HLAs (Figures S1B and S1C).^[32] For lung cancer cells NCI-H520, pancreatic cancer cells PANC-1 and breast cancer cells MDA-MB-468, the non-tagged HLA was engineered (Figure S1B).

To evaluate the potential presentation of p53R273H-derived neoepitope by the patient’s HLA class I alleles, engineered SW620 was pulsed with a long p53R273H 25-mer peptide (20 μM). T cell dependent cytotoxicity was assessed by co-culturing with T cells and TCE01A across a concentration range of 10−4 to 101 nM. At 24 h, superior T cell-mediated cytotoxicity was observed for the cells presenting HLA-C∗01:02/p53R273H complex, achieving up to approximately 74.86% lysis, compared with moderate lysis for HLA-B∗54:01, and minimal to negligible lysis for HLA-A∗02:06, HLA-A∗11:01, HLA-B∗40:01, HLA-C∗07:02, and the mock control (Figure 1C). The high cytotoxicity observed in HLA-C∗01:02 was independent from the presence or absence of epitope tags used in our constructs (Figure S1D). In addition, these observations also highlight the allele-specific potency of p53R273H-targeting TCE in neoepitope recognition.

Although the 25-mer peptide was designed to encompass the core epitopes, its intracellular processing and loading onto HLA-C∗01:02 is less efficient and more variable. In contrast, the small, processed 8–11 peptide can bind directly to this HLA and result in robust T cell-mediated cytotoxicity. Thus, it is necessary to identify the specific p53-derived peptide(s) presented by HLA-C∗01:02. For initial screening, we used NetHMCScan 4.1 to predict and compare the binding affinity of p53R273H-derived peptides (8–14 mers) to HLA-C∗01:02 as well as other patient-derived HLA alleles (Tables S1). We selected the top 5 peptide candidates with the strongest predicted presentation potential for further evaluation (Figures 1D; Tables S2 and S3). In our experiments, HLA-C∗01:02-expressing SW620 cells pulsed with these candidate peptides exhibited varying degrees of T cell-mediated lysis efficiency. Notably, the 9-mer peptides H5 and H6 induced significant cytotoxicity, whereas the other three candidates show no difference from mock control (Figure 1E). Interestingly, HLA-C∗01:02-expressing cells pulsed with 9-mer peptide H5 exhibited more robust T cell-dependent cytotoxicity at dose of 0.25 and 0.0625 nM than the cells pulsed with the 25-mer peptide under the same conditions (Figure 1E).### Engineering TCR01 into different TCE formats

In the present study, we further expanded the previous TCE platform in terms of relative spatial positions between single-chain TCR (sTCR) and single-chain anti-CD3 (sCD3), while also aimed to balance between molecular size and stability. TCE variants are designated as TCE01XrN, where X denotes the molecular architecture (A–D; see Figure 1F for domain schematics and molecular weights) and rN indicates the optimization iteration on TCE01 scaffold (r1 = lead candidate from first round of engineering, r2 = lead candidate from second round, r3 = lead candidate from third round). We previously developed and validated two TCE formats targeting the KRASG12V neoepitope presented by HLA-A∗11:01, designated as A and B.^[22] Format A was described briefly in the preceding section, whereas format B consists of a KIH sFc is fused N-terminally to sTCR and C-terminally to sCD3. In both formats, the constant regions of the TCR α- and β-chains were included to promote proper chain pairing and maintain structural similarity to the native TCR.

In this study, we sought to reduce the molecular size of our TCE molecules by minimizing domain content while preserving the antibody binding affinity and functionality (Figure 1F). Despite similarities between antibodies and TCRs, which all composed of variable and constant regions, TCR’s variable domains are inherently unstable and prone to misfolding or aggregation without their constant domains, especially for the case of TCE where the transmembrane regions are also removed.^[30],^[33],^[34] In a recent study, Boucher et al. demonstrated that retaining only the variable domains of an antibody is feasible by introducing a cysteine residue into each variable fragment and incorporating a hinge-mimicking linker of the Fc region.^[26] This approach resulted in the so-called stapled single-chain variable fragment (stapled scFv or spFv). A similar approach has been done during the development of ABBV-184, a Survivin-targeting bispecific TCE engineered by Chervin et al., however, such a format was only mentioned in their patent but not in the main publication.^[35] Overall, by the time of this study, there are limited applications to miniTCE molecules have been reported to date.

In this study, we adopted an approach similar to that described by Boucher et al. to generate the stapled-single-chain TCE carrying TCR01. The methionine residue at position 43 on β chain was substituted with cysteine (M43C), and alanine residue 100 was substituted with cysteine (A100C) on α chain. These residues are located outside of the complementary-determining region (CDR) on both chains and are positioned to form disulfide bonds with the hinge-like linker that connects two chains. The linker consisted of two copies of G4S flexible linker, followed by a CPPC hinge-like sequence, and then one additional copy of G4S linker (Figure S1E). The predicted distance between the Cα atoms of M42C and A100C was 4.0 Å, and the predicted distance between two cysteine residues in the linker was 6.8 Å (Figures S1E and S1F). These values fall within the typical range observed for Cα-Cα distances in the hinge region of standard IgG antibody approximately 5.0–7.0 Å. Implementation of the stapled approach generated two additional platforms used in this study (Figure 1F). This strategy reduced the theoretical molecular weight of TCE01 from 130.38 kDa (in TCE01A and TCE01B) to approximately 104.80 kDa (in TCE01C and TCE01D), corresponding to a reduction of 24.40% (Figures 1F and S1G).

Bio-layer interferometry (BLI) was used to evaluate the binding affinity of TCE01 variants for the pHLA-C∗01:02/p53R273H 9-mer (H5) complex. The pHLA-C∗01:02/p53R273H 9-mer H5 was immobilized on the biosensor and binding of TCE01 variants was performed in a 3-fold serial dilution. All TCE01 variants (A–D) showed comparable binding affinities to the immobilized pHLA-C∗01:02/p53R273H 9-mer H5 complex, with dissociation constant (Kd) values of 17.43 nM (TCE01A), 18.58 nM (TCE01B), 20.22 nM (TCE01C), and 25.21 nM (TCE01D). These results indicate that the stapled format effectively preserves the parental binding activity (Figure 1G). The Kd values obtained for TCE01 variants (17–25 nM) are higher than the affinity range typically reported for naturally occurring TCR-pHLA interactions (1–100 μM). This is consistent with TCR01’s isolation via functional avidity-based screening rather than conventional or tetramer-sorted repertoire sampling (see above). We also note that because the pHLA ligand was immobilized as a streptavidin-based tetramer, the reported Kd values represent apparent affinities that may be influenced by local avidity/rebinding effects at the sensor surface rather than strictly reflecting intrinsic monomeric 1:1 affinity.

Next, we assessed the on-cell binding affinities of the TCE constructs to CD3 and p53R273H HLA-C∗01:02 target cells using flow cytometry. All TCE01A, TCE01B, TCE01C, and TCE01D exhibited high binding affinity against CD3ε on T cells with Kd values of 2.544 nM, 2.322 nM, 2.467 nM, and 2.763 nM, respectively. All of these values are higher (lower affinity) than the parental anti-CD3 antibody at 136.3 pM (Figure 1H).^[22] Consistent with BLI results, the binding of the TCE01 variants to SW620 cells expressing HLA-C∗01:02 pulsed with peptide p53R273H 9-mer H5 showed no significant difference among TCE01A-D with the Kd of 121.42 nM (TCE01A), 134.11 nM (TCE01B), 101.24 nM (TCE01C), and 112.52 nM (TCE01D) (Figures 1I and 1J).

It is worth noting the performance differences between the stapled versions (TCE01C and TCE01D) and non-stapled versions (designated with ∗ sign, e.g., TCE01C∗ and TCE01D∗). Under similar conditions, we observed a noticeable reduction in binding affinity with Kd values of 60.42 nM for TCE01C∗ (versus 20.22 nM for TCE01C) and 78.73 nM for TCE01D∗ (versus of 25.21 nM in TCE01D) (Figure S1H). In addition, the non-stapled TCE∗ variants (TCE01C∗ and TCE01D∗) again showed relatively weaker on-cell binding to SW620 cells expressing HLA-C∗01:02 and pulsed with p53R273H peptide, compared with their stapled versions TCE01C and TCE01D, consistent with BLI approach (Figure S1I). For simplicity, we hereafter only focus on the stapled versions with TCE01C and TCE01D. These data together demonstrate that TCE01A-D retain the bispecific binding to both pHLA-C∗01:02/p53R273H 9-mer H5 target via TCR moiety as well as CD3ε via anti-CD3 moiety, and importantly the stapled approach can be utilized as a potential option to further reduce the size of TCR region to be incorporated into TCE formats.### p53R273H-targeting TCE01 promotes T cell-mediated cytotoxicity in vitro

To investigate the capacity of TCE01 variants to activate T cells and mediate target cell killing, TCE molecules were co-cultured with p53R273H 9-mer H5 peptide-pulsed SW620 CRC cells expressing HLA-C∗01:02 and primary human T cells in a dose-dependent manner. Target cell viability was quantified at 24 h post co-culture using CellTiter-Glo luminescence ATP assay. Despite comparable binding affinities across the variants, TCE01A demonstrated reduced cytotoxic efficacy, with a maximum specific lysis of 72.24%. TCE01B achieved 78.62% maximum cytotoxicity, whereas TCE01C and TCE01D induced higher cytotoxicity, with maxima of 92.52% and 95.26%, respectively (Figure 2A). All four TCE01 variants achieve near-maximal cytotoxicity at concentration ≥0.01 nM, with optimal specific lysis observed at 0.625 nM and higher. T cell activation induced by TCE01 variants was further evaluated by measuring CD69 expression via flow cytometry. TCE01A and TCE01B yield maximum CD69+ frequencies of 62.14% and 70.22%, respectively, which were substantially lower than those observed for TCE01C (87.24%) and TCE01D (86.82%) under equivalent conditions. This pattern aligned with the observed trend T cell-mediated cytotoxicity (Figure 2B).

We additionally profiled the release of cytolytic granules from TCE-activated T cells using flow cytometry to quantify perforin and granzyme B. Granzyme B expression increased in TCE01-treated samples, with frequencies of 44.15% for TCE01A, 62.51% for TCE01B, 78.42% for TCE01C, and 71.22% for TCE01D, consistent with CD69 activation profile and T cell-mediated cytotoxicity outcome (Figure 2C). Similarly, perforin expression followed a similar trend, albeit lower frequencies (40.05% for TCE01A, 48.51% for TCE01B, 62.21% for TCE01C, and 59.25% for TCE01D) (Figure 2D). In line with these markers for activation and cytolytic function, interferon gamma (IFN-γ) secretion exhibited a comparable dose-dependent pattern across the TCE01 variants (Figure 2E). Collectively, the ability of TCE01 variants to bind, recruit, and activate T enables them to effectively promote cytotoxicity against CRC cells expressing pHLA-C∗01:02/p53R273H (Figure 2F).### Exploring sensitivity of p53R273H-targeting TCE variants in 2D and 3D models

To investigate the sensitivity of TCE01 variants, primary human T cells were co-incubated with target cells at various E:T ratios (0:1, 0.1:1, 0.5:1, 1:1, 2:1, 5:1, and 10:1), while maintaining constant TCE concentrations. The data showed that TCE01A and TCE01B achieved optimal cytotoxicity starting at an E:T ratio of 5:1, reaching 68.52% and 76.24%, respectively. In contrast, TCE01C and TCE01D elicited significant cytotoxicity at an E:T ratio of 2:1, with killing efficiencies exceeding 70% and further increasing to approximately 92.56% and 94.52%, respectively (Figure S2A). In time course analyses, TCE01C and TCE01D attained near-complete T cell-mediated cell killing by 24 h post-co-culture, whereas TCE01A and TCE01B required up to 48 h to surpass 90% cytotoxicity (Figure S2B). Collectively, these findings indicate that the miniTCE (C and D platforms) not only reduces molecular weight of T cell-engaging receptor molecules but also substantially enhances their functional efficacy. Accordingly, subsequent analyses prioritized the TCE01C variant due to its superior profile in terms of molecular weight and potency.

We next examined the efficacy of TCE01C using primary T cells derived from multiple healthy donors, in recognition of the substantial variability in immune profiles. Patients exhibit differential responses to identical immunotherapies, highlighting the importance of assessing TCE01C’s ability to recruit and activate T cells from multiple donors.^[36] Across seven healthy donors, TCE01C EC50 for T cell-mediated cytotoxicity ranged from 0.62 pM (donor #004, 95% CI 0.58–0.68 pM) to 2,862 pM (donor #006; 95% CI, 1,863–6,183 pM), approximately 4,600-fold range (one-way ANOVA on log-transformed EC50, F6’14 = 490.3, p < 0.0001; Welch’s t test for donor #004 vs. #006, t = 25.5, df = 2.2, p = 0.001), underscoring substantial inter-donor variability in T cell responsiveness (Figure 2G).

To further investigate the potency of TCE01C in physiologically relevant models, we evaluated its activity in 3D spheroid systems. Given the increasing adoption of 3D models for preclinical evaluation of drug efficacy in cancer and other diseases,^[37],^[38],^[39] we generated spheroids composed of HLA-C∗01:02-expressing CRC cells pulsed with p53 R273H peptide, human adipose-derived stem cells (ASCs), and human umbilical vein endothelial cells (HUVECs). Of note, only the colorectal cancer cells within the spheroids were engineered to express HLA-C∗01:02, while the entire spheroid was pulsed with p53 R273H 9-mer peptide (H5). This experimental design partially recapitulates the heterogeneity observed in patient-derived tumor samples, in which HLA expression is variable across cell subsets and not all cells within the tumor microenvironment present the target antigen. These spheroids were co-cultured with primary T cells and TCE01C in a dose-dependent manner (Figure 2H). Therapeutic effects were quantified via fluorescence-based assays and IFN-γ secretion. TCE01C eliminated 82.45% of CRC cells within 48 h, accompanied by a maximal IFN-γ release of 9.2 ng/mL, while exhibiting minimal to no cytotoxicity toward non-malignant ASCs and HUVECs (Figures 2I–2K). Collectively, these findings highlight the robust efficacy of TCE01C across diverse patient-derived T cells and in complex 3D tumor microenvironments.### Specificity of TCE01C

Investigating the specificity of bispecific T cell engagers is critical for their successful therapeutic development. For the p53R273H-targeting TCE01C, specificity was evaluated across multiple dimensions: (1) peptide specificity, assessed via analysis of key interaction residues and homologous peptides, (2) HLA restriction, determined by comparing recognition of p53R273H presented by different HLAs alleles, and (3) TCE sensitivity to peptides presented specifically by HLA-C∗01:02. Given that the p53R273H 9-mer H5 epitope is presented by HLA-C∗01:02, structurally similar peptides may also be presented by this HLA molecule, potentially inducing conformational changes that facilitate TCE binding and subsequent T cell-dependent cytotoxicity. To explore this possibility, we first designed and synthesized a panel of 9-mer peptides harboring substitutions at the p53 codon 273 position, analogous to the p53R273H 9-mer H5 used herein. These included variants with documented or predicted associations to the onset and progression of diverse cancers (Tables S2 and S4), encompassing R273H, R273C, R273L, R273G, R273P, and R273S. Among these p53R273 variants, TCE01C elicited T cell-mediated killing against four variants: the primary target R273H as well as R273C, R273L, and unexpectedly wild-type R273 (R273WT). On-cell binding assays corroborated these findings, showing clear population shifts indicative of TCE01C engagement with the same four targets (Figures 3A and 3B). In addition, we titrated p53 R273H, R273WT, and R273C peptides from 0.002 to 20,000 nM. All three variants achieved near-maximal efficacy at 2,000 nM, with R273H exhibiting superior potency compared with others (Figure S2C). These results demonstrated that TCE01C recognizes and functionally engages additional p53 R273 variants beyond the intended p53R273H target.^[1],^[40] Notably, because wild-type p53 protein undergoes continuous proteasomal degradation and MHC presentation in normal cells, the observed recognition of the p53 R273WT peptide by TCE01C is particularly concerning, as it demonstrates the high sensitivity of the molecule toward the wild-type sequence. In addition, we subsequently examined TCE01C activity against homologous peptides to the p53 R273H 9-mer H5 (SFEVHVCAC) (Table S4). Homology searches were conducted using NCBI Blast and ExPAsy’s ScanProsite, permitting up to seven mismatches as the threshold.^[15],^[22],^[35],^[41],^[42] Notably, no off-target peptides with a single mismatch outside R273 position (fifth residue in the 9-mer) were identified. Among these off-targets, TCE01C exhibited detectable activity solely against off-target 01 (SFEIHVCSC) (Figure 3C).

The observed activity of TCE01C toward R273 variants, but limited response to potential off-targets, suggests possible off-target effects potentially attributed to (1) intrinsic TCE binding properties, (2) critical peptide interaction residues, or (3) HLA-C∗01:02 presenting capacity. To probe this, we performed alanine scanning by generating single alanine-substituted variants at each position (P1–P9) of the p53R273H 9-mer H5 peptide and evaluated TCE01C functional activity in CellTiter-Glo assays (Table S4). Of the eight alanine-substituted peptides, only V6A variant retained full activity comparable to the original p53R273H 9-mer H5. Substitutions S1A, H5A, and C7A reduced TCE01C efficacy by approximately 50%, whereas E3A and C9A substitutions caused substantial reduction. Notably, alanine replacements at positions F2A or V4A completely abolished TCE01C functional activity (Figure 3D). As TCE01C also recognized p53 R273H 9-mer H6 with moderate efficacy (Figure 1E), we extended alanine scanning to each position of this peptide (Table S4). Complete abrogation was observed at S2A, E4A, and C8A substitutions, while N1A, V5A induced partial reduction in efficacy, mirroring patterns in the p53 R273H 9-mer H5 alanine variants. Conversely, F3A, H6A, and V7A substitutions did not impair TCE01C function (Figure 3E). Collectively, these data indicate that the second, fourth, and eighth residues on p53 R273H 9-mer peptide are pivotal for HLA-C∗01:02 presentation and TCE01C binding efficacy.

To elucidate the structural basis of TCE01C interactions, we employed state-of-the-art computational modeling approaches, specifically Boltz-2 and AlphaFold 3, to predict ternary complex comprising TCE01C, HLA-C∗01:02 with p53R273H 9-mer H5, p53R273H 9-mer H6, or p53R273H 10-mer H9.^[43],^[44] These models were generated to identify residues critical for complex. In agreement with our in vitro findings, the second and eighth residues were predicted to be essential for the stable presentation of 9-mer H5, 9-mer H6 peptides, whereas the second residue of 10-mer H9 exhibited instability within the ternary complex. Additionally, the fourth residue in 9-mer H5 was also predicted to contribute to complex stabilization, such that alanine substitution at this position markedly impaired both peptide presentation and TCE01C binding efficacy (Figures S2D–S2F). Across these models, a congruent pattern emerged regarding the key governing complex integrity.### Computational optimization of p53R273H-targeting TCE

The observed cross-reactivity of TCE01C toward R273 variants, coupled with limited responses to potential off-targets, suggests its potential utility in targeting additional other p53 R273 variants across relevant cancer types, particularly given that p53 R273 represents a hotspot mutation in human malignancies. A TCE capable of recognizing multiple variants could function as an off-the-shelf therapy, applicable to a broader patient cohort spanning diverse tumor indications. Supporting evidence demonstrates that TCE molecules engaging multiple tumor-associated antigens or variants confer superior tumor recognition and a broader antitumor efficacy relative to those with stringent single-target specificity.^[45] However, such broadened specificity heightens the risk of cross-reactivity with wild-type p53 or other unrelated self-peptides, which may emerge stochastically in heterogeneous patient population contingent on individual conditions.^[46],^[47] Although mutant p53 R273H is generally absent in normal cells, inadvertent recognition of wild-type p53 could potentially trigger severe systemic autoimmunity or on-target/off-tumor toxicity.^[5],^[48],^[49] Furthermore, TCRs that can target several variants may exhibit diminished affinity for individual targets, thereby constraining T cell activation and tumoricidal potency. To address this potential trade-off between breadth and potency, we pursued further optimization of our TCE platform, prioritizing refinements to the TCR domain in context of TCE architecture.

We first predicted the structure of the pHLA-C∗01:02/p53R273H 9-mer H5 complex, by employing three independent computational modeling programs: AlphaFold 3, Boltz-2, and TCRmodel2.^[43],^[44],^[50] Even though AlphaFold3 and Boltz-2 are more generalist models, Boltz-2 is open, fast, and controllable while Alphafold3 provides the highest general accuracy. Several models were generated in parallel using these two models and ranked by ipTM score. On the other hand, TCRmodel2 was trained based on TCR docking database and used in this approach for cross validation. Top pHLA structure will then be used to predict the complex structure with addition of TCE01C using above-mentioned models. Leveraging this structure, we performed ML-guided directed evolution using three complementary computational protein design tools EvoPro, EvoProtGrad, and Antibody Evol^[27],^[51],^[52],^[53] (Figure 4A). Emergent candidates were evaluated for (1) predicted binding affinity (ddG, Prodigy, and BoltzGen), (2) structural stability (DeepSP and Aggrescan 3D), (3) immunogenicity (ABImmpred and Hu-mAb), and (4) toxicity (TAP).^[44],^[54],^[55],^[56],^[57] Among these criteria, the binding affinity was used to rank top candidates, while the later ones, including stability and immunogenicity, were included to predict safety and efficacy of generated candidates. The highest-ranking candidate from each iteration served as the parental sequence for subsequent rounds of directed evolution, with optimized residues fixed thereafter. This iterative workflow was complete after three cycles of computational directed evolution (Figures 4A and S3A; Table S5). Notably, evolutionary modifications were not confined to CDRs, as accumulating evidence underscores the contributions of non-CDR domains to conformation dynamics, binding affinity, interface stability, and overall functional efficacy in antibodies and related engagers.^[58],^[59],^[60],^[61] Modeling of the TCE01C/pHLA-C∗01:02/p53R273H 9-mer H5 ternary complex reveal asymmetric engagement restricted to one flank of the HLA-peptide binding groove, alongside disproportionate α/β chain positioning, with α chain oriented (Figures 4B and S3B). Prior investigations have established the pivotal role of CDR3β in TCR functionality. ML-optimized candidates exhibited progressive conformational refinements, facilitating comprehensive interaction across the peptide-binding groove and equitable alignment of both TCR chains toward the epitope (Figures S3C–S3E).### Optimized engagers displayed enhanced on-cell binding efficacy toward pHLA-C∗01:02/p53R273H 9-mer H5

One of the current challenges in developing novel proteins through computational programming is confirming whether the predicted performance of new variants in silico translates faithfully to experimental reality in both 2D and 3D models. To mitigate this risk and bridge the in silico-in vitro divide, top candidates from each evolutionary round of directed evolution underwent empirical validation in cell-based assays prior to advancement. Candidates were evaluated for binding efficiency and T cells-dependent cytotoxicity. BLI revealed progressive improvements in TCR affinity for stabilized pHLA-C∗01:02/p53R273H 9-mer H5 complex, with Kd values declining from 20.22 nM (TCE01C) to 12.51 nM (TCE01Cr1), 3.12 nM (TCE01Cr2), and for the final lead variant 2.51 nM (TCE01Cr3) (Figure 4C). This increase in in vitro binding affinity aligns closely with the corresponding decrease in predicted Kd values obtained from in silico evaluations across the optimization rounds (Table S5). On-cell binding assays further confirmed these improvements, demonstrating progressively enhanced affinities for the target pHLA-C∗01:02/p53R273H complex across successive variants. The final lead candidate, TCE01Cr3, exhibited the strongest on-cell binding signal, consistent with its superior biophysical profile (Figures 4C–4E).

To compare the functional potencies of optimized p53R273H-targeting TCE variants, we quantified T cell-mediated cytotoxicity and cytokine release in co-cultures with HLA-C∗01:02-expressing SW620 pulsed with p53R273H 9-mer H5 peptide. Dose-response curves for specific lysis revealed sigmoidal profiles for all TCE constructs, with maximal lysis approaching 100% at higher concentrations. The lead optimized variant, TCE01Cr3 (orange) exhibited the highest potency (EC50 ≈ 10–5 nM), followed by TCE01Cr2 (blue; EC50 ≈ 2 × 10−5 nM), TCE01Cr1 (green; EC^[44] ≈ 6 × 10−4 nM), and the parental TCE01C (red; EC50 ≈ 5 × 10−3 nM). Control conditions (TCR-Fc, CD3-Fc, and cells only) showed negligible activity (Figure 5A). Similarly, IFN-γ secretion followed analogous trends, with TCE01Cr3 achieving maximal release (∼7.85 ng/mL; EC50 ≈ 5 × 10−4 nM), outperforming TCE01Cr2 (EC50 ≈ 10–3 nM), TCE01Cr1 (EC50 ≈ 5 × 10−3 nM), and TCE01C (EC50 ≈ 10–2 nM), while controls remained baseline (Figure 5B). We next assessed the dependence of TCE-mediated killing on effector-to-target ratios (0:1, 0.1:1, 0.5:1, 1:1, 2:1, 5:1, and 10:1). All four variants TCE01C, TCE01Cr1, TCE01Cr2, and TCE01Cr3 induced strong target cell lysis even at low E:T ratios (as low as 1:1), with maximal killing above 85% reached at E:T ratios of 5:1 and 10:1. At an E:T ratio of 1:1, TCE01Cr2 and TCE01Cr3 achieved significantly higher cytotoxicity than TCE01C and TCE01Cr1 (one-way ANOVA, F3’8 = 56.5, p < 0.001), with pairwise differences ranging from 1.8-fold (TCE01Cr2 vs. TCE01C, Tukey-adjusted p = 0.0007) to 2.8-fold (TCE01Cr3 vs. TCE01Cr1, Tukey-adjusted p < 0.001). At an E:T ratio of 2:1, the same rank order was observed, but differences were smaller and only partially significant (F3’8 = 7.5, p = 0.010 overall). Only TCE01Cr3 vs. TCE01Cr1 remained significant after correction (1.4-fold, p = 0.007), while TCE01Cr2 vs. TCE01C did not (1.06-fold, p = 0.85) (Figure 5C). Integrating computational predictions, biophysical binding data, and preliminary cellular data, TCE01Cr3 emerged as the lead candidate for advanced evaluation. Thereafter, evaluations focused primarily on TCE01Cr3, incorporating ancillary variants where pertinent.### TCE01Cr3 demonstrated enhanced fidelity on HLA-C∗01:02/p53R273H 9-mer H5 tetramer

Similar to the parental TCE01C, comprehensive specificity profiling of ML-optimized TCE01Cr3 was imperative. The co-culture assays demonstrated robust T cell-mediated lysis of target cells presenting the p53R273H neoantigen with maximal specific lysis exceeding 80%. Residual cytotoxicity was still detectable against cells pulsed with the R273C and wild-type R273 peptides, although the magnitude of lysis was substantially reduced compared with the parental TCE01C (Figure 3A). In contrast, no measurable response was elicited against cells pulsed with R273L, R273G, R273P, R273S, or mock (no peptide) controls (Figure 5D). TCE01Cr3 exhibited strong binding to SW620 cells pulsed with p53 R273H 9-mer peptide, moderate binding to R273C and wild-type R273 peptides, and no detectable binding to R273L, R273G, R273P, R273S, or mock controls (Figure 5E). It is worth mentioning that in this approach, peptide carrying these p53 R273 variants were pulsed at the dose of 20 μM, which is among the highest doses of peptide being pulsed across similar studies^[23],^[39],^[62] and well beyond the normal presenting capacity of native neoantigen. For p53 R273 9-mer carrying either R273H, R273WT, or R273C, those that triggered TCE-mediated cytotoxicity of TCE01Cr3, sensitivity was evaluated in the peptide dose-dependent manner ranging from 20,000 to 0.002 nM. Specific lysis reached a maximum of ∼88%–90% at higher R273H peptide concentrations starting from 2 to 20,000 nM, with an EC50 in the nanomolar range. In contrast, both R273WT and R273C mutant peptide only trigger TCE-mediated killing at the dose of 20,000 nM peptide and dramatically reduce at lower dose (Figure 5F). Pulsing with the wild-type R273WT peptide resulted in negligible lysis across the entire concentration range. The closely related R273C mutant peptide induced only minimal lysis at the highest concentrations tested, confirming strong discrimination against both the wild-type sequence and this single-amino-acid variant.

Off-target evaluations yielded robust lysis (>80% maxima) exclusively for the R273H target, with negligible activity against on homologs or mock (Figure 5G). HLA allele restriction assays confirmed potent and specific lysis exclusively for HLA-C∗01:02 pulsed with p53R273H 9-mer H5, with negligible responses to HLA-A∗02:06, HLA-A∗11:01, HLA-B∗40:01, HLA-B∗54:01, HLA-C∗07:02, or mock (Figure 5H). These results confirm the exquisite allele and peptide specificity of TCE01Cr3 for the HLA-C∗01:02–p53R273H complex.

Similarly, Alanine scanning of the core p53R273H 9-mer H5 epitope (SFEVHVCAC) pinpointed residues F2, E3, V4, H5, V6, and C7 as critical, with substitutions eliciting marked reductions or abrogation of lysis, while S1A and C9A still exhibited weak cytotoxicity (Figure 5I). In silico ternary complex mapping of TCE, B2M/HLA-C∗01:02, and p53R273H 9-mer H5 highlighted predominant CDR involvement, with CDRβ1, CDRβ2, and CDRα3 favoring HLA-C∗01:02 contacts, and CDR3α/β involved in both HLA and peptide interaction, similar to previous studies on other TCR/HLA/peptide interactions.^[22],^[35],^[63] Computational optimization notably amplified CDR3 engagement with presenting peptide residues, concurrent with augmented HLA-C∗01:02 interaction (Figure S3B; Table S6). On the other hand, while Alanine scanning on presenting peptide allows us to quickly identify critical hotspot residues where side-chain interactions can contribute significantly to binding affinity or specificity, we further investigated the full specificity profile particularly for the key residues essential for anchoring p53 R273H 9-mer peptide onto HLA-C∗01:02 or residue involved in interacting with TCR, namely, positions 2, 5, 6, and 8 (P2, P5, P6, and P8). Structural studies on other HLA/TCR complexes have suggested a pattern where P2 and P8 are essential form proper presentation of peptide onto HLA while P5 and P6 are more involved in interaction with TCR.^[62],^[64],^[65],^[66] This approach is necessary especially for the case of p53 R273H 9-mer H5 peptide in this study as the position P8 is the Alanine residue (SFEVHVCAC), which cannot be further examined from Alanine scanning perspective. A combinatorial peptide library was generated where each of the residues at position 2, 5, 6, 8, corresponding to F2, H5, V6, A8, respectively, was substituted with 20 amino acids (Figure 6A). The responses of TCE01C and TCE01Cr3 were tested with target cells cytotoxicity via CellTiter-Glo 2D. For position #2, we observed a high tolerance of TCE01C for hydrophobic amino acids, especially for the ones with aromatic side chain similar to the original residue Phenylalanine (F) (Figure 6B). The optimized version, TCE01Cr3, exhibited a substantial reduction in tolerance; however, still has some detectable off-target activity if this position is switched to the hydrophobic category, even though at a lower level (Figure 6B). For position #5, TCE01C showed some levels of tolerance toward hydrophobic amino acids, especially for the negatively charged residues besides some hydrophobic ones like Alanine (A) or Leucine (L). Optimized TCE01Cr3 was in general more sensitive to substitution at this position with low-to-moderate activity for hydrophilic residues like C5, N5, K5 (Figure 6C). Position #6 showed a tremendous off-target issue for TCE01C if substituted to hydrophobic category, especially for the aliphatic group. This issue was completely solved in TCE01Cr3 when it was only sensitive to Valine at the 6th position. Cross reactivity at the theoretical C-terminal anchoring A8 was observed against a substitution with other amino acids also belong to the hydrophobic family beside hydrophilic T8 and Q8 for TCE01C. On the other hand, TCE01Cr3 shows more stringent fidelity at this position with only Alanine showed the robust cytotoxicity (Figure 6E).

We then further tried to use ScanProsite to analyze homologous peptides that meet the responsive motifs of TCE01Cr3 observed from the peptide combinatorial library analysis above. This analysis results in four additional homologous peptides (Uniprot: A0R2I7, SAEVHVVTC; Uniprot: A3QFJ5, SSEVKVCAD; Uniprot: E1BBQ2, SEEVHVCLP; and Uniprot: P34388, IPEVHVCAV). These four additional homologous peptides carry 3–4 mismatches comparing with p53 R273H 9-mer H5 peptide we used throughout this study. However, none of them were from human genome (Table S4). In addition, sensitivity against peptides presented by HLA-C∗01:02 were also investigated. The interaction with HLA-C∗01:02 will anchor TCR regions of TCE01Cr3 in the positions where they are exposed to peptides presented by this HLA, and thus, can trigger potential off-target activity. As a result, evaluating peptides presented by this HLA can provide us another insight toward TCE01Cr3 specificity. Using Immune Epitope Database & Tools (IEDB), we found 2,924 peptides reported to be presented by HLA-C∗01:02. Based on the off-target pattern when performing 20 amino acid-scanning for TCE01Cr3 on p53 R273H 9-mer, we picked the top 10 candidates for quick evaluations (Table S4, off-target 17–26). These candidates all carry Phenylalanine at the second position (P2), which matches the p53 R273H 9-mer peptide and can serve as an important anchor point for peptide presentation. Furthermore, they also carry matches at the fifth (Off-target 18), sixth (Off-target 17–24), or eighth (Off-target 17, 25, 26) positions, making them more vulnerable to TCE01Cr3 recognition. A T cell-dependent cytotoxicity assay was carried out for these candidates in comparing with p53 R273H 9-mer H5 peptide. In our hands, these 10 candidates did not trigger significant off-target activities from TCE01Cr3, with an exception of Off-target 17, which also carry F2, V6, and A8 residues, similar to p53 R273H 9-mer peptide (Figure 6F).### Immune response of TCE01Cr3 across diverse cell types

We further examined TCE01Cr3 efficacy against endogenous p53R273H expression using the SW620 colorectal cancer line, which harbors the heterozygous p53R273H mutation.^[54] Initial cytotoxicity assays using SW620 cells engineered to express HLA-C∗01:02 and pulsed with p53 R273H peptide revealed a baseline-specific lysis of 15%–20% killing even at low doses of optimized variants (TCE01Cr1, TCE01Cr2, and TCE01Cr3), whereas the parental TCE01C showed no detectable activity under the same conditions (Figure 5A), prompting scrutiny of hypersensitivity or cross-reactivity. While cross-reactivity for TCE01Cr3 was partially investigated via off-target screening, Alanine scanning, and combinatorial peptide screening at essential residues (Figures 5D–5I and 6A–6E), it remains unknown whether TCE01Cr3 is hypersensitive, especially to endogenous p53 R273H as SW620 cell line used throughout this study is heterozygous for TP53R273H. To determine whether heterozygosity contributed to this sensitivity, CRISPR/Cas9 was employed to edit p53 exon 8 flanking R273 with dual gRNAs and a repair template encoding 818819GT > AC modification (Figure S4A). Cells transfected with plasmid expressing Cas9 and gRNAs were isolated for further evaluation (Figure S4B). Among these cells, editing was observed in approximately 12.27% population (Figure S4C). Clonal derivatives were vetted by variant-specific PCR and p21 expression (indicative of wild-type p53 ablation) (Figures S4D–S4F). Resultant lines included homozygous wild-type p53 and compound heterozygous p53R273H (native 818G > A plus CRISPR 818819GT > AC). Modifications were verified by Sanger sequencing, followed by NGS confirming on-target modification and negligible off-target edits (Figures S4G and S4H). Notably, homozygous p53R273H clones (818G > A or 818_819GT > AC) proved inviable for sustained culture (data not shown), while the compound heterozygous cell line with two different mutated alleles encoding the same p53 R273H was obtained. As a result, in related experiments, we used this cell line in parallel with homozygous p53R273WT- and heterozygous p53R273H-cell lines. In these endogenous models, TCE01Cr3 achieved 47.81% cytotoxicity at 3.9 × 10−2 nM in compound heterozygous p53R273H-expressing cells versus 26.44% in heterozygotes. Comparatively, TCE01C yielded 26.82% and 23.56%, respectively, at 4.0 nM. Negligible killing was observed in SW620 expressing HLA-C∗01:02 and harboring homozygous wt-p53 as well as p53R273H SW620 expressing HLA-A∗11:01 (Figure 6G).

In addition, we further evaluated the potency of TCE01Cr3 across a panel of cell lines, including cancer cell lines SW620 (colorectal carcinoma), NCI-H520 (lung cancer), MDA-MB-468 (breast carcinoma), and PANC-1 (pancreatic cancer), as well as non-cancerous lines HEK293 (human embryonic kidney) and CCL-75 WI-38 (human lung fibroblasts). SW620 is heterozygous for TP53 R273H, MDA-MB-468 and PANC-1 are homozygous for TP53 R273H while NCI-H520 is homozygous wild-type at this locus, similar to non-malignant cells. Target cells with endogenous p53 R273H peptide were co-cultured with TCE01Cr3 and T cells at the E:T ratio of 10:1 and assessed for viability at 0, 24, 48, and 72 h. We observed a significant T cell-mediated cytotoxicity in all three cancer cell lines, albeit with varying efficiency. The strongest killing was observed in MDA-MB-468 and PANC-1 cells expressing HLA-C∗01:02 at 48–72 h, followed by SW620 under similar condition. In contrast, neither significant cytotoxicity was detected in cancer cell lines missing HLA-C∗01:02 nor the non-cancerous HEK293 or WI-38 cell lines (Figure 6H).### Stability evaluation for TCE targeting HLA-C∗01:02/p53 R273H

We further probed TCE01Cr3 stability. TCE01Cr3 employs a truncated (mini) single-chain architecture, in which the TCR constant domains are omitted, potentially influencing not only functionality but also stability. As the stapled-single-chain approach remains emergent in the field, this feature was carefully considered in our designs.^[26] Independent studies have suggested that the Fc domain can extend the half-life of bispecific T cell engager molecules.^[22],^[35],^[63] To determine whether this stability profile was specific to the TCE01C-derived lineage or generalizable across formats, each of the four architectures used in this study was intravenously administered to Swiss mice, and stability was evaluated by measuring residual serum concentrations over 72 h by ELISA, with half-life derived by noncompartmental analysis (Phoenix WinNonlin 8.5.2.4, Model 201) from the terminal log-linear elimination phase of each animal’s concentration-time profile. Specifically, TCE01Ar3 (platform A) and TCE01Br3 (platform B) retain both the variable and constant TCR regions, whereas TCE01Cr3 (platform C) and TCE01Dr.3 (platform D) employ the truncated, stapled single-chain architecture. Serum concentrations of all four formats were compared over the same time course.

Our results showed that TCE01Cr3 had a half-life of 39.40 ± 10.50 h (Figure 7A; Table 1). To determine whether the stapling modification itself affects systemic stability, we compared TCE01Cr3 with other three formats and observed comparable half-lives across all four (TCE01Ar3: 43.72 ± 10.71 h, N = 4; TCE01Br3: 48.12 ± 24.62 h, N = 6; TCE01Dr.3: 30.85 ± 16.36 h, N = 6) (Figures 7A; Table S1), with no evidence that the stapled formats (TCE01Cr3 and TCE01Dr.3) cleared faster than their non-stapled counterparts (TCE01Ar3 and TCE01Br3). These data indicate that within this matched set of optimized formats, the stapling strategy did not compromise systemic stability.

To further investigate the serum stability of these formats in vitro, each molecule was pre-incubated in human serum at 37°C for 0–128 h. Following pre-incubation, the samples were co-cultured with primary T cells and pHLA-C∗01:02-expressing target cells pulsed with p53 R273H peptide for an additional 24 h, and stability was evaluated indirectly by measuring IFN-γ release as a functional readout. Prolonged pre-incubation led to lower IFN-γ release, particularly in samples pre-incubated for over 16 h.^[22] To accurately assess how changes in serum TCE stability affect downstream functionality, IFN-γ levels were normalized to values obtained from the corresponding 0-h pre-incubation samples (defined as 100% activity) with mock-treated controls set to 0%. TCE01Cr3 activity decreased by ∼15% by the first hour, followed by a further reduction to 20%–40% loss by 2–4 h and ∼30% loss by 8 h. With extended incubation, the rate of decline slowed considerably, stabilizing at 128 h with 20.43% residual activity. The other three formats showed a comparable overall decline pattern over the same time course (Figure 7B). These results indicate that the stapling strategy had only a minimal impact on the long-term serum stability of TCE01Cr3, consistent with the comparable four-format profile observed in Figure 7A.### ML-derived TCE01Cr3 requires structural refinement on both CDR and non-CDR regions

Even though in-silico ternary complex mapping of TCE01Cr3, B2M/HLA-C∗01:02, and p53R273H 9-mer H5 highlighted predominant CDR involvement, it overlooked contributions to conformational folding, as evidenced by distinct spatial orientations between TCE01C and TCE01Cr3 in relation with B2M/HLA-C∗01:02/p53R273H 9-mer H5 (Figures S3B–S3E; Table S6). In addition, for the optimized candidates, we also noticed some mutations outside of the CDR regions and wondered how much these replacements contribute to the activity of the optimized versions (Figure 4B). To determine whether non-CDR modifications are essential for the binding and functionality of p53-targeting TCEs, we generated hybrid variants by swapping CDR and non-CDR regions between TCE01C and TCE01Cr3. Specifically, the non-CDR framework of TCE01Cr3 was integrated into TCE01C to yield TCE01C-S, while the non-CDR framework of TCE01C was integrated into TCE01Cr3 to yield TCE01Cr3-S (Figure S5A). In silico 3D structure prediction revealed greater similarity between TCE01C-S and TCE01Cr3, and between TCE01C and TCE01Cr3-S (Figures S5B–S5E). Moreover, dose-response analyses of cell killing and IFN-γ release assays demonstrated reduced efficacy for both TCE01C-S and TCE01Cr3-S relative to TCE01Cr3 efficacy, although both hybrids outperformed TCE01C (Figures S5F and S5G). Notably, removal of the TCR non-CDR regions from TCE01Cr3 markedly diminished killing activity, whereas grafting these non-CDR regions onto TCE01C-S substantially enhanced its potency. On the other hand, specificity was evaluated via co-culture with SW620 expressing HLA-C∗01:02 and pulsed with p53R273C 9-mer C5, with TCE01C-S and TCE01Cr3-S show intermediate level of off-target activity (Figure S5H). These findings together indicate that optimizations confined to either CDRs or non-CDRs are insufficient to fully enhance the activity and specificity of ML-derived TCEs, necessitating coordinated modifications across both regions.

Notably, given TCE01C sensitivity to HLA-C∗01:02/p53R273C 9-mer C5, we applied analogous to ML-directed procedures and criteria to develop TCE04C with improve sensitivity and specificity on this target (Figure S5I). TCE04C exhibited enhanced cytotoxicity and specificity toward HLA-C∗01:02/p53R273C (Figures S5J and S5K), warranting further investigation in subsequent studies.### Optimized TCE variants elicit potent T cell-mediated cytotoxicity in 3D multicellular tumor spheroids

To evaluate the antitumor efficacy of optimized TCE variants in complex 3D tumor microenvironments, multicellular spheroids were assembled comprising p53R273H 9-mer H5-pulsed CRC expressing HLA-C∗01:02 (green fluorescence), human umbilical vein endothelial cells (HUVECs; red), and adipose-derived stem cells (ASCs; blue). Spheroids were then treated with TCE01Cr3 in the presence or absence of primary T cells, either alone or in combination with chemotherapeutic agents (oxaliplatin or irinotecan) and monitored via brightfield and fluorescence microscopy at 0, 24, and 48 h. Viability of cancer cells expressing HLA-C∗01:02 and pulsed with p53R273H peptide, non-malignant HUVECs and ASCs were assessed via changes in green, red, and blue fluorescence signal over incubation time. TCE01Cr3 induced progressive disassembly and loss of green fluorescence in T cell-containing conditions, indicative of targeted CRC lysis. Co-treatment with oxaliplatin or irinotecan did not augment this effect, and controls lacking TCE or T cells preserved spheroid integrity (Figure 7C). TCE01Cr3 induced dose-dependent, T cell-mediated spheroid lysis, with superior efficacy when used alone compared to combinations with oxaliplatin or irinotecan, while no-TCE and mock conditions remained baseline (Figure 7D). IFN-γ secretion mirrored these trends, with TCE01Cr3 and TCE01Cr3/oxaliplatin yielding maxima ∼8.42 ng/mL, but different EC50 values (0.051 and 0.49 nM, respectively), followed by TCE01Cr3/irinotecan (∼5.24 ng/mL, EC50 = 3.3 nM) (Figure 7E).

Spheroids harboring CRC cells with endogenous p53R273H mutations (homozygous or heterozygous) were similarly assessed for up to 72 h. TCE01Cr3 induced rapid spheroid disruption and loss of live-cell fluorescence in p53R273H compound heterozygous (R273H/R273H) CRC spheroids, with intermediate effects in heterozygous (R273H/R273WT) spheroids; parental TCE01C showed only weak activity (Figure 7F). Dose-response curves revealed higher potency against homozygous targets (79.85% max lysis, EC50 ≈ 2.3 nM) than heterozygous targets (54.32% max lysis, EC50 ≈ 8.2 nM), underscoring robust efficacy of ML-derived variant (Figure 7G). These findings highlight strong chemotherapy-independent efficacy of the optimized variant in a 3D endogenous tumor model.

Characterization and evaluation of p53-targeting TCR in context of T cell engaging receptor

In our previous study, in an attempt to establish a procedure for identifying a shared neoantigen-reactive T cell receptor, samples from patients diagnosed with various cancers were collected and processed.^[31] Among them was one sample with colorectal cancer harboring an endogenous hotspot mutation in TP53 gene, resulting in substitution of arginine at codon 273 with histidine (R273H). Peripheral blood mononuclear cells (PBMCs) from this patient were exposed to 25-mer peptides corresponding to the p53 R273 region carrying wild-type or R273H residues (SGNLLGRNSFEVRVCACPGRDRRTE or SGNLLGRNSFEVHVCACPGRDRRTE). IFN-γ assessment via ELISPOT was used to evaluate T cells reactivity.^[31] T cell populations exhibiting more than 2-fold increase when exposed to mutant peptide were subjected for single cell isolation and next-generation sequencing (NGS). Sequence analysis of this sample identified a novel T cell receptor, designated TCR01 (corresponding to TCR12.1 in patent application 12442046).^[31] Because clone selection required functional threshold (>2-fold increase in IFN-γ secretion against the mutant peptide), this approach enriches for T cell clones with higher relative avidity for their cognate pHLA compared with unselected or tetramer-sorted repertoires, consistent with the established phenomenon of avidity maturation during antigen-driven T cell clonal selection. TCR01 was used in its naturally isolated form, without subsequent affinity engineering. In addition, HLA class I typing of the patient revealed the following alleles: HLA-A∗02:06, HLA-A∗11:01, HLA-B∗40:01, HLA-B∗54:01, HLA-C∗01:02, and HLA-C∗07:02 (Figures 1B; Table S1). These data collectively suggest that p53R273H can serve as a potential neoepitope presented by one or more of the patient’s HLA-class-I alleles.^[31] In this study, the sequence of identified TCR01 was used to engineer p53R273H-targeting TCE. Although preliminary studies have proposed candidate TCRs capable of recognizing the p53R273H neoepitope, by the time of this study, only limited data on pHLA/p53R273H tetramer staining or on TCR01 itself has been reported, particularly in the context of its incorporation into TCE format. To further evaluate the therapeutic potential of this approach, we first characterized the p53 epitope presentation capacity of the identified patient-derived HLA alleles as well as the efficacy of the TCR01-based TCE, which we refer as TCE01 class.

Previously, we have established a T cell engaging receptor platform by incorporating TCR 1-2C, which targeted KRASG12V presented by HLA-A∗11:01.^[22] Using the similar approach, we generated TCE01A, in which a single-chain TCR01 (sTCR01) is fused to N-terminus of single-chain variable fragment anti-CD3 (sCD3) and followed by a single-chain Fc tail in knob-into-hole [KIH] format. The designated TCE01A label is used to distinguish from other TCE formats we will use later in this study (Figure S1A). In addition, we engineered SW620 colorectal cancer cell line to stably express relevant HLAs identified in patient’s sample. There were two different versions of the same HLA engineered into CRC SW620, including the original sequence, and DYKDDDDK (FLAG)-6xHis-DYKDDDDK tagged version. The epitope tags were inserted at the C-terminus of the alpha 3 (α3) domain in the extracellular region, immediately preceding the N-terminus of the transmembrane (TM) domain. This design ensures proper membrane localization of engineered HLAs (Figures S1B and S1C).^[32] For lung cancer cells NCI-H520, pancreatic cancer cells PANC-1 and breast cancer cells MDA-MB-468, the non-tagged HLA was engineered (Figure S1B).

To evaluate the potential presentation of p53R273H-derived neoepitope by the patient’s HLA class I alleles, engineered SW620 was pulsed with a long p53R273H 25-mer peptide (20 μM). T cell dependent cytotoxicity was assessed by co-culturing with T cells and TCE01A across a concentration range of 10−4 to 101 nM. At 24 h, superior T cell-mediated cytotoxicity was observed for the cells presenting HLA-C∗01:02/p53R273H complex, achieving up to approximately 74.86% lysis, compared with moderate lysis for HLA-B∗54:01, and minimal to negligible lysis for HLA-A∗02:06, HLA-A∗11:01, HLA-B∗40:01, HLA-C∗07:02, and the mock control (Figure 1C). The high cytotoxicity observed in HLA-C∗01:02 was independent from the presence or absence of epitope tags used in our constructs (Figure S1D). In addition, these observations also highlight the allele-specific potency of p53R273H-targeting TCE in neoepitope recognition.

Although the 25-mer peptide was designed to encompass the core epitopes, its intracellular processing and loading onto HLA-C∗01:02 is less efficient and more variable. In contrast, the small, processed 8–11 peptide can bind directly to this HLA and result in robust T cell-mediated cytotoxicity. Thus, it is necessary to identify the specific p53-derived peptide(s) presented by HLA-C∗01:02. For initial screening, we used NetHMCScan 4.1 to predict and compare the binding affinity of p53R273H-derived peptides (8–14 mers) to HLA-C∗01:02 as well as other patient-derived HLA alleles (Tables S1). We selected the top 5 peptide candidates with the strongest predicted presentation potential for further evaluation (Figures 1D; Tables S2 and S3). In our experiments, HLA-C∗01:02-expressing SW620 cells pulsed with these candidate peptides exhibited varying degrees of T cell-mediated lysis efficiency. Notably, the 9-mer peptides H5 and H6 induced significant cytotoxicity, whereas the other three candidates show no difference from mock control (Figure 1E). Interestingly, HLA-C∗01:02-expressing cells pulsed with 9-mer peptide H5 exhibited more robust T cell-dependent cytotoxicity at dose of 0.25 and 0.0625 nM than the cells pulsed with the 25-mer peptide under the same conditions (Figure 1E).

Engineering TCR01 into different TCE formats

In the present study, we further expanded the previous TCE platform in terms of relative spatial positions between single-chain TCR (sTCR) and single-chain anti-CD3 (sCD3), while also aimed to balance between molecular size and stability. TCE variants are designated as TCE01XrN, where X denotes the molecular architecture (A–D; see Figure 1F for domain schematics and molecular weights) and rN indicates the optimization iteration on TCE01 scaffold (r1 = lead candidate from first round of engineering, r2 = lead candidate from second round, r3 = lead candidate from third round). We previously developed and validated two TCE formats targeting the KRASG12V neoepitope presented by HLA-A∗11:01, designated as A and B.^[22] Format A was described briefly in the preceding section, whereas format B consists of a KIH sFc is fused N-terminally to sTCR and C-terminally to sCD3. In both formats, the constant regions of the TCR α- and β-chains were included to promote proper chain pairing and maintain structural similarity to the native TCR.

In this study, we sought to reduce the molecular size of our TCE molecules by minimizing domain content while preserving the antibody binding affinity and functionality (Figure 1F). Despite similarities between antibodies and TCRs, which all composed of variable and constant regions, TCR’s variable domains are inherently unstable and prone to misfolding or aggregation without their constant domains, especially for the case of TCE where the transmembrane regions are also removed.^[30],^[33],^[34] In a recent study, Boucher et al. demonstrated that retaining only the variable domains of an antibody is feasible by introducing a cysteine residue into each variable fragment and incorporating a hinge-mimicking linker of the Fc region.^[26] This approach resulted in the so-called stapled single-chain variable fragment (stapled scFv or spFv). A similar approach has been done during the development of ABBV-184, a Survivin-targeting bispecific TCE engineered by Chervin et al., however, such a format was only mentioned in their patent but not in the main publication.^[35] Overall, by the time of this study, there are limited applications to miniTCE molecules have been reported to date.

In this study, we adopted an approach similar to that described by Boucher et al. to generate the stapled-single-chain TCE carrying TCR01. The methionine residue at position 43 on β chain was substituted with cysteine (M43C), and alanine residue 100 was substituted with cysteine (A100C) on α chain. These residues are located outside of the complementary-determining region (CDR) on both chains and are positioned to form disulfide bonds with the hinge-like linker that connects two chains. The linker consisted of two copies of G4S flexible linker, followed by a CPPC hinge-like sequence, and then one additional copy of G4S linker (Figure S1E). The predicted distance between the Cα atoms of M42C and A100C was 4.0 Å, and the predicted distance between two cysteine residues in the linker was 6.8 Å (Figures S1E and S1F). These values fall within the typical range observed for Cα-Cα distances in the hinge region of standard IgG antibody approximately 5.0–7.0 Å. Implementation of the stapled approach generated two additional platforms used in this study (Figure 1F). This strategy reduced the theoretical molecular weight of TCE01 from 130.38 kDa (in TCE01A and TCE01B) to approximately 104.80 kDa (in TCE01C and TCE01D), corresponding to a reduction of 24.40% (Figures 1F and S1G).

Bio-layer interferometry (BLI) was used to evaluate the binding affinity of TCE01 variants for the pHLA-C∗01:02/p53R273H 9-mer (H5) complex. The pHLA-C∗01:02/p53R273H 9-mer H5 was immobilized on the biosensor and binding of TCE01 variants was performed in a 3-fold serial dilution. All TCE01 variants (A–D) showed comparable binding affinities to the immobilized pHLA-C∗01:02/p53R273H 9-mer H5 complex, with dissociation constant (Kd) values of 17.43 nM (TCE01A), 18.58 nM (TCE01B), 20.22 nM (TCE01C), and 25.21 nM (TCE01D). These results indicate that the stapled format effectively preserves the parental binding activity (Figure 1G). The Kd values obtained for TCE01 variants (17–25 nM) are higher than the affinity range typically reported for naturally occurring TCR-pHLA interactions (1–100 μM). This is consistent with TCR01’s isolation via functional avidity-based screening rather than conventional or tetramer-sorted repertoire sampling (see above). We also note that because the pHLA ligand was immobilized as a streptavidin-based tetramer, the reported Kd values represent apparent affinities that may be influenced by local avidity/rebinding effects at the sensor surface rather than strictly reflecting intrinsic monomeric 1:1 affinity.

Next, we assessed the on-cell binding affinities of the TCE constructs to CD3 and p53R273H HLA-C∗01:02 target cells using flow cytometry. All TCE01A, TCE01B, TCE01C, and TCE01D exhibited high binding affinity against CD3ε on T cells with Kd values of 2.544 nM, 2.322 nM, 2.467 nM, and 2.763 nM, respectively. All of these values are higher (lower affinity) than the parental anti-CD3 antibody at 136.3 pM (Figure 1H).^[22] Consistent with BLI results, the binding of the TCE01 variants to SW620 cells expressing HLA-C∗01:02 pulsed with peptide p53R273H 9-mer H5 showed no significant difference among TCE01A-D with the Kd of 121.42 nM (TCE01A), 134.11 nM (TCE01B), 101.24 nM (TCE01C), and 112.52 nM (TCE01D) (Figures 1I and 1J).

It is worth noting the performance differences between the stapled versions (TCE01C and TCE01D) and non-stapled versions (designated with ∗ sign, e.g., TCE01C∗ and TCE01D∗). Under similar conditions, we observed a noticeable reduction in binding affinity with Kd values of 60.42 nM for TCE01C∗ (versus 20.22 nM for TCE01C) and 78.73 nM for TCE01D∗ (versus of 25.21 nM in TCE01D) (Figure S1H). In addition, the non-stapled TCE∗ variants (TCE01C∗ and TCE01D∗) again showed relatively weaker on-cell binding to SW620 cells expressing HLA-C∗01:02 and pulsed with p53R273H peptide, compared with their stapled versions TCE01C and TCE01D, consistent with BLI approach (Figure S1I). For simplicity, we hereafter only focus on the stapled versions with TCE01C and TCE01D. These data together demonstrate that TCE01A-D retain the bispecific binding to both pHLA-C∗01:02/p53R273H 9-mer H5 target via TCR moiety as well as CD3ε via anti-CD3 moiety, and importantly the stapled approach can be utilized as a potential option to further reduce the size of TCR region to be incorporated into TCE formats.

p53R273H-targeting TCE01 promotes T cell-mediated cytotoxicity in vitro

To investigate the capacity of TCE01 variants to activate T cells and mediate target cell killing, TCE molecules were co-cultured with p53R273H 9-mer H5 peptide-pulsed SW620 CRC cells expressing HLA-C∗01:02 and primary human T cells in a dose-dependent manner. Target cell viability was quantified at 24 h post co-culture using CellTiter-Glo luminescence ATP assay. Despite comparable binding affinities across the variants, TCE01A demonstrated reduced cytotoxic efficacy, with a maximum specific lysis of 72.24%. TCE01B achieved 78.62% maximum cytotoxicity, whereas TCE01C and TCE01D induced higher cytotoxicity, with maxima of 92.52% and 95.26%, respectively (Figure 2A). All four TCE01 variants achieve near-maximal cytotoxicity at concentration ≥0.01 nM, with optimal specific lysis observed at 0.625 nM and higher. T cell activation induced by TCE01 variants was further evaluated by measuring CD69 expression via flow cytometry. TCE01A and TCE01B yield maximum CD69+ frequencies of 62.14% and 70.22%, respectively, which were substantially lower than those observed for TCE01C (87.24%) and TCE01D (86.82%) under equivalent conditions. This pattern aligned with the observed trend T cell-mediated cytotoxicity (Figure 2B).

We additionally profiled the release of cytolytic granules from TCE-activated T cells using flow cytometry to quantify perforin and granzyme B. Granzyme B expression increased in TCE01-treated samples, with frequencies of 44.15% for TCE01A, 62.51% for TCE01B, 78.42% for TCE01C, and 71.22% for TCE01D, consistent with CD69 activation profile and T cell-mediated cytotoxicity outcome (Figure 2C). Similarly, perforin expression followed a similar trend, albeit lower frequencies (40.05% for TCE01A, 48.51% for TCE01B, 62.21% for TCE01C, and 59.25% for TCE01D) (Figure 2D). In line with these markers for activation and cytolytic function, interferon gamma (IFN-γ) secretion exhibited a comparable dose-dependent pattern across the TCE01 variants (Figure 2E). Collectively, the ability of TCE01 variants to bind, recruit, and activate T enables them to effectively promote cytotoxicity against CRC cells expressing pHLA-C∗01:02/p53R273H (Figure 2F).

Exploring sensitivity of p53R273H-targeting TCE variants in 2D and 3D models

To investigate the sensitivity of TCE01 variants, primary human T cells were co-incubated with target cells at various E:T ratios (0:1, 0.1:1, 0.5:1, 1:1, 2:1, 5:1, and 10:1), while maintaining constant TCE concentrations. The data showed that TCE01A and TCE01B achieved optimal cytotoxicity starting at an E:T ratio of 5:1, reaching 68.52% and 76.24%, respectively. In contrast, TCE01C and TCE01D elicited significant cytotoxicity at an E:T ratio of 2:1, with killing efficiencies exceeding 70% and further increasing to approximately 92.56% and 94.52%, respectively (Figure S2A). In time course analyses, TCE01C and TCE01D attained near-complete T cell-mediated cell killing by 24 h post-co-culture, whereas TCE01A and TCE01B required up to 48 h to surpass 90% cytotoxicity (Figure S2B). Collectively, these findings indicate that the miniTCE (C and D platforms) not only reduces molecular weight of T cell-engaging receptor molecules but also substantially enhances their functional efficacy. Accordingly, subsequent analyses prioritized the TCE01C variant due to its superior profile in terms of molecular weight and potency.

We next examined the efficacy of TCE01C using primary T cells derived from multiple healthy donors, in recognition of the substantial variability in immune profiles. Patients exhibit differential responses to identical immunotherapies, highlighting the importance of assessing TCE01C’s ability to recruit and activate T cells from multiple donors.^[36] Across seven healthy donors, TCE01C EC50 for T cell-mediated cytotoxicity ranged from 0.62 pM (donor #004, 95% CI 0.58–0.68 pM) to 2,862 pM (donor #006; 95% CI, 1,863–6,183 pM), approximately 4,600-fold range (one-way ANOVA on log-transformed EC50, F6’14 = 490.3, p < 0.0001; Welch’s t test for donor #004 vs. #006, t = 25.5, df = 2.2, p = 0.001), underscoring substantial inter-donor variability in T cell responsiveness (Figure 2G).

To further investigate the potency of TCE01C in physiologically relevant models, we evaluated its activity in 3D spheroid systems. Given the increasing adoption of 3D models for preclinical evaluation of drug efficacy in cancer and other diseases,^[37],^[38],^[39] we generated spheroids composed of HLA-C∗01:02-expressing CRC cells pulsed with p53 R273H peptide, human adipose-derived stem cells (ASCs), and human umbilical vein endothelial cells (HUVECs). Of note, only the colorectal cancer cells within the spheroids were engineered to express HLA-C∗01:02, while the entire spheroid was pulsed with p53 R273H 9-mer peptide (H5). This experimental design partially recapitulates the heterogeneity observed in patient-derived tumor samples, in which HLA expression is variable across cell subsets and not all cells within the tumor microenvironment present the target antigen. These spheroids were co-cultured with primary T cells and TCE01C in a dose-dependent manner (Figure 2H). Therapeutic effects were quantified via fluorescence-based assays and IFN-γ secretion. TCE01C eliminated 82.45% of CRC cells within 48 h, accompanied by a maximal IFN-γ release of 9.2 ng/mL, while exhibiting minimal to no cytotoxicity toward non-malignant ASCs and HUVECs (Figures 2I–2K). Collectively, these findings highlight the robust efficacy of TCE01C across diverse patient-derived T cells and in complex 3D tumor microenvironments.

Specificity of TCE01C

Investigating the specificity of bispecific T cell engagers is critical for their successful therapeutic development. For the p53R273H-targeting TCE01C, specificity was evaluated across multiple dimensions: (1) peptide specificity, assessed via analysis of key interaction residues and homologous peptides, (2) HLA restriction, determined by comparing recognition of p53R273H presented by different HLAs alleles, and (3) TCE sensitivity to peptides presented specifically by HLA-C∗01:02. Given that the p53R273H 9-mer H5 epitope is presented by HLA-C∗01:02, structurally similar peptides may also be presented by this HLA molecule, potentially inducing conformational changes that facilitate TCE binding and subsequent T cell-dependent cytotoxicity. To explore this possibility, we first designed and synthesized a panel of 9-mer peptides harboring substitutions at the p53 codon 273 position, analogous to the p53R273H 9-mer H5 used herein. These included variants with documented or predicted associations to the onset and progression of diverse cancers (Tables S2 and S4), encompassing R273H, R273C, R273L, R273G, R273P, and R273S. Among these p53R273 variants, TCE01C elicited T cell-mediated killing against four variants: the primary target R273H as well as R273C, R273L, and unexpectedly wild-type R273 (R273WT). On-cell binding assays corroborated these findings, showing clear population shifts indicative of TCE01C engagement with the same four targets (Figures 3A and 3B). In addition, we titrated p53 R273H, R273WT, and R273C peptides from 0.002 to 20,000 nM. All three variants achieved near-maximal efficacy at 2,000 nM, with R273H exhibiting superior potency compared with others (Figure S2C). These results demonstrated that TCE01C recognizes and functionally engages additional p53 R273 variants beyond the intended p53R273H target.^[1],^[40] Notably, because wild-type p53 protein undergoes continuous proteasomal degradation and MHC presentation in normal cells, the observed recognition of the p53 R273WT peptide by TCE01C is particularly concerning, as it demonstrates the high sensitivity of the molecule toward the wild-type sequence. In addition, we subsequently examined TCE01C activity against homologous peptides to the p53 R273H 9-mer H5 (SFEVHVCAC) (Table S4). Homology searches were conducted using NCBI Blast and ExPAsy’s ScanProsite, permitting up to seven mismatches as the threshold.^[15],^[22],^[35],^[41],^[42] Notably, no off-target peptides with a single mismatch outside R273 position (fifth residue in the 9-mer) were identified. Among these off-targets, TCE01C exhibited detectable activity solely against off-target 01 (SFEIHVCSC) (Figure 3C).

The observed activity of TCE01C toward R273 variants, but limited response to potential off-targets, suggests possible off-target effects potentially attributed to (1) intrinsic TCE binding properties, (2) critical peptide interaction residues, or (3) HLA-C∗01:02 presenting capacity. To probe this, we performed alanine scanning by generating single alanine-substituted variants at each position (P1–P9) of the p53R273H 9-mer H5 peptide and evaluated TCE01C functional activity in CellTiter-Glo assays (Table S4). Of the eight alanine-substituted peptides, only V6A variant retained full activity comparable to the original p53R273H 9-mer H5. Substitutions S1A, H5A, and C7A reduced TCE01C efficacy by approximately 50%, whereas E3A and C9A substitutions caused substantial reduction. Notably, alanine replacements at positions F2A or V4A completely abolished TCE01C functional activity (Figure 3D). As TCE01C also recognized p53 R273H 9-mer H6 with moderate efficacy (Figure 1E), we extended alanine scanning to each position of this peptide (Table S4). Complete abrogation was observed at S2A, E4A, and C8A substitutions, while N1A, V5A induced partial reduction in efficacy, mirroring patterns in the p53 R273H 9-mer H5 alanine variants. Conversely, F3A, H6A, and V7A substitutions did not impair TCE01C function (Figure 3E). Collectively, these data indicate that the second, fourth, and eighth residues on p53 R273H 9-mer peptide are pivotal for HLA-C∗01:02 presentation and TCE01C binding efficacy.

To elucidate the structural basis of TCE01C interactions, we employed state-of-the-art computational modeling approaches, specifically Boltz-2 and AlphaFold 3, to predict ternary complex comprising TCE01C, HLA-C∗01:02 with p53R273H 9-mer H5, p53R273H 9-mer H6, or p53R273H 10-mer H9.^[43],^[44] These models were generated to identify residues critical for complex. In agreement with our in vitro findings, the second and eighth residues were predicted to be essential for the stable presentation of 9-mer H5, 9-mer H6 peptides, whereas the second residue of 10-mer H9 exhibited instability within the ternary complex. Additionally, the fourth residue in 9-mer H5 was also predicted to contribute to complex stabilization, such that alanine substitution at this position markedly impaired both peptide presentation and TCE01C binding efficacy (Figures S2D–S2F). Across these models, a congruent pattern emerged regarding the key governing complex integrity.

Computational optimization of p53R273H-targeting TCE

The observed cross-reactivity of TCE01C toward R273 variants, coupled with limited responses to potential off-targets, suggests its potential utility in targeting additional other p53 R273 variants across relevant cancer types, particularly given that p53 R273 represents a hotspot mutation in human malignancies. A TCE capable of recognizing multiple variants could function as an off-the-shelf therapy, applicable to a broader patient cohort spanning diverse tumor indications. Supporting evidence demonstrates that TCE molecules engaging multiple tumor-associated antigens or variants confer superior tumor recognition and a broader antitumor efficacy relative to those with stringent single-target specificity.^[45] However, such broadened specificity heightens the risk of cross-reactivity with wild-type p53 or other unrelated self-peptides, which may emerge stochastically in heterogeneous patient population contingent on individual conditions.^[46],^[47] Although mutant p53 R273H is generally absent in normal cells, inadvertent recognition of wild-type p53 could potentially trigger severe systemic autoimmunity or on-target/off-tumor toxicity.^[5],^[48],^[49] Furthermore, TCRs that can target several variants may exhibit diminished affinity for individual targets, thereby constraining T cell activation and tumoricidal potency. To address this potential trade-off between breadth and potency, we pursued further optimization of our TCE platform, prioritizing refinements to the TCR domain in context of TCE architecture.

We first predicted the structure of the pHLA-C∗01:02/p53R273H 9-mer H5 complex, by employing three independent computational modeling programs: AlphaFold 3, Boltz-2, and TCRmodel2.^[43],^[44],^[50] Even though AlphaFold3 and Boltz-2 are more generalist models, Boltz-2 is open, fast, and controllable while Alphafold3 provides the highest general accuracy. Several models were generated in parallel using these two models and ranked by ipTM score. On the other hand, TCRmodel2 was trained based on TCR docking database and used in this approach for cross validation. Top pHLA structure will then be used to predict the complex structure with addition of TCE01C using above-mentioned models. Leveraging this structure, we performed ML-guided directed evolution using three complementary computational protein design tools EvoPro, EvoProtGrad, and Antibody Evol^[27],^[51],^[52],^[53] (Figure 4A). Emergent candidates were evaluated for (1) predicted binding affinity (ddG, Prodigy, and BoltzGen), (2) structural stability (DeepSP and Aggrescan 3D), (3) immunogenicity (ABImmpred and Hu-mAb), and (4) toxicity (TAP).^[44],^[54],^[55],^[56],^[57] Among these criteria, the binding affinity was used to rank top candidates, while the later ones, including stability and immunogenicity, were included to predict safety and efficacy of generated candidates. The highest-ranking candidate from each iteration served as the parental sequence for subsequent rounds of directed evolution, with optimized residues fixed thereafter. This iterative workflow was complete after three cycles of computational directed evolution (Figures 4A and S3A; Table S5). Notably, evolutionary modifications were not confined to CDRs, as accumulating evidence underscores the contributions of non-CDR domains to conformation dynamics, binding affinity, interface stability, and overall functional efficacy in antibodies and related engagers.^[58],^[59],^[60],^[61] Modeling of the TCE01C/pHLA-C∗01:02/p53R273H 9-mer H5 ternary complex reveal asymmetric engagement restricted to one flank of the HLA-peptide binding groove, alongside disproportionate α/β chain positioning, with α chain oriented (Figures 4B and S3B). Prior investigations have established the pivotal role of CDR3β in TCR functionality. ML-optimized candidates exhibited progressive conformational refinements, facilitating comprehensive interaction across the peptide-binding groove and equitable alignment of both TCR chains toward the epitope (Figures S3C–S3E).

Optimized engagers displayed enhanced on-cell binding efficacy toward pHLA-C∗01:02/p53R273H 9-mer H5

One of the current challenges in developing novel proteins through computational programming is confirming whether the predicted performance of new variants in silico translates faithfully to experimental reality in both 2D and 3D models. To mitigate this risk and bridge the in silico-in vitro divide, top candidates from each evolutionary round of directed evolution underwent empirical validation in cell-based assays prior to advancement. Candidates were evaluated for binding efficiency and T cells-dependent cytotoxicity. BLI revealed progressive improvements in TCR affinity for stabilized pHLA-C∗01:02/p53R273H 9-mer H5 complex, with Kd values declining from 20.22 nM (TCE01C) to 12.51 nM (TCE01Cr1), 3.12 nM (TCE01Cr2), and for the final lead variant 2.51 nM (TCE01Cr3) (Figure 4C). This increase in in vitro binding affinity aligns closely with the corresponding decrease in predicted Kd values obtained from in silico evaluations across the optimization rounds (Table S5). On-cell binding assays further confirmed these improvements, demonstrating progressively enhanced affinities for the target pHLA-C∗01:02/p53R273H complex across successive variants. The final lead candidate, TCE01Cr3, exhibited the strongest on-cell binding signal, consistent with its superior biophysical profile (Figures 4C–4E).

To compare the functional potencies of optimized p53R273H-targeting TCE variants, we quantified T cell-mediated cytotoxicity and cytokine release in co-cultures with HLA-C∗01:02-expressing SW620 pulsed with p53R273H 9-mer H5 peptide. Dose-response curves for specific lysis revealed sigmoidal profiles for all TCE constructs, with maximal lysis approaching 100% at higher concentrations. The lead optimized variant, TCE01Cr3 (orange) exhibited the highest potency (EC50 ≈ 10–5 nM), followed by TCE01Cr2 (blue; EC50 ≈ 2 × 10−5 nM), TCE01Cr1 (green; EC^[44] ≈ 6 × 10−4 nM), and the parental TCE01C (red; EC50 ≈ 5 × 10−3 nM). Control conditions (TCR-Fc, CD3-Fc, and cells only) showed negligible activity (Figure 5A). Similarly, IFN-γ secretion followed analogous trends, with TCE01Cr3 achieving maximal release (∼7.85 ng/mL; EC50 ≈ 5 × 10−4 nM), outperforming TCE01Cr2 (EC50 ≈ 10–3 nM), TCE01Cr1 (EC50 ≈ 5 × 10−3 nM), and TCE01C (EC50 ≈ 10–2 nM), while controls remained baseline (Figure 5B). We next assessed the dependence of TCE-mediated killing on effector-to-target ratios (0:1, 0.1:1, 0.5:1, 1:1, 2:1, 5:1, and 10:1). All four variants TCE01C, TCE01Cr1, TCE01Cr2, and TCE01Cr3 induced strong target cell lysis even at low E:T ratios (as low as 1:1), with maximal killing above 85% reached at E:T ratios of 5:1 and 10:1. At an E:T ratio of 1:1, TCE01Cr2 and TCE01Cr3 achieved significantly higher cytotoxicity than TCE01C and TCE01Cr1 (one-way ANOVA, F3’8 = 56.5, p < 0.001), with pairwise differences ranging from 1.8-fold (TCE01Cr2 vs. TCE01C, Tukey-adjusted p = 0.0007) to 2.8-fold (TCE01Cr3 vs. TCE01Cr1, Tukey-adjusted p < 0.001). At an E:T ratio of 2:1, the same rank order was observed, but differences were smaller and only partially significant (F3’8 = 7.5, p = 0.010 overall). Only TCE01Cr3 vs. TCE01Cr1 remained significant after correction (1.4-fold, p = 0.007), while TCE01Cr2 vs. TCE01C did not (1.06-fold, p = 0.85) (Figure 5C). Integrating computational predictions, biophysical binding data, and preliminary cellular data, TCE01Cr3 emerged as the lead candidate for advanced evaluation. Thereafter, evaluations focused primarily on TCE01Cr3, incorporating ancillary variants where pertinent.

TCE01Cr3 demonstrated enhanced fidelity on HLA-C∗01:02/p53R273H 9-mer H5 tetramer

Similar to the parental TCE01C, comprehensive specificity profiling of ML-optimized TCE01Cr3 was imperative. The co-culture assays demonstrated robust T cell-mediated lysis of target cells presenting the p53R273H neoantigen with maximal specific lysis exceeding 80%. Residual cytotoxicity was still detectable against cells pulsed with the R273C and wild-type R273 peptides, although the magnitude of lysis was substantially reduced compared with the parental TCE01C (Figure 3A). In contrast, no measurable response was elicited against cells pulsed with R273L, R273G, R273P, R273S, or mock (no peptide) controls (Figure 5D). TCE01Cr3 exhibited strong binding to SW620 cells pulsed with p53 R273H 9-mer peptide, moderate binding to R273C and wild-type R273 peptides, and no detectable binding to R273L, R273G, R273P, R273S, or mock controls (Figure 5E). It is worth mentioning that in this approach, peptide carrying these p53 R273 variants were pulsed at the dose of 20 μM, which is among the highest doses of peptide being pulsed across similar studies^[23],^[39],^[62] and well beyond the normal presenting capacity of native neoantigen. For p53 R273 9-mer carrying either R273H, R273WT, or R273C, those that triggered TCE-mediated cytotoxicity of TCE01Cr3, sensitivity was evaluated in the peptide dose-dependent manner ranging from 20,000 to 0.002 nM. Specific lysis reached a maximum of ∼88%–90% at higher R273H peptide concentrations starting from 2 to 20,000 nM, with an EC50 in the nanomolar range. In contrast, both R273WT and R273C mutant peptide only trigger TCE-mediated killing at the dose of 20,000 nM peptide and dramatically reduce at lower dose (Figure 5F). Pulsing with the wild-type R273WT peptide resulted in negligible lysis across the entire concentration range. The closely related R273C mutant peptide induced only minimal lysis at the highest concentrations tested, confirming strong discrimination against both the wild-type sequence and this single-amino-acid variant.

Off-target evaluations yielded robust lysis (>80% maxima) exclusively for the R273H target, with negligible activity against on homologs or mock (Figure 5G). HLA allele restriction assays confirmed potent and specific lysis exclusively for HLA-C∗01:02 pulsed with p53R273H 9-mer H5, with negligible responses to HLA-A∗02:06, HLA-A∗11:01, HLA-B∗40:01, HLA-B∗54:01, HLA-C∗07:02, or mock (Figure 5H). These results confirm the exquisite allele and peptide specificity of TCE01Cr3 for the HLA-C∗01:02–p53R273H complex.

Similarly, Alanine scanning of the core p53R273H 9-mer H5 epitope (SFEVHVCAC) pinpointed residues F2, E3, V4, H5, V6, and C7 as critical, with substitutions eliciting marked reductions or abrogation of lysis, while S1A and C9A still exhibited weak cytotoxicity (Figure 5I). In silico ternary complex mapping of TCE, B2M/HLA-C∗01:02, and p53R273H 9-mer H5 highlighted predominant CDR involvement, with CDRβ1, CDRβ2, and CDRα3 favoring HLA-C∗01:02 contacts, and CDR3α/β involved in both HLA and peptide interaction, similar to previous studies on other TCR/HLA/peptide interactions.^[22],^[35],^[63] Computational optimization notably amplified CDR3 engagement with presenting peptide residues, concurrent with augmented HLA-C∗01:02 interaction (Figure S3B; Table S6). On the other hand, while Alanine scanning on presenting peptide allows us to quickly identify critical hotspot residues where side-chain interactions can contribute significantly to binding affinity or specificity, we further investigated the full specificity profile particularly for the key residues essential for anchoring p53 R273H 9-mer peptide onto HLA-C∗01:02 or residue involved in interacting with TCR, namely, positions 2, 5, 6, and 8 (P2, P5, P6, and P8). Structural studies on other HLA/TCR complexes have suggested a pattern where P2 and P8 are essential form proper presentation of peptide onto HLA while P5 and P6 are more involved in interaction with TCR.^[62],^[64],^[65],^[66] This approach is necessary especially for the case of p53 R273H 9-mer H5 peptide in this study as the position P8 is the Alanine residue (SFEVHVCAC), which cannot be further examined from Alanine scanning perspective. A combinatorial peptide library was generated where each of the residues at position 2, 5, 6, 8, corresponding to F2, H5, V6, A8, respectively, was substituted with 20 amino acids (Figure 6A). The responses of TCE01C and TCE01Cr3 were tested with target cells cytotoxicity via CellTiter-Glo 2D. For position #2, we observed a high tolerance of TCE01C for hydrophobic amino acids, especially for the ones with aromatic side chain similar to the original residue Phenylalanine (F) (Figure 6B). The optimized version, TCE01Cr3, exhibited a substantial reduction in tolerance; however, still has some detectable off-target activity if this position is switched to the hydrophobic category, even though at a lower level (Figure 6B). For position #5, TCE01C showed some levels of tolerance toward hydrophobic amino acids, especially for the negatively charged residues besides some hydrophobic ones like Alanine (A) or Leucine (L). Optimized TCE01Cr3 was in general more sensitive to substitution at this position with low-to-moderate activity for hydrophilic residues like C5, N5, K5 (Figure 6C). Position #6 showed a tremendous off-target issue for TCE01C if substituted to hydrophobic category, especially for the aliphatic group. This issue was completely solved in TCE01Cr3 when it was only sensitive to Valine at the 6th position. Cross reactivity at the theoretical C-terminal anchoring A8 was observed against a substitution with other amino acids also belong to the hydrophobic family beside hydrophilic T8 and Q8 for TCE01C. On the other hand, TCE01Cr3 shows more stringent fidelity at this position with only Alanine showed the robust cytotoxicity (Figure 6E).

We then further tried to use ScanProsite to analyze homologous peptides that meet the responsive motifs of TCE01Cr3 observed from the peptide combinatorial library analysis above. This analysis results in four additional homologous peptides (Uniprot: A0R2I7, SAEVHVVTC; Uniprot: A3QFJ5, SSEVKVCAD; Uniprot: E1BBQ2, SEEVHVCLP; and Uniprot: P34388, IPEVHVCAV). These four additional homologous peptides carry 3–4 mismatches comparing with p53 R273H 9-mer H5 peptide we used throughout this study. However, none of them were from human genome (Table S4). In addition, sensitivity against peptides presented by HLA-C∗01:02 were also investigated. The interaction with HLA-C∗01:02 will anchor TCR regions of TCE01Cr3 in the positions where they are exposed to peptides presented by this HLA, and thus, can trigger potential off-target activity. As a result, evaluating peptides presented by this HLA can provide us another insight toward TCE01Cr3 specificity. Using Immune Epitope Database & Tools (IEDB), we found 2,924 peptides reported to be presented by HLA-C∗01:02. Based on the off-target pattern when performing 20 amino acid-scanning for TCE01Cr3 on p53 R273H 9-mer, we picked the top 10 candidates for quick evaluations (Table S4, off-target 17–26). These candidates all carry Phenylalanine at the second position (P2), which matches the p53 R273H 9-mer peptide and can serve as an important anchor point for peptide presentation. Furthermore, they also carry matches at the fifth (Off-target 18), sixth (Off-target 17–24), or eighth (Off-target 17, 25, 26) positions, making them more vulnerable to TCE01Cr3 recognition. A T cell-dependent cytotoxicity assay was carried out for these candidates in comparing with p53 R273H 9-mer H5 peptide. In our hands, these 10 candidates did not trigger significant off-target activities from TCE01Cr3, with an exception of Off-target 17, which also carry F2, V6, and A8 residues, similar to p53 R273H 9-mer peptide (Figure 6F).

Immune response of TCE01Cr3 across diverse cell types

We further examined TCE01Cr3 efficacy against endogenous p53R273H expression using the SW620 colorectal cancer line, which harbors the heterozygous p53R273H mutation.^[54] Initial cytotoxicity assays using SW620 cells engineered to express HLA-C∗01:02 and pulsed with p53 R273H peptide revealed a baseline-specific lysis of 15%–20% killing even at low doses of optimized variants (TCE01Cr1, TCE01Cr2, and TCE01Cr3), whereas the parental TCE01C showed no detectable activity under the same conditions (Figure 5A), prompting scrutiny of hypersensitivity or cross-reactivity. While cross-reactivity for TCE01Cr3 was partially investigated via off-target screening, Alanine scanning, and combinatorial peptide screening at essential residues (Figures 5D–5I and 6A–6E), it remains unknown whether TCE01Cr3 is hypersensitive, especially to endogenous p53 R273H as SW620 cell line used throughout this study is heterozygous for TP53R273H. To determine whether heterozygosity contributed to this sensitivity, CRISPR/Cas9 was employed to edit p53 exon 8 flanking R273 with dual gRNAs and a repair template encoding 818819GT > AC modification (Figure S4A). Cells transfected with plasmid expressing Cas9 and gRNAs were isolated for further evaluation (Figure S4B). Among these cells, editing was observed in approximately 12.27% population (Figure S4C). Clonal derivatives were vetted by variant-specific PCR and p21 expression (indicative of wild-type p53 ablation) (Figures S4D–S4F). Resultant lines included homozygous wild-type p53 and compound heterozygous p53R273H (native 818G > A plus CRISPR 818819GT > AC). Modifications were verified by Sanger sequencing, followed by NGS confirming on-target modification and negligible off-target edits (Figures S4G and S4H). Notably, homozygous p53R273H clones (818G > A or 818_819GT > AC) proved inviable for sustained culture (data not shown), while the compound heterozygous cell line with two different mutated alleles encoding the same p53 R273H was obtained. As a result, in related experiments, we used this cell line in parallel with homozygous p53R273WT- and heterozygous p53R273H-cell lines. In these endogenous models, TCE01Cr3 achieved 47.81% cytotoxicity at 3.9 × 10−2 nM in compound heterozygous p53R273H-expressing cells versus 26.44% in heterozygotes. Comparatively, TCE01C yielded 26.82% and 23.56%, respectively, at 4.0 nM. Negligible killing was observed in SW620 expressing HLA-C∗01:02 and harboring homozygous wt-p53 as well as p53R273H SW620 expressing HLA-A∗11:01 (Figure 6G).

In addition, we further evaluated the potency of TCE01Cr3 across a panel of cell lines, including cancer cell lines SW620 (colorectal carcinoma), NCI-H520 (lung cancer), MDA-MB-468 (breast carcinoma), and PANC-1 (pancreatic cancer), as well as non-cancerous lines HEK293 (human embryonic kidney) and CCL-75 WI-38 (human lung fibroblasts). SW620 is heterozygous for TP53 R273H, MDA-MB-468 and PANC-1 are homozygous for TP53 R273H while NCI-H520 is homozygous wild-type at this locus, similar to non-malignant cells. Target cells with endogenous p53 R273H peptide were co-cultured with TCE01Cr3 and T cells at the E:T ratio of 10:1 and assessed for viability at 0, 24, 48, and 72 h. We observed a significant T cell-mediated cytotoxicity in all three cancer cell lines, albeit with varying efficiency. The strongest killing was observed in MDA-MB-468 and PANC-1 cells expressing HLA-C∗01:02 at 48–72 h, followed by SW620 under similar condition. In contrast, neither significant cytotoxicity was detected in cancer cell lines missing HLA-C∗01:02 nor the non-cancerous HEK293 or WI-38 cell lines (Figure 6H).

Stability evaluation for TCE targeting HLA-C∗01:02/p53 R273H

We further probed TCE01Cr3 stability. TCE01Cr3 employs a truncated (mini) single-chain architecture, in which the TCR constant domains are omitted, potentially influencing not only functionality but also stability. As the stapled-single-chain approach remains emergent in the field, this feature was carefully considered in our designs.^[26] Independent studies have suggested that the Fc domain can extend the half-life of bispecific T cell engager molecules.^[22],^[35],^[63] To determine whether this stability profile was specific to the TCE01C-derived lineage or generalizable across formats, each of the four architectures used in this study was intravenously administered to Swiss mice, and stability was evaluated by measuring residual serum concentrations over 72 h by ELISA, with half-life derived by noncompartmental analysis (Phoenix WinNonlin 8.5.2.4, Model 201) from the terminal log-linear elimination phase of each animal’s concentration-time profile. Specifically, TCE01Ar3 (platform A) and TCE01Br3 (platform B) retain both the variable and constant TCR regions, whereas TCE01Cr3 (platform C) and TCE01Dr.3 (platform D) employ the truncated, stapled single-chain architecture. Serum concentrations of all four formats were compared over the same time course.

Our results showed that TCE01Cr3 had a half-life of 39.40 ± 10.50 h (Figure 7A; Table 1). To determine whether the stapling modification itself affects systemic stability, we compared TCE01Cr3 with other three formats and observed comparable half-lives across all four (TCE01Ar3: 43.72 ± 10.71 h, N = 4; TCE01Br3: 48.12 ± 24.62 h, N = 6; TCE01Dr.3: 30.85 ± 16.36 h, N = 6) (Figures 7A; Table S1), with no evidence that the stapled formats (TCE01Cr3 and TCE01Dr.3) cleared faster than their non-stapled counterparts (TCE01Ar3 and TCE01Br3). These data indicate that within this matched set of optimized formats, the stapling strategy did not compromise systemic stability.

To further investigate the serum stability of these formats in vitro, each molecule was pre-incubated in human serum at 37°C for 0–128 h. Following pre-incubation, the samples were co-cultured with primary T cells and pHLA-C∗01:02-expressing target cells pulsed with p53 R273H peptide for an additional 24 h, and stability was evaluated indirectly by measuring IFN-γ release as a functional readout. Prolonged pre-incubation led to lower IFN-γ release, particularly in samples pre-incubated for over 16 h.^[22] To accurately assess how changes in serum TCE stability affect downstream functionality, IFN-γ levels were normalized to values obtained from the corresponding 0-h pre-incubation samples (defined as 100% activity) with mock-treated controls set to 0%. TCE01Cr3 activity decreased by ∼15% by the first hour, followed by a further reduction to 20%–40% loss by 2–4 h and ∼30% loss by 8 h. With extended incubation, the rate of decline slowed considerably, stabilizing at 128 h with 20.43% residual activity. The other three formats showed a comparable overall decline pattern over the same time course (Figure 7B). These results indicate that the stapling strategy had only a minimal impact on the long-term serum stability of TCE01Cr3, consistent with the comparable four-format profile observed in Figure 7A.

ML-derived TCE01Cr3 requires structural refinement on both CDR and non-CDR regions

Even though in-silico ternary complex mapping of TCE01Cr3, B2M/HLA-C∗01:02, and p53R273H 9-mer H5 highlighted predominant CDR involvement, it overlooked contributions to conformational folding, as evidenced by distinct spatial orientations between TCE01C and TCE01Cr3 in relation with B2M/HLA-C∗01:02/p53R273H 9-mer H5 (Figures S3B–S3E; Table S6). In addition, for the optimized candidates, we also noticed some mutations outside of the CDR regions and wondered how much these replacements contribute to the activity of the optimized versions (Figure 4B). To determine whether non-CDR modifications are essential for the binding and functionality of p53-targeting TCEs, we generated hybrid variants by swapping CDR and non-CDR regions between TCE01C and TCE01Cr3. Specifically, the non-CDR framework of TCE01Cr3 was integrated into TCE01C to yield TCE01C-S, while the non-CDR framework of TCE01C was integrated into TCE01Cr3 to yield TCE01Cr3-S (Figure S5A). In silico 3D structure prediction revealed greater similarity between TCE01C-S and TCE01Cr3, and between TCE01C and TCE01Cr3-S (Figures S5B–S5E). Moreover, dose-response analyses of cell killing and IFN-γ release assays demonstrated reduced efficacy for both TCE01C-S and TCE01Cr3-S relative to TCE01Cr3 efficacy, although both hybrids outperformed TCE01C (Figures S5F and S5G). Notably, removal of the TCR non-CDR regions from TCE01Cr3 markedly diminished killing activity, whereas grafting these non-CDR regions onto TCE01C-S substantially enhanced its potency. On the other hand, specificity was evaluated via co-culture with SW620 expressing HLA-C∗01:02 and pulsed with p53R273C 9-mer C5, with TCE01C-S and TCE01Cr3-S show intermediate level of off-target activity (Figure S5H). These findings together indicate that optimizations confined to either CDRs or non-CDRs are insufficient to fully enhance the activity and specificity of ML-derived TCEs, necessitating coordinated modifications across both regions.

Notably, given TCE01C sensitivity to HLA-C∗01:02/p53R273C 9-mer C5, we applied analogous to ML-directed procedures and criteria to develop TCE04C with improve sensitivity and specificity on this target (Figure S5I). TCE04C exhibited enhanced cytotoxicity and specificity toward HLA-C∗01:02/p53R273C (Figures S5J and S5K), warranting further investigation in subsequent studies.

Optimized TCE variants elicit potent T cell-mediated cytotoxicity in 3D multicellular tumor spheroids

To evaluate the antitumor efficacy of optimized TCE variants in complex 3D tumor microenvironments, multicellular spheroids were assembled comprising p53R273H 9-mer H5-pulsed CRC expressing HLA-C∗01:02 (green fluorescence), human umbilical vein endothelial cells (HUVECs; red), and adipose-derived stem cells (ASCs; blue). Spheroids were then treated with TCE01Cr3 in the presence or absence of primary T cells, either alone or in combination with chemotherapeutic agents (oxaliplatin or irinotecan) and monitored via brightfield and fluorescence microscopy at 0, 24, and 48 h. Viability of cancer cells expressing HLA-C∗01:02 and pulsed with p53R273H peptide, non-malignant HUVECs and ASCs were assessed via changes in green, red, and blue fluorescence signal over incubation time. TCE01Cr3 induced progressive disassembly and loss of green fluorescence in T cell-containing conditions, indicative of targeted CRC lysis. Co-treatment with oxaliplatin or irinotecan did not augment this effect, and controls lacking TCE or T cells preserved spheroid integrity (Figure 7C). TCE01Cr3 induced dose-dependent, T cell-mediated spheroid lysis, with superior efficacy when used alone compared to combinations with oxaliplatin or irinotecan, while no-TCE and mock conditions remained baseline (Figure 7D). IFN-γ secretion mirrored these trends, with TCE01Cr3 and TCE01Cr3/oxaliplatin yielding maxima ∼8.42 ng/mL, but different EC50 values (0.051 and 0.49 nM, respectively), followed by TCE01Cr3/irinotecan (∼5.24 ng/mL, EC50 = 3.3 nM) (Figure 7E).

Spheroids harboring CRC cells with endogenous p53R273H mutations (homozygous or heterozygous) were similarly assessed for up to 72 h. TCE01Cr3 induced rapid spheroid disruption and loss of live-cell fluorescence in p53R273H compound heterozygous (R273H/R273H) CRC spheroids, with intermediate effects in heterozygous (R273H/R273WT) spheroids; parental TCE01C showed only weak activity (Figure 7F). Dose-response curves revealed higher potency against homozygous targets (79.85% max lysis, EC50 ≈ 2.3 nM) than heterozygous targets (54.32% max lysis, EC50 ≈ 8.2 nM), underscoring robust efficacy of ML-derived variant (Figure 7G). These findings highlight strong chemotherapy-independent efficacy of the optimized variant in a 3D endogenous tumor model.

Discussion

We describe the development, optimization, and functional characterization of a bispecific T cell engaging receptor (TCE) targeting the recurrent p53R273H neoantigen presented by HLA-C∗01:02. Starting from a TCR isolated from a colorectal cancer patient whose T cells had been exposed with p53 R273H 25-mer peptide, we generated a panel of TCE constructs and systematically evaluated their antigen-presentation requirements, binding kinetics, specificity, potency, and activity in both in vitro 2D and 3D tumor models.

A central finding was the highly allele-specific presentation of the p53R273H epitope. Among the patient’s six HLA class I alleles, only HLA-C∗01:02 efficiently presented the immunodominant 9-mer peptide (SFEVHVCAC), resulting in robust T cell-dependent cytotoxicity. This selectivity explains the inconsistent clinical responses observed with p53-targeted therapies and highlights the necessity of HLA genotyping for patient stratification.

Conversion of the TCR into a stapled single-chain TCE format (miniTCE) reduced the molecular weight by ≈ 24% while preserving binding affinity to pHLA-C∗01:02/p53R273H and significantly enhancing cytotoxic potency, T cell activation, and cytokine release compared with full-length formats. The stapling approach, adapted from recent innovations in scFv stabilization, introduces disulfide bonds via targeted cysteine substitutions (e.g., M43C in β-chain and A100C in α-chain) and a hinge-like linker, promoting proper chain pairing and conformational integrity outside the CDR regions. This not only mitigates aggregation risks associated with traditional single-chain constructs but also improves serum stability, as evidenced by sustained IFN-γ induction after prolonged pre-incubation. Such enhancements could extend half-life and bioavailability in vivo, addressing a common limitation of bispecific engagers.

The pHLA-binding affinity of TCR01, as measured within the TCE01 scaffold, exceeds the range typically reported for naturally occurring TCR-pHLA interaction (1–100 μM). This likely reflects both the functional selection strategy used to isolate TCR01-which enriched for clones with higher relative avidity for pHLA-C∗01:02/p53R273H based on IFN-γ response, consistent with avidity maturation during antigen-driven T cell clonal selection, and the streptavidin-based tetramerization of the immobilized pHLA ligand in the BLI assay, which can promote analyte rebinding and yield apparent affinities tighter than the true monomeric interaction. Prior work has established that monovalent TCEs incorporating physiological-affinity TCRs (μM range) fail to elicit robust T cell-mediated cytotoxicity, necessitating affinity-engineered TCRs or TCR-mimic antibodies for effective design, as exemplified by tebentafusp (ImmTAC), which pairs an affinity-engineered TCR with an anti-CD3 scFv, and by our previously reported KRASG12V-targeting TCEs, where high-affinity TCR incorporation was similarly required for potent T cell activation and cytotoxicity. The naturally elevated affinity of TCR01, arising from functional neoantigen-reactivity screening rather than deliberate engineering, is therefore consistent with – and likely necessary for – the potent monovalent TCE activity observed in this study. Building on this favorable starting affinity, we next sought to further refine specificity and potency through structure-guided engineering.

Subsequent machine-learning-guided directed evolution yielded the lead molecule TCE01Cr3, which exhibited markedly improved affinity (Kd = 2.51 nM), exceptional potency (EC50 ≈ 10–5 nM), and enhanced specificity relative to the parental TCE01C. Leveraging tools like EvoPro, EvoProtGrad, and Antibody Evol on predicted pMHC-TCR structures (from AlphaFold 3, Boltz-2, etc.), the iterative process refined non-CDR residues to optimize groove engagement and chain alignment, amplifying CDR3 interactions with the peptide while bolstering HLA contacts. Furthermore, cross-reactive parental TCE01C can be systematically engineered into highly specific candidates like TCE01Cr3 for TP53 R273H or TCE04C for TP53 R273C with substantially improved sensitivity and efficacy. This ML-driven strategy exemplifies how computational protein design can accelerate affinity maturation and specificity tuning, reducing off-target liabilities in a fraction of the time required for empirical library screening.

Alanine-scanning mutagenesis and AlphaFold 3/Boltz-2 modeling identified the second and fourth residues of the 9-mer peptide as critical anchors for both HLA-C∗01:02 binding and TCE01C recognition, providing a structural rationale for its selectivity profile. Furthermore, positional scanning mutagenesis at key residues combined with cell-based functional assays (Figures 6A–6E) provides deeper insight into specificity profiles of these engineered TCEs. Overall, TCE01C displayed broad tolerance to conservative substitutions, accommodating peptides in which amino acids with similar physicochemical properties replaced the original residues at corresponding positions within the p53R273H 9-mer H5 epitope (SFEVHVCAC). In contrast, the optimized TCE01Cr3 exhibited higher stringency, showing strong selectivity even for substitutions with similar physicochemical characteristics, thereby enhancing overall specificity while retaining potent toxicity. TCE01Cr3 thus represents a significant advance in targeting p53R273H, a hotspot mutation driving oncogenesis in diverse cancers. By achieving exquisite specificity for the mutant epitope with minimal cross-reactivity to wild-type p53 or homologs, it mitigates risks of autoimmunity while retaining broad applicability as an off-the-shelf agent for HLA-C∗01:02-positive patients. Its superior performance against endogenous p53R273H in heterozygous, compound heterozygous, and homozygous mutant models while keeping minimal activity in wild-type models underscores its potential to address tumor heterogeneity.

Nevertheless, several limitations should be acknowledged. The positional scanning approach in this study focused primarily on four key positions (F2, H5, V6, and A8) within the 9-mer epitope. While this provided critical insights into TCR-pHLA interactions, a more comprehensive combinational mutagenesis library across all epitope positions, combined with deeper structural analysis (e.g., cryo-EM or high-resolution crystallography of the ternary complex), would further refine specificity predictions and identify additional tolerated or intolerant residues. In addition, early-stage validation during TCE evaluation relied on peptide-pulsing HLA-engineered models. While this approach enables rapid assessment of TCE efficacy and specificity, it does not fully recapitulate physiological conditions, in which endogenously processed neoantigens are presented at substantially lower densities on the cell surface. Although we evaluated TCE01Cr3 activity against cells harboring endogenous TP53 R273H mutations in both 2D and 3D spheroid models, limitations remain regarding efficacy, durability, and the frequency of patient samples that co-express HLA-C∗01:02 and the TP53 R273H mutation. Future studies will be required to further assess TCE01Cr3 sensitivity and therapeutic window under these more clinically relevant conditions.

Importantly, TCE01Cr3 retained high efficacy in multicellular 3D spheroids containing CRC cells, HUVECs, and adipose-derived stem cells, achieving >80% elimination of peptide-pulsed and endogenous p53R273H-expressing tumor cells within 48 h while sparing non-malignant stroma. Strikingly, TCE01Cr3 as monotherapy induced more rapid and extensive spheroid disassembly and higher maximum specific lysis than combinations with oxaliplatin or irinotecan in the 3D model. This superior performance in a physiologically relevant setting raises the possibility that affinity-matured TCE molecules targeting recurrent neoantigens such as p53R273H may achieve meaningful clinical benefit without the need for combination chemotherapy, thereby reducing the risk of overlapping toxicities and improving tolerability in patients. TCE01Cr3 showed a measurable safety margin against wild-type p53 or R273C. In peptide dose-titration assays (Figure 5F), R273H-mediated cytotoxicity approached its maximum at peptide concentrations as low as ∼2 nM, whereas detectable killing against the wild-type R273 or R273C mutant peptides required a ∼10,000-fold higher concentration (20,000nM) and remained near baseline at all lower doses tested. Because peptide pulsing loads HLA molecules well above the density achieved through endogenous antigen processing, the 20 μM dose used for specificity screening was deliberately chosen to exceed physiological neoantigen presentation. This in vitro margin likely understates the true separation between the mutant-activation threshold and the far lower density of wild-type p53 peptide naturally presented on normal tissue. This remains an estimate derived from a peptide-pulsed surrogate system, however, rather than a direct measurement of endogenous wild-type p53 presentation on normal cells, and should be confirmed with HLA immunopeptidomics or normal-tissue/humanized in vivo models before clinical translation. The molecule also exhibited good mouse serum stability for nearly 40 h, supporting its suitability for systemic administration.

Clinical implications

TCE01Cr3 holds promise for precision oncology in p53-mutated cancers, where mutations affect >50% of malignancies (e.g., colorectal, lung, breast, ovarian). Targeting the frequent R273H neoantigen via HLA-C∗01:02 (∼5–10% global prevalence, higher in Asian/European populations) enables patient stratification through HLA typing and tumor sequencing, broadening access while avoiding futile treatments.

Its off-the-shelf format bypasses autologous therapy delays, ideal for relapsed/refractory cases with p53-driven chemoresistance. Nanomolar potency and stroma-sparing in 3D models imply strong tumor penetration. They also point to a favorable safety window: killing plateaued at ∼2 nM of mutant peptide, but needed ∼10,000-fold more wild-type peptide (20,000 nM) before any activity appeared (Figure 5F). This is an early, in vitro estimate of the therapeutic window against on-target/off-tumor toxicity in heterogeneous solid tumors, not a clinical safety guarantee. It should be confirmed by measuring actual wild-type p53 presentation density on normal tissue and by testing tolerability in vivo. The absence of added benefit from oxaliplatin or irinotecan indicates that TCE can achieve robust anti-tumor efficacy as a single agent, potentially obviating the need for concomitant chemotherapy and thereby avoiding chemotherapy-associated toxicities in patients.

The stapling and ML paradigms enhance biophysical stability, dosing, and pipeline efficiency for neoantigen therapies. Importantly, while this study has tremendously investigated the efficacy, sensitivity and specificity profile of TCE01Cr3 both in 2D and 3D models, we have not characterized its performance in animal models. The 3D spheroid system is more physiologically complicated than standard 2D co-culture, but it cannot fully substitute for an intact tumor microenvironment like patient-derived organoid or organ-on-a-chip models. We have therefore designed two complementary using NOD- or NOG-mouse xenograft studies using the HLA-C∗01:02/p53R273H-positive SW620 model with human PBMC effector cells: (1) an early-start design in which tumor cells and PBMCs/T cells are co-engrafted subcutaneously and a single intravenous dose is administered at 0.2, 1.0, or 5 mg/kg versus controls, with survival, body weight, and tumor growth inhibition monitored over time, and (2) an established-tumor design, where tumors are growth for 10 days with human PBMCs infusion on day 3, then a single therapeutic dose of TCE01Cr3 (1 mg/kg) administration on day 10, assessing survival, tumor regression, and T cells infiltration. Cytokine release syndrome risk, biodistribution, and immunogenicity will require dedicated follow-up studies. Furthermore, TCE01Cr3 efficacy on endogenously presented p53 R273H neoantigen did not achieve maximal killing, suggesting a potential need for combination with other treatment targeting p53 mutation. Nonetheless, these collective results establish TCE01Cr3 as a promising off-the-shelf candidate for patients whose tumors harbor the hotspot p53R273H mutation and the HLA-C∗01:02 allele.

Clinical implications

TCE01Cr3 holds promise for precision oncology in p53-mutated cancers, where mutations affect >50% of malignancies (e.g., colorectal, lung, breast, ovarian). Targeting the frequent R273H neoantigen via HLA-C∗01:02 (∼5–10% global prevalence, higher in Asian/European populations) enables patient stratification through HLA typing and tumor sequencing, broadening access while avoiding futile treatments.

Its off-the-shelf format bypasses autologous therapy delays, ideal for relapsed/refractory cases with p53-driven chemoresistance. Nanomolar potency and stroma-sparing in 3D models imply strong tumor penetration. They also point to a favorable safety window: killing plateaued at ∼2 nM of mutant peptide, but needed ∼10,000-fold more wild-type peptide (20,000 nM) before any activity appeared (Figure 5F). This is an early, in vitro estimate of the therapeutic window against on-target/off-tumor toxicity in heterogeneous solid tumors, not a clinical safety guarantee. It should be confirmed by measuring actual wild-type p53 presentation density on normal tissue and by testing tolerability in vivo. The absence of added benefit from oxaliplatin or irinotecan indicates that TCE can achieve robust anti-tumor efficacy as a single agent, potentially obviating the need for concomitant chemotherapy and thereby avoiding chemotherapy-associated toxicities in patients.

The stapling and ML paradigms enhance biophysical stability, dosing, and pipeline efficiency for neoantigen therapies. Importantly, while this study has tremendously investigated the efficacy, sensitivity and specificity profile of TCE01Cr3 both in 2D and 3D models, we have not characterized its performance in animal models. The 3D spheroid system is more physiologically complicated than standard 2D co-culture, but it cannot fully substitute for an intact tumor microenvironment like patient-derived organoid or organ-on-a-chip models. We have therefore designed two complementary using NOD- or NOG-mouse xenograft studies using the HLA-C∗01:02/p53R273H-positive SW620 model with human PBMC effector cells: (1) an early-start design in which tumor cells and PBMCs/T cells are co-engrafted subcutaneously and a single intravenous dose is administered at 0.2, 1.0, or 5 mg/kg versus controls, with survival, body weight, and tumor growth inhibition monitored over time, and (2) an established-tumor design, where tumors are growth for 10 days with human PBMCs infusion on day 3, then a single therapeutic dose of TCE01Cr3 (1 mg/kg) administration on day 10, assessing survival, tumor regression, and T cells infiltration. Cytokine release syndrome risk, biodistribution, and immunogenicity will require dedicated follow-up studies. Furthermore, TCE01Cr3 efficacy on endogenously presented p53 R273H neoantigen did not achieve maximal killing, suggesting a potential need for combination with other treatment targeting p53 mutation. Nonetheless, these collective results establish TCE01Cr3 as a promising off-the-shelf candidate for patients whose tumors harbor the hotspot p53R273H mutation and the HLA-C∗01:02 allele.

Materials and methods

Ethics Statement

Human blood samples were collected for research purposes in accordance with the protocols approved by the Ethics Committee of the Medical Genetics Institute (approval number 02/2024/CT-VDTYH) and conducted in accordance with the ethical principles of the Declaration of Helsinki (1964). Written informed consent was obtained from each participant in accordance with the Declaration of Helsinki. All procedures complied with the guidelines and regulations, and the sample collection was performed at the Medical Genetics Institute.### TCE engineering, cloning, and plasmid construction

For sequences of TCE variants used in this study, see Table S7. Plasmids used in this study are listed in Table S8. The original plasmids were sourced from BioIntron (pcDNA 3.4 (+)), Takara Bio (pLVX Puro 632164), Invitrogen (GeneArt CRISPR nuclease OFP plasmid A21174), and Invitrogen (pcDNA 3.1 (+) V79020). All constructs were designed in-house and assembled using Gibson-cloning (Table S8). TCE or HLA fragments were amplified using Q5 High Fidelity DNA polymerase (NEB M0491S) or GXL DNA polymerase (Takara R050A), followed by DpnI (NEB R0176L) treatment and isopropanol-based PCR cleanup. These fragments were then inserted into pcDNA 3.4 (+) backbones (for TCE structures) or pLVX (for lentivirus expressing plasmids) via Gibson assembly (Thermo Fisher A46628) following standard procedure. Assembled fragments were transformed into 10-beta competent cells (NEB C3019I) for TCE plasmids or Stbl3 (Thermo Fisher C737303) for lentivirus plasmids. Plasmids were verified via Sanger Sequencing at DNA Sequencing Vietnam.### Production and purification of TCEs

TCEs were produced and purified in-house. All TCEs were generated in HEK293T cells, using polyethyleneimine (PEI) 10 kDa (Sigma-Aldrich 765090) or Lipofectamine 3000 (Invitrogen L3000015) in OptiMEM medium (Gibco 11058521) at 37°C, and 5% CO2 for 4–7 days. Expression was assessed pre-purification by dot-blot or SDS-PAGE with silver staining or ELISA using anti-Human IgG (Fc specific)-HRP.

In-house purification was conducted using an ÄKTA system (Cytiva) with a HiTrap MabSelect PrismA column (Cytiva 17549853) following the manufacturer’s recommendation with some adjustments where column was subjected to gradual decrease in Citrate pH from 7.2 to 3.5. Fractions were evaluated by 8%–10% SDS-PAGE with silver staining and monomers were further purified via size-exclusion chromatography. High-yield, high-purity fractions were pooled and buffer-exchanged into storage buffer (0.55 mg mL−1 glutamic acid, 0.03 mg mL−1 polysorbate 80, 28.5 mg mL−1 sucrose; pH 5.2).### Cell lines and culturing

All cell lines were maintained at 37°C with 5% CO2. Human embryonic kidney cells for lentiviral production (HEK293T, Takara Bio 632273), colorectal carcinoma cells (SW620 and ATCC CCL-227), lung carcinoma cells (NCI-H520 and ATCC HTB-182), breast cancer cells (MDA-MB-468 and ATCC HTB-132), pancreatic cancer cells (PANC-1 and ATCC CRL-1469), human adipose-derived stem cells (ASC), lung fibroblast (CCL-75 WI-38) were adapted and cultured in DMEM (Thermo Fisher 11995081) supplemented with 10% heat-inactivated fetal bovine serum (FBS; Thermo Fisher A5256701). Cells revived from cryopreserved stocks were expanded for 3–4 passages prior to use.### PBMC collection and T cell isolation

PBMCs were isolated from donor blood samples using Lymphoprep (Stemcell Technologies 07861) according to the manufacturer’s protocol with minor modifications. Briefly, diluted blood was layered over Lymphoprep and centrifuged at 1,000 × g for 30 min at room temperature with acceleration and brake off. PBMCs were then either frozen immediately for storage or subjected to T cells isolation within no more than 2 days.

CD3+ T cells were isolated from PBMCs using the EasySep Human T cell Isolation Kit (Stemcell Technologies 17951) following the manufacturer’s instructions. Lymphocytes were resuspended at 5 × 107 cells mL−1 and incubated with Isolation Cocktail (50 μL mL−1) for 5 min at room temperature, followed by addition of RapidSpheres (50 μL mL−1). The mixture was topped up to 2.5 mL with wash medium (PBS containing 2% FBS and 1 mM EDTA, Ca2+/Mg2+-free). Non-CD3+ cells were magnetically captured, and CD3+ T cells in the supernatant were collected and resuspended in AIM-V medium.### Cell culture transient transfection

For PEI-mediated transfection, cells were cultured in antibiotic-free medium, briefly washed with pre-warmed 1× PBS (Cytiva SH30258.02), and cultured in 80% volume of Opti-MEM (Thermo Fisher 11058021). Plasmid DNA was complexed with PEI (10  kDa; Sigma-Aldrich 765090) at a 1:3 (w/w) ratio in Opti-MEM for 20–25 min at room temperature or 37°C and added dropwise to cells. For Lipofectamine 3000-mediated transfection, plasmids were prepared according to the manufacturer’s protocol (Thermo Fisher L3000015) and incubated with cells for 6 h before replacement with fresh complete medium.### Lentivirus production and transduction

HEK293T cells (passage number 95%. Peptides were resuspended in appropriate solvent following the manufacturer’s certificate of analysis (CoA) at the concentration of 20 mM and stored as aliquots at −80°C. Each peptide batch was subjected to no more than two freeze-thaw cycles.### Peptide pulsing and co-culturing

For peptide pulsing, NCI-H520, MDA-MB-468, and SW620 cells expressing the HLA allele of interest were seeded at a density of 6.25 × 104 cells per cm2 in AIM-V medium supplemented with 0%–2% FBS and the relevant peptides at the indicated concentrations. Cells were incubated for 18–24 h at 37°C under 5% CO2. Pulsed cells were then washed once with phosphate-buffered saline (PBS) and once with AIM-V medium supplemented with 5% FBS, resuspended, and seeded at a density of 2 × 105 cells per ml.

For co-culture experiments, effector cells (PBMCs or T cells) were expanded in AIM-V medium supplemented with 5% FBS and subsequently mixed with target cells at the specified effector:target (E:T) ratios in the presence of relevant T cell-engaging receptors (TCEs). Co-cultures were maintained for 24–48 h prior to downstream analysis.### Flow cytometry and cell-based staining

For a complete list of antibodies used in this study, see Table S9. For cell-surface staining, cells were washed three times in ice-cold 1× PBS, resuspended in ice-cold 1× PBS, and incubated with fluorophore-conjugated antibodies for 30 min at 4°C in the dark. Cells were washed three times in ice-cold 1× PBS before analysis. For intracellular perforin and granzyme B staining, T cells were incubated with the protein transport inhibitor monensin (eBioscience, cat. no. 00–4505-51) for 4–6 h at 37°C under 5% CO2. Cells were then permeabilized using Intracellular Staining Permeabilization Wash Buffer (BioLegend, cat. no. 421002) for 45 min, washed twice, and stained with the relevant antibodies. Flow cytometry data were acquired on an Agilent NovoCyte flow cytometer.

Gating strategy: Singlet cells were identified using forward scatter (FSC) and side scatter (SSC) parameters. After gating on singlets, live cells of interest were defined using specific markers (CD3-PE-Cy7 for T cells and His-APC for SW620 cells). Markers of T cell activation and cytotoxic granule production were evaluated by measuring CD69-APC, perforin-AF700, and granzyme B-PerCP, with comparisons made between mock-treated and experimental samples under the indicated conditions.### Cell-based binding assay

For cell binding assays, target cells were incubated with relevant TCEs for 1 h in defined media and then proceeded with a secondary antibody for assessment via flow cytometry.### Cell-based functional assays

Peptide-pulsed cells were washed once in 1x PBS and once in AIM-V containing 2% FBS. Effector cells were cultured in AIM-V media supplemented with 2% for at least 24 h and checked for viability. Cells were then co-cultured with effector cells in the presence of relevant TCEs. The supernatant was collected for IFN-γ-release ELISA assay (Biolegend 430104). Simultaneously, T cells were collected for T cell activation assays as described earlier, and target cells were subjected to luminescence-based cytotoxicity assay using Celltiter-Glo (Promega G9242). Both functional assays were done following the manufacturer’s instructions.### In silico peptide evaluation and protein complex prediction

Peptide homolog identification was performed using NCBI Blast and Expasy ProScan with the UniProt protein database for Homo sapiens, using the cut-off of 4 mismatches. Homologous and alanine-scanned peptides were then evaluated with p53R273H 9-mer H5 peptide for binding affinity with HLA-C∗01:02 using NetMHC Pan 4.1 or MHC-I prediction. The binding models of TCEs with p53R273H/HLA-C∗01:02 and CD3 were done using a combination of Python-scripted Alphafold 3.0, TCR model 2.3, Boltz-2, and TCR dock. PyMOL was used to evaluate and optimize models’ visualization.### Machine learning-based directed evolution

In brief, a complex between B2M, HLA-C∗01:02, and p53 R273H peptide was generated via Boltz-2, Alphafold 3.0, and TCR model to ensure structure prediction consistency. Parental TCE was then added to the pHLA complex and new complex spatial structure was then generated and cross-evaluated across three above models. This complex was then subjected for in silico directed evolution via EvoPro, EvoProtGrad with maximum 50 mutations per round and 20 cycles of optimization per run, focusing on TCR region on TCE while leaving the anti-CD3, Fc of TCE, and pHLA remained unchanged. Per directed evolution rounds, top 10 TCE candidates with lower free energy (ΔG) and lower free binding energy (ddG) were selected for further evaluation, including binding affinity, stability, and aggregation using independent computational programs. For example, these candidates were tested for binding affinity against HLA and presenting peptide at both 25°C as well as 37°C via PRODIGY. Aggregation score was assessed via evaluating distribution of hydrophobic and hydrophilic residues along TCR sequences using Aggrescan 3D and DeepSP. In addition, candidates were also evaluated for immunogenicity potential using ABImmpred and Hu-mAb. Top candidate per round was chosen among the top performers for each category and served as the starting construct for the next directed evolution round.### Spheroid formation and drug administration

To evaluate the cytotoxic effects of T cell engagers (TCEs) alone and in combination with chemotherapy, 3D colorectal cancer (CRC) spheroids were generated. Oxaliplatin (MedChem Express, HY-17371) and irinotecan (MedChem Express, HY-16562) were employed as chemotherapeutic agents alongside TCEs. Spheroids were created by co-culturing SW620 CRC cells (ATCC, CCL-227), human umbilical endothelial cells (HUVECs) and human adipose-derived stromal cells (ASCs, Cellosaurus CVCL_4W37) at a 3:1:1 ratio, totaling 6,000 cells, in ultra-low attachment 96-well round-bottom plates (Corning Costar, 7007). The cells were suspended in 200 μL of AIM-V medium (Gibco, 12055091) supplemented with 5% FBS (Gibco, 10437028) and either the presence of 2 μM or the absence p53R273H 9-mer peptide H5 to promote TCE-specific antigen presentation on SW620 cells in pulsed or endogenous models. SW620 cells and ASCs were stably transduced with green fluorescent protein (GFP) and red fluorescent protein (RFP), respectively, to enable visualization of tumor-stromal interactions. The plates were centrifuged at 350 × g for 5 min and incubated at 37°C with 5% CO2 for 48 h to form compact spheroids.

Spheroids were co-cultured with 1 × 105 T cells and exposed to single agents (TCE01C, TCE01Cr3, oxaliplatin, or irinotecan) or their combinations. For single TCE treatments, TCE01C and TCE01Cr3 were diluted 4-fold across a 10-point range (0.015–60 nM). For single chemotherapy treatments, oxaliplatin (0.2–200 μM) and irinotecan (0.1–100 μM) were diluted 10-fold. In combination studies, TCEs (0.00015–60 μM) were combined with fixed sublethal doses of oxaliplatin (10 μM) or irinotecan (100 μM). Controls consisted of spheroids alone, spheroids with T cells, T cells alone, and T cells with oxaliplatin or irinotecan. After 48 h, spheroid morphology and GFP fluorescence were evaluated using an EVOS M5000 Imaging System (Thermo Fisher Scientific). GFP fluorescence intensity was quantified with ImageJ software (NIH, v1.53). Supernatants were collected for IFN-γ measurement via ELISA assay, as previously described.^[22]### Statistical analysis

T cell-mediated cytotoxicity was evaluated using the Celltiter-Glo assay, in which the amount of ATP from live cells was used to determine live cell density. The percentage of specific cell lysis was calculated using the following equation ^[1]:

For IFN-γ release ELISA, readouts were normalized to those of co-cultured samples without TCEs. Flow cytometry signals were normalized to T cells only, target cells only, or co-cultured samples without TCEs. Error bars represent the standard error of the mean (SEM). For multiple comparisons to the same control (in the donor’s evaluation experiment), we used one-way ANOVA, followed by Dunnett’s test. For multiple pairwise comparisons (in tumor cells assessment experiment), we applied one-way ANOVA coupled with Tukey’s honestly significant difference (HSD) test. At least three biological samples and three technical replicates were analyzed for each condition.### CRISPR genome editing

Region surrounding R273 residue of TP53 gene was analyzed and cross-compared via Cas-Designer (http://www.rgenome.net/cas-designer/), Geneious Prime, and Benchling CRISPR platform. Potential target sites were evaluated for Cas9 activity score and potential off-targets via Cas-OFFinder (http://www.rgenome.net/cas-offinder/). Top two candidates covering both sides of R273 residue were selected, and corresponding gRNAs were cloned into the multiplex system separated by tRNA. Plasmid carrying Cas9, multiplex gRNA was co-transfected with a single-strand donor DNA repair template into target cells of interest using PEI 10 kDa. DNA from individual clones was extracted using lysis buffer (50 mM KCl, 10 mM Tris-HCl pH 8.3, 2.5 mM MgCl2, 0.5% Tween 20, 1× Proteinase K). Extracted DNA was used for downstream analysis, including genomic PCR and NGS.### Pharmacokinetics study

For studying the PK, six male Swiss albino mice were randomly assigned to each experimental TCE group that represents platform A, B, C, and D, with ten mice in the PBS control group. Mice were injected i.v. through the tail vein with 5 mg kg−1 drug or PBS. Blood samples were collected at 1, 2, 5, 24, 30, 48, and 72 h. Samples were centrifuged, and serum fractions were stored at −20°C for later ELISA. Animal experimentation was approved by the Animal Ethics Committee at the University of Science in Ho Chi Minh City, Vietnam (protocol number 562/KHTN-ACUCUS).

For ELISA, samples (100 μL) diluted in PBS and dilution series of known concentrations were added to immuno 96-well plates and incubated for 2 h at RT. The wells were washed 3 times in PBS +0.05% Tween 20 (PBST), 300 μL per well. TCEs were detected using anti-Fc peroxidase antibody (Thermo Fisher Scientific, A18817) diluted 1:5,000 in 1% BSA in PBST, while TCER03 was detected using anti-DYKDDDDK TAG conjugated HRP (1:1,000, Thermo Fisher Scientific, A01428100). The wells were washed four times with PBS and developed with TMB (eBioscience, 00-4201-56). Each sample was measured in triplicate.

Pharmacokinetic parameters were calculated by noncompartmental analysis (NCA) using Phoenix WinNonlin 8.5.2.4 with model 201 (IV bolus administration). For each animal, the terminal elimination rate constant (λz) was estimated by log-linear regression of the terminal (declining) phase of the serum concentration over time profile, and the terminal half-life t1/2 was calculated as ln(2)/λz. Animals for which a terminal log-linear phase could not be reliably identified (e.g., due to non-declining or non-monotonic concentrations in the terminal window, consistent with assay noise near the lower limit of quantification or an atypical early-phase concentration rise suggestive of a dosing artifact) were excluded from λz and t1/2 estimation for that animal. The number of animals included in each group-level half-life estimate is reported in Table 1.

Ethics Statement

Human blood samples were collected for research purposes in accordance with the protocols approved by the Ethics Committee of the Medical Genetics Institute (approval number 02/2024/CT-VDTYH) and conducted in accordance with the ethical principles of the Declaration of Helsinki (1964). Written informed consent was obtained from each participant in accordance with the Declaration of Helsinki. All procedures complied with the guidelines and regulations, and the sample collection was performed at the Medical Genetics Institute.

TCE engineering, cloning, and plasmid construction

For sequences of TCE variants used in this study, see Table S7. Plasmids used in this study are listed in Table S8. The original plasmids were sourced from BioIntron (pcDNA 3.4 (+)), Takara Bio (pLVX Puro 632164), Invitrogen (GeneArt CRISPR nuclease OFP plasmid A21174), and Invitrogen (pcDNA 3.1 (+) V79020). All constructs were designed in-house and assembled using Gibson-cloning (Table S8). TCE or HLA fragments were amplified using Q5 High Fidelity DNA polymerase (NEB M0491S) or GXL DNA polymerase (Takara R050A), followed by DpnI (NEB R0176L) treatment and isopropanol-based PCR cleanup. These fragments were then inserted into pcDNA 3.4 (+) backbones (for TCE structures) or pLVX (for lentivirus expressing plasmids) via Gibson assembly (Thermo Fisher A46628) following standard procedure. Assembled fragments were transformed into 10-beta competent cells (NEB C3019I) for TCE plasmids or Stbl3 (Thermo Fisher C737303) for lentivirus plasmids. Plasmids were verified via Sanger Sequencing at DNA Sequencing Vietnam.

Production and purification of TCEs

TCEs were produced and purified in-house. All TCEs were generated in HEK293T cells, using polyethyleneimine (PEI) 10 kDa (Sigma-Aldrich 765090) or Lipofectamine 3000 (Invitrogen L3000015) in OptiMEM medium (Gibco 11058521) at 37°C, and 5% CO2 for 4–7 days. Expression was assessed pre-purification by dot-blot or SDS-PAGE with silver staining or ELISA using anti-Human IgG (Fc specific)-HRP.

In-house purification was conducted using an ÄKTA system (Cytiva) with a HiTrap MabSelect PrismA column (Cytiva 17549853) following the manufacturer’s recommendation with some adjustments where column was subjected to gradual decrease in Citrate pH from 7.2 to 3.5. Fractions were evaluated by 8%–10% SDS-PAGE with silver staining and monomers were further purified via size-exclusion chromatography. High-yield, high-purity fractions were pooled and buffer-exchanged into storage buffer (0.55 mg mL−1 glutamic acid, 0.03 mg mL−1 polysorbate 80, 28.5 mg mL−1 sucrose; pH 5.2).

Cell lines and culturing

All cell lines were maintained at 37°C with 5% CO2. Human embryonic kidney cells for lentiviral production (HEK293T, Takara Bio 632273), colorectal carcinoma cells (SW620 and ATCC CCL-227), lung carcinoma cells (NCI-H520 and ATCC HTB-182), breast cancer cells (MDA-MB-468 and ATCC HTB-132), pancreatic cancer cells (PANC-1 and ATCC CRL-1469), human adipose-derived stem cells (ASC), lung fibroblast (CCL-75 WI-38) were adapted and cultured in DMEM (Thermo Fisher 11995081) supplemented with 10% heat-inactivated fetal bovine serum (FBS; Thermo Fisher A5256701). Cells revived from cryopreserved stocks were expanded for 3–4 passages prior to use.

PBMC collection and T cell isolation

PBMCs were isolated from donor blood samples using Lymphoprep (Stemcell Technologies 07861) according to the manufacturer’s protocol with minor modifications. Briefly, diluted blood was layered over Lymphoprep and centrifuged at 1,000 × g for 30 min at room temperature with acceleration and brake off. PBMCs were then either frozen immediately for storage or subjected to T cells isolation within no more than 2 days.

CD3+ T cells were isolated from PBMCs using the EasySep Human T cell Isolation Kit (Stemcell Technologies 17951) following the manufacturer’s instructions. Lymphocytes were resuspended at 5 × 107 cells mL−1 and incubated with Isolation Cocktail (50 μL mL−1) for 5 min at room temperature, followed by addition of RapidSpheres (50 μL mL−1). The mixture was topped up to 2.5 mL with wash medium (PBS containing 2% FBS and 1 mM EDTA, Ca2+/Mg2+-free). Non-CD3+ cells were magnetically captured, and CD3+ T cells in the supernatant were collected and resuspended in AIM-V medium.

Cell culture transient transfection

For PEI-mediated transfection, cells were cultured in antibiotic-free medium, briefly washed with pre-warmed 1× PBS (Cytiva SH30258.02), and cultured in 80% volume of Opti-MEM (Thermo Fisher 11058021). Plasmid DNA was complexed with PEI (10  kDa; Sigma-Aldrich 765090) at a 1:3 (w/w) ratio in Opti-MEM for 20–25 min at room temperature or 37°C and added dropwise to cells. For Lipofectamine 3000-mediated transfection, plasmids were prepared according to the manufacturer’s protocol (Thermo Fisher L3000015) and incubated with cells for 6 h before replacement with fresh complete medium.

Lentivirus production and transduction

HEK293T cells (passage number 95%. Peptides were resuspended in appropriate solvent following the manufacturer’s certificate of analysis (CoA) at the concentration of 20 mM and stored as aliquots at −80°C. Each peptide batch was subjected to no more than two freeze-thaw cycles.

Peptide pulsing and co-culturing

For peptide pulsing, NCI-H520, MDA-MB-468, and SW620 cells expressing the HLA allele of interest were seeded at a density of 6.25 × 104 cells per cm2 in AIM-V medium supplemented with 0%–2% FBS and the relevant peptides at the indicated concentrations. Cells were incubated for 18–24 h at 37°C under 5% CO2. Pulsed cells were then washed once with phosphate-buffered saline (PBS) and once with AIM-V medium supplemented with 5% FBS, resuspended, and seeded at a density of 2 × 105 cells per ml.

For co-culture experiments, effector cells (PBMCs or T cells) were expanded in AIM-V medium supplemented with 5% FBS and subsequently mixed with target cells at the specified effector:target (E:T) ratios in the presence of relevant T cell-engaging receptors (TCEs). Co-cultures were maintained for 24–48 h prior to downstream analysis.

Flow cytometry and cell-based staining

For a complete list of antibodies used in this study, see Table S9. For cell-surface staining, cells were washed three times in ice-cold 1× PBS, resuspended in ice-cold 1× PBS, and incubated with fluorophore-conjugated antibodies for 30 min at 4°C in the dark. Cells were washed three times in ice-cold 1× PBS before analysis. For intracellular perforin and granzyme B staining, T cells were incubated with the protein transport inhibitor monensin (eBioscience, cat. no. 00–4505-51) for 4–6 h at 37°C under 5% CO2. Cells were then permeabilized using Intracellular Staining Permeabilization Wash Buffer (BioLegend, cat. no. 421002) for 45 min, washed twice, and stained with the relevant antibodies. Flow cytometry data were acquired on an Agilent NovoCyte flow cytometer.

Gating strategy: Singlet cells were identified using forward scatter (FSC) and side scatter (SSC) parameters. After gating on singlets, live cells of interest were defined using specific markers (CD3-PE-Cy7 for T cells and His-APC for SW620 cells). Markers of T cell activation and cytotoxic granule production were evaluated by measuring CD69-APC, perforin-AF700, and granzyme B-PerCP, with comparisons made between mock-treated and experimental samples under the indicated conditions.

Cell-based binding assay

For cell binding assays, target cells were incubated with relevant TCEs for 1 h in defined media and then proceeded with a secondary antibody for assessment via flow cytometry.

Cell-based functional assays

Peptide-pulsed cells were washed once in 1x PBS and once in AIM-V containing 2% FBS. Effector cells were cultured in AIM-V media supplemented with 2% for at least 24 h and checked for viability. Cells were then co-cultured with effector cells in the presence of relevant TCEs. The supernatant was collected for IFN-γ-release ELISA assay (Biolegend 430104). Simultaneously, T cells were collected for T cell activation assays as described earlier, and target cells were subjected to luminescence-based cytotoxicity assay using Celltiter-Glo (Promega G9242). Both functional assays were done following the manufacturer’s instructions.

In silico peptide evaluation and protein complex prediction

Peptide homolog identification was performed using NCBI Blast and Expasy ProScan with the UniProt protein database for Homo sapiens, using the cut-off of 4 mismatches. Homologous and alanine-scanned peptides were then evaluated with p53R273H 9-mer H5 peptide for binding affinity with HLA-C∗01:02 using NetMHC Pan 4.1 or MHC-I prediction. The binding models of TCEs with p53R273H/HLA-C∗01:02 and CD3 were done using a combination of Python-scripted Alphafold 3.0, TCR model 2.3, Boltz-2, and TCR dock. PyMOL was used to evaluate and optimize models’ visualization.

Machine learning-based directed evolution

In brief, a complex between B2M, HLA-C∗01:02, and p53 R273H peptide was generated via Boltz-2, Alphafold 3.0, and TCR model to ensure structure prediction consistency. Parental TCE was then added to the pHLA complex and new complex spatial structure was then generated and cross-evaluated across three above models. This complex was then subjected for in silico directed evolution via EvoPro, EvoProtGrad with maximum 50 mutations per round and 20 cycles of optimization per run, focusing on TCR region on TCE while leaving the anti-CD3, Fc of TCE, and pHLA remained unchanged. Per directed evolution rounds, top 10 TCE candidates with lower free energy (ΔG) and lower free binding energy (ddG) were selected for further evaluation, including binding affinity, stability, and aggregation using independent computational programs. For example, these candidates were tested for binding affinity against HLA and presenting peptide at both 25°C as well as 37°C via PRODIGY. Aggregation score was assessed via evaluating distribution of hydrophobic and hydrophilic residues along TCR sequences using Aggrescan 3D and DeepSP. In addition, candidates were also evaluated for immunogenicity potential using ABImmpred and Hu-mAb. Top candidate per round was chosen among the top performers for each category and served as the starting construct for the next directed evolution round.

Spheroid formation and drug administration

To evaluate the cytotoxic effects of T cell engagers (TCEs) alone and in combination with chemotherapy, 3D colorectal cancer (CRC) spheroids were generated. Oxaliplatin (MedChem Express, HY-17371) and irinotecan (MedChem Express, HY-16562) were employed as chemotherapeutic agents alongside TCEs. Spheroids were created by co-culturing SW620 CRC cells (ATCC, CCL-227), human umbilical endothelial cells (HUVECs) and human adipose-derived stromal cells (ASCs, Cellosaurus CVCL_4W37) at a 3:1:1 ratio, totaling 6,000 cells, in ultra-low attachment 96-well round-bottom plates (Corning Costar, 7007). The cells were suspended in 200 μL of AIM-V medium (Gibco, 12055091) supplemented with 5% FBS (Gibco, 10437028) and either the presence of 2 μM or the absence p53R273H 9-mer peptide H5 to promote TCE-specific antigen presentation on SW620 cells in pulsed or endogenous models. SW620 cells and ASCs were stably transduced with green fluorescent protein (GFP) and red fluorescent protein (RFP), respectively, to enable visualization of tumor-stromal interactions. The plates were centrifuged at 350 × g for 5 min and incubated at 37°C with 5% CO2 for 48 h to form compact spheroids.

Spheroids were co-cultured with 1 × 105 T cells and exposed to single agents (TCE01C, TCE01Cr3, oxaliplatin, or irinotecan) or their combinations. For single TCE treatments, TCE01C and TCE01Cr3 were diluted 4-fold across a 10-point range (0.015–60 nM). For single chemotherapy treatments, oxaliplatin (0.2–200 μM) and irinotecan (0.1–100 μM) were diluted 10-fold. In combination studies, TCEs (0.00015–60 μM) were combined with fixed sublethal doses of oxaliplatin (10 μM) or irinotecan (100 μM). Controls consisted of spheroids alone, spheroids with T cells, T cells alone, and T cells with oxaliplatin or irinotecan. After 48 h, spheroid morphology and GFP fluorescence were evaluated using an EVOS M5000 Imaging System (Thermo Fisher Scientific). GFP fluorescence intensity was quantified with ImageJ software (NIH, v1.53). Supernatants were collected for IFN-γ measurement via ELISA assay, as previously described.^[22]

Statistical analysis

T cell-mediated cytotoxicity was evaluated using the Celltiter-Glo assay, in which the amount of ATP from live cells was used to determine live cell density. The percentage of specific cell lysis was calculated using the following equation ^[1]:

For IFN-γ release ELISA, readouts were normalized to those of co-cultured samples without TCEs. Flow cytometry signals were normalized to T cells only, target cells only, or co-cultured samples without TCEs. Error bars represent the standard error of the mean (SEM). For multiple comparisons to the same control (in the donor’s evaluation experiment), we used one-way ANOVA, followed by Dunnett’s test. For multiple pairwise comparisons (in tumor cells assessment experiment), we applied one-way ANOVA coupled with Tukey’s honestly significant difference (HSD) test. At least three biological samples and three technical replicates were analyzed for each condition.

CRISPR genome editing

Region surrounding R273 residue of TP53 gene was analyzed and cross-compared via Cas-Designer (http://www.rgenome.net/cas-designer/), Geneious Prime, and Benchling CRISPR platform. Potential target sites were evaluated for Cas9 activity score and potential off-targets via Cas-OFFinder (http://www.rgenome.net/cas-offinder/). Top two candidates covering both sides of R273 residue were selected, and corresponding gRNAs were cloned into the multiplex system separated by tRNA. Plasmid carrying Cas9, multiplex gRNA was co-transfected with a single-strand donor DNA repair template into target cells of interest using PEI 10 kDa. DNA from individual clones was extracted using lysis buffer (50 mM KCl, 10 mM Tris-HCl pH 8.3, 2.5 mM MgCl2, 0.5% Tween 20, 1× Proteinase K). Extracted DNA was used for downstream analysis, including genomic PCR and NGS.

Pharmacokinetics study

For studying the PK, six male Swiss albino mice were randomly assigned to each experimental TCE group that represents platform A, B, C, and D, with ten mice in the PBS control group. Mice were injected i.v. through the tail vein with 5 mg kg−1 drug or PBS. Blood samples were collected at 1, 2, 5, 24, 30, 48, and 72 h. Samples were centrifuged, and serum fractions were stored at −20°C for later ELISA. Animal experimentation was approved by the Animal Ethics Committee at the University of Science in Ho Chi Minh City, Vietnam (protocol number 562/KHTN-ACUCUS).

For ELISA, samples (100 μL) diluted in PBS and dilution series of known concentrations were added to immuno 96-well plates and incubated for 2 h at RT. The wells were washed 3 times in PBS +0.05% Tween 20 (PBST), 300 μL per well. TCEs were detected using anti-Fc peroxidase antibody (Thermo Fisher Scientific, A18817) diluted 1:5,000 in 1% BSA in PBST, while TCER03 was detected using anti-DYKDDDDK TAG conjugated HRP (1:1,000, Thermo Fisher Scientific, A01428100). The wells were washed four times with PBS and developed with TMB (eBioscience, 00-4201-56). Each sample was measured in triplicate.

Pharmacokinetic parameters were calculated by noncompartmental analysis (NCA) using Phoenix WinNonlin 8.5.2.4 with model 201 (IV bolus administration). For each animal, the terminal elimination rate constant (λz) was estimated by log-linear regression of the terminal (declining) phase of the serum concentration over time profile, and the terminal half-life t1/2 was calculated as ln(2)/λz. Animals for which a terminal log-linear phase could not be reliably identified (e.g., due to non-declining or non-monotonic concentrations in the terminal window, consistent with assay noise near the lower limit of quantification or an atypical early-phase concentration rise suggestive of a dosing artifact) were excluded from λz and t1/2 estimation for that animal. The number of animals included in each group-level half-life estimate is reported in Table 1.

Data and code availability

Datasets supporting the conclusions of this study are available in the source data file. The relevant materials will be shared upon request.

Acknowledgments

We are grateful to Que-Tran Bui Nguyen, Le Son Tran, Trung Quan Nguyen, Lyna Tran, Phuc M. Phung, and Lan Tu for kindly providing materials used in this study.

Patient and healthy-donor blood samples were collected under the ethical approval described in the Ethics Statement above (Medical Genetics Institute, approval number 02/2024/CT-VDTYH). HUVECs and adipose-derived stem cells (ASCs) used in this study were obtained from ATCC and did not require additional institutional ethical approval.

This study was funded by 10.13039/100032442Gene Solutions.

Author contributions

N.H., H.-N.N., and C.T.N. conceptualized the study. N.H., T.-M.T.N., N.T.B., D.T.P.T., and C.T.N. developed the methodology and designed the experiments. N.H., T.-M.T.N., N.T.B., T.-N.H.N., and C.T.N. performed the experiments. N.H., T.-M.T.N., N.T.B., and T.-N.H.N. conducted the visualization. N.H., C.T.N., and H.-N.N. project managed the administration, while N.H., C.T.N., and H.-N.N. provided the supervision. N.H. wrote the original draft, and N.H., T.-M.T.N., N.T.B., D.T.P.T., T.-N.H.N., H.N.N., and C.T.N. reviewed and edited the manuscript.

Declaration of interests

D.T.P.T. is a co-inventor on patent application 12442046, which incorporates the discovery of TCR12.1 targeting p53 R273H neoantigen N.H., T.-M.T.N., N.T.B., H.-N.N., C.T.N. are co-inventors on patent application 19059303, which incorporates the ML-derived TCRs in this study and the generation of p53 R273H cell lines. All authors are employees of Gene Solutions Inc.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of this manuscript, the authors used ChatGPT (OpenAI) to assist with improving the language, grammar, readability, and clarity of the manuscript. The authors reviewed, edited, and verified all content generated with this tool and take full responsibility for the accuracy, originality, and integrity of the final manuscript.