PRECISE-seq reveals disease-relevant TCR repertoires with phenotypic plasticity

T cells are central to immune surveillance against infections and cancer (Cox et al., 2019; Waldman et al., 2020). T cell–based immunotherapies have achieved remarkable clinical success, relying on specific interactions between T cell receptors (TCRs) and peptide–major histocompatibility complexes (pMHCs). Alongside TCR specificity, another important consideration for clinical development is TCR potency (Chan et al., 2023; Klebanoff et al., 2023; Zhao et al., 2022), which refers to the ability of TCRs to initiate T cell responses to naturally processed peptides displayed on target cells. TCR potency is typically measured through functional avidity assays that preserve physiological parameters, including TCR-pMHC density, affinity, and coreceptor engagement. Importantly, TCR potency shapes T cell fate and functional phenotypes in vivo (Huang et al., 2010; Liu et al., 2014). Linking TCR specificity and potency with phenotypic states is therefore essential for understanding T cell responses during immunotherapies and optimizing TCR-based therapies (Oliveira and Wu, 2023).

Current approaches for identifying disease-associated TCRs rely on either pMHC multimer staining (Ma et al., 2021; Newell et al., 2013) or in vitro functional screening of TCR clonotypes (Arnaud et al., 2022; Chandran et al., 2022; Dijkstra et al., 2018; Lowery et al., 2022; Rosenberg and Restifo, 2015). While pMHC multimers allow simultaneous assessment of antigen specificity and phenotype, their reliance on predefined antigens limits multiplex capacity, and binding affinity alone often poorly predicts functional outcomes (Sibener et al., 2018). Functional assays based on coculture of T cells and antigen-presenting cells (APCs) circumvent predefined epitope requirements but require prolonged in vitro culture, leading to loss of native phenotypic information (Dijkstra et al., 2018; Lowery et al., 2022). Thus, there remains a need for high-throughput platforms capable of simultaneously decoding antigen specificity, quantifying functional avidity under physiological conditions, and preserving endogenous T cell phenotypes. Contact-dependent proximity labeling enables mapping of live T cell–target interactions within confined spatial contexts, offering a potential solution (Liu et al., 2020; Pasqual et al., 2018). However, current strategies suffer from variable labeling efficiency due to heterogeneous substrate expression on prey cells (Liu et al., 2020; Nakandakari-Higa et al., 2024; Pasqual et al., 2018), and nonspecific background beyond the immunological synapse (Oakley et al., 2022; Zhou and Zou, 2021), limiting sensitivity for detecting differences in TCR potency.

Here, we present PRECISE-seq, a contact-dependent proximity labeling sequencing platform that integrates Sortase A (SrtA)–catalyzed ligation to simultaneously profile antigen specificity, TCR potency, and phenotypic landscape. In PRECISE-seq, T cells are uniformly modified with acceptor peptides (APs) via chemical conjugation to ensure consistent substrate availability. Spatially restricted SrtA-catalyzed ligation confines signal deposition to the immunological synapse, minimizing off-target background. PRECISE-seq robustly identifies antigen-specific T cells, with labeling intensity quantitatively reflecting TCR potency. Application to human cytomegalovirus (CMV)-specific T cells revealed that high-potency clonotypes preferentially exhibit an exhaustion phenotype. In tumors, PRECISE-seq identified a Ly49⁺ CD8⁺ T cell (T_Ly49) state characterized by inhibitory killer cell lectin-like receptor expression and immunoregulatory activity in suppressing host antitumor response. PD-1 blockade rejuvenates tumor-specific T cells, driving a transition from the T_Ly49 state to an effector-like state that is associated with improved immunotherapy responsiveness in colorectal cancer (CRC), hepatocellular carcinoma (HCC), and melanoma. Thus, PRECISE-seq enables rapid characterization of in vivo T cell responsiveness across diseases, facilitating the advancement of T cell–based therapies against infections and cancers.

PRECISE-seq comprises two primary processes: the proximity-based labeling of T cells during antigen recognition, followed by high-throughput single-cell sequencing of the labeled T cells. For the labeling step, PRECISE-seq employs SrtA-mediated transpeptidation (sortagging), wherein engineered SrtA on the cell membrane of bait cells covalently transfers the Biotin-LPETGG probe to proximal APs on neighboring prey cells (Fig. 1 A and Fig. S1, A and B). The sensitivity and unbiased nature of PRECISE-seq are highly dependent on the level of the AP present on the surfaces of neighboring prey cells. To establish a precise and unbiased contact-dependent labeling method, we developed a two-step chemical reaction to introduce a homogeneous level of the AP (G5 tag) on the surfaces of prey cells. The prey cells were modified with the maleimide group through conjugation of primary amine with NHS-PEG4-maleimide, followed by cysteine–maleimide conjugation to attach the AP-HA tag (peptide sequence: GGGGGYPYDVPDYASSC) (Fig. S1 C). We then assessed the conjugation efficiency of the two-step chemical reaction. Flow cytometry data demonstrated that the two-step chemical reaction successfully yielded uniform and abundant AP-HA tag display on the cell surface, enabling efficient recognition and labeling by SrtA (Fig. 1 B and Fig. S1 D).

To assess the capability of PRECISE-seq in deciphering physical intercellular interactions, we established a cellular contact system based on CD40 and CD40L recognition within HEK293T cells (Fig. 1, C and D). We co-expressed SrtA and CD40L on the surface of HEK293T cells, which served as bait cells (SrtA⁺CD40L⁺ HEK293T), and chemically attached prey cells (CD40-GFP⁺ HEK293T) with the AP-HA tag. The two populations were mixed in the presence of the biotinylated SrtA substrate (Biotin-LPETGG) for 30 min, followed by flow cytometry analysis to detect Biotin signals (Fig. S1 E). We observed efficient labeling of prey cells (CD40-GFP⁺ HEK293T) by PRECISE-seq when cocultured with SrtA⁺CD40L⁺ HEK293T cells, as evidenced by a robust Biotin signal (Fig. 1, E–G; and Fig. S1, F–H). In contrast, minimal Biotin signals were detected on prey cells cocultured with SrtA⁺ HEK293T cells lacking CD40L expression. Overall, these results indicate that PRECISE-seq effectively records intercellular interactions mediated by specific ligand–receptor pairs.

Given the relatively short labeling radius of PRECISE-seq (Lin et al., 2023, Preprint; Oakley et al., 2022; Zhang et al., 2023; Zhou and Zou, 2021), we inferred that this method can accurately record the specific recognition between a live T cell and its target cell (Fig. S1 I). For this purpose, we employed the specific recognition between 1G4-TCR and NY-ESO-1_157–165 presented by HLA-A2 (Jäger et al., 1998; Li et al., 2005) (Fig. 2 A). We cloned 1G4-TCR into a TCR-deficient Jurkat T cell line (1G4-JC5 cells). Additionally, we constructed an artificial APC (aAPC), which was established on the K562 cell line with stable expression of HLA-A2 and SrtA. Subsequently, we cocultured AP-HA tag⁺ 1G4-JC5 cells with SrtA⁺ aAPCs presenting NY-ESO-1_157–165, in the presence of Biotin-LPETGG substrate for 2 h (Fig. 2 B). We observed functional activation of 1G4-JC5 cells, evidenced by the upregulation of the TCR downstream signaling marker CD69. Notably, nearly all CD69⁺ 1G4-JC5 cells were labeled with Biotin (Fig. 2, C and D). Conversely, when incubated with either irrelevant CMV pp65_495–503-pulsed or mock-pulsed SrtA⁺ aAPC, few 1G4-JC5 cells displayed Biotin signals. In addition, PRECISE efficiently labeled K562-HLA-A2 cells that endogenously present NY-ESO-1 peptides when cocultured with SrtA⁺ 1G4-JC5 cells, but not with SrtA⁺ JC5 cells. This confirms the method’s specificity and efficacy in labeling APCs that present endogenous peptides (Fig. S1 J). These findings underscore the capacity of PRECISE-seq to accurately decode antigen specificity, with the Biotin signals correlating with functional activation.

We then extended the application of PRECISE-seq to determine the antigen specificity of primary CD8⁺ T cells. We evaluated the efficiency of AP-HA tag conjugation using naïve T cells (T_N) and effector T cells (T_EFF), two representative T cell subsets with distinct transcriptional profiles. Flow cytometry data indicated that the abundance of the AP-HA tag on T_EFF was comparable to that on T_N (Fig. S1 K), suggesting that PRECISE-seq mitigates potential labeling discrepancies arising from diverse transcriptional activities across cellular states. Additionally, the cellular viability and antigen sensitivity of T cells remained unaffected after the conjugation of the AP-HA tag (Fig. S1, L and M).

Subsequently, we cocultured SrtA-expressing B16-OVA tumor cells with primary OT-I T cells conjugated with the AP-HA tag, whose TCRs specifically recognize the OVA antigen. Flow cytometry analysis revealed that the majority of CD69⁺ OT-I T cells activated by cognate OVA antigen were simultaneously labeled with Biotin probes by PRECISE-seq (Fig. 2, E and F). Conversely, minimal Biotin signals were detected on OT-I T cells when cocultured with B16F10-SrtA cells, confirming the ability of PRECISE-seq to capture primary antigen-specific T cells. Of note, the abundance of AP-HA tags on T cells was critical for labeling efficiency, as sub-optimal conjugation significantly reduced downstream labeling (Fig. S1, N and O). To assess the sensitivity of PRECISE-seq, we mixed target OT-I T cells with antigen-irrelevant polyclonal CD8⁺ T cells from wild-type mice in various ratios, ranging from 1:10 to 1:10,000 (OT-I: polyclonal CD8⁺ T cells). Analysis of Biotin signals on OT-I T cells demonstrated consistent performance, even when the proportion of antigen-specific T cells was as low as 0.01% (Fig. 2 G and Fig. S1 P), indicating the sensitivity of PRECISE-seq in detecting low-frequency antigen-specific T cells.

The ex vivo assessment of TCR potency within the cellular context provides superior metrics for evaluating T cell responsiveness. However, its application is limited by relatively low throughput and a tedious screening and validation process. To assess the capacity of PRECISE-seq in quantifying TCR potency, we utilized an established model employing altered peptide ligands (APLs) to modulate the strength of APC–T cell interactions across a range of antigenic potencies (Daniels et al., 2006). The original ligand OVA_257–264 (SIINFEKL, N4) is most potent in stimulating OT-I cells, while the remaining five APLs (Q4, T4, Q4H7, G4, and E1), derived from N4, were listed in a descending order based on their potency in stimulating OT-I cells (Fig. S1 Q). A gradual decrease in labeling intensity of OT-I T cells corresponding to the recognition of APLs with potencies ranging from high to low was observed (Fig. 2 H). We further examined whether PRECISE-seq could resolve dose-dependent differences by exposing T cells to increasing doses of APLs. The Biotin signals detected by PRECISE-seq faithfully reflected the binding strength of TCR-pMHC interactions and demonstrated consistency with T cell activation, as measured by CD69 upregulation (Fig. 2 I and Fig. S1 R). These data indicate that PRECISE-seq can accurately quantify TCR potency guided by the strength of TCR-pMHC interactions.

We then utilized PRECISE-seq to systematically analyze the in vivo T cell response against human CMV, a common herpesvirus that periodically reactivates and thus repeatedly stimulates the antigen-specific T cells (Griffiths and Reeves, 2021) (Fig. 3 A). For this purpose, we isolated peripheral blood mononuclear cells (PBMCs) from a CMV-seropositive donor carrying HLA-A2 and labeled them with the AP-HA tag (Fig. S2 A). To identify CMV-specific CD8⁺ T cells, we incubated these AP-HA tag⁺ PBMCs with SrtA⁺ HLA-A2⁺ K562 cells presenting CMV pp65_495–503 (Wills et al., 1996). We compared the TCR repertoire before and after coculture and found no significant alteration, thus ruling out potential artifacts from the coculture process (Fig. S2 B). Flow cytometry analysis revealed ∼1% of CD8⁺ T cells labeled with Biotin signals (Fig. 3 B). In addition, we performed parallel tetramer staining for direct comparison of the labeling efficiency using PBMCs from two CMV-seropositive donors carrying HLA-A*24:02. Flow cytometry analysis showed strong concordance between Biotin labeling and tetramer staining, confirming the efficacy of the PRECISE-seq approach (Fig. S2 C). The Biotin⁺ CD8⁺ T cells from HLA-A2⁺ donors were isolated and subsequently subjected to 5′ single-cell RNA sequencing (scRNA-seq), coupled with TCR profiling (Fig. S2 D). Additionally, nonconjugated CD8⁺ T cells were also isolated for single-cell sequencing to characterize the baseline transcriptional profile and TCR repertoire of polyclonal T cells in fresh PBMCs (hereafter referred to as the baseline group).

From the PRECISE-seq data, we recovered 1,450 clonotypes for the baseline group and 345 clonotypes for the Biotin⁺ group, respectively. Notably, the top clones in the Biotin⁺ group exhibited larger clone size compared with those in the baseline group, suggesting a selective enrichment of antigen-experienced, highly expanded TCR clonotypes (Fig. S2, E and F). To identify high-confidence CMV-specific clonotypes, we applied stringent criteria: (1) a minimum cell number threshold of 2 in the Biotin⁺ group; (2) significant enrichment in the Biotin⁺ group compared with the baseline group, with a false discovery rate (FDR) of <0.05 by Fisher’s exact test; and (3) an odds ratio (OR) >4. This strategy yielded six high-confidence clonotypes in the Biotin⁺ group, namely, clones 4, 13, 35, 40, 52, and 67. Notably, three of these clonotypes (clones 35, 52, and 67) were undetectable at baseline, highlighting the sensitivity of PRECISE-seq in detecting low-frequency antigen-specific clonotypes that might be overlooked by conventional single-cell TCR-sequencing (scTCR-seq) methods (Fig. 3 C).

It is important to note that when analyzing samples with an extensive TCR repertoire such as PBMCs, the number of cells captured by single-cell sequencing is limited by the maximum throughput capacity (typically up to 10⁴ cells in the 10X droplet-based platform), potentially resulting in a representation that does not fully reflect the starting population. Consequently, antigen-specific clonotypes with a low frequency (<0.01%) cannot be accurately quantified. To achieve more precise quantification of baseline frequency, we conducted bulk TCRα/β chain repertoire sequencing on baseline CD8⁺ T cells (bulk TCR-seq). The baseline frequencies of high-confidence clonotypes detected in bulk TCR-seq and PRECISE-seq were highly correlated (Fig. S2 G). However, bulk TCR-seq enabled the detection of clones 52 and 67 with baseline frequencies as low as 0.046% and 0.0012%, respectively. Notably, these two clonotypes were identified as high-confidence clonotypes by PRECISE-seq, demonstrating the ability of our method to identify extremely low-frequency antigen-specific T cells (Fig. 3 D).

Next, we evaluated the accuracy of PRECISE-seq using a TCR-pMHC activation screening assay. We selected 245 clonotypes to generate an adeno-associated virus (AAV) vector-based TCR library, which contained: (1) six high-confidence clonotypes and 221 low-confidence clonotypes identified by PRECISE-seq in the Biotin⁺ group; (2) 18 clonotypes only detected in the baseline group. In addition, three reported CMV pp65_495–503-specific clonotypes as positive controls (Shugay et al., 2018). Subsequently, we transduced this TCRαβ expression library into TCR-deficient human primary T cells. We performed the activation screening assay in the presence of CMV pp65_495–503 peptide and then sorted activated cells (4-1BB⁺) to identify CMV-specific clonotypes (Gros et al., 2016) (Fig. 3 E). The CMV-specific TCR clonotypes were defined as those enriched in the CMV pp65_495–503-stimulated group compared with control groups (Materials and methods). While three reported TCRs were enriched in the activation screening assay, four of six high-confidence clonotypes (clones 4, 40, 52, and 67) identified by PRECISE-seq were also enriched in this screening assay (Fig. 3 F and Fig. S2 H). Only one out of 221 low-confidence clonotypes (clone 60) was enriched. None from the baseline group were enriched in the screening assay. Additionally, three out of four true-positive CMV-specific clones were reproducibly identified in an independent experimental batch of PRECISE-seq (Fig. S2, I–K). Moreover, applying PRECISE-seq to two additional CMV-seropositive donors successfully retrieved their respective CMV pp65_495–503-specific clonotypes, as validated by IFN-γ ELISpot assay (Fig. S2, L–O). Overall, these results indicate that PRECISE-seq effectively and accurately retrieves antigen-specific TCRs from a large pool of TCR repertoire in circulating blood upon CMV infection.

To quantify the potency of TCR-pMHC interactions for individual clonotypes, we assigned a potency score to each clonotype by calculating the fold change in frequency between the Biotin⁺ group and the baseline group (determined by bulk TCR-seq) and then adjusting it based on the distribution of irrelevant TCRs (Materials and methods). Among the four CMV-specific clonotypes, clone 4 was classified as high-potency, and clones 40 and 52 as intermediate-potency, and clone 67 was excluded from the analysis due to its extremely low baseline frequency (0.0012%), as such a small denominator makes the fold-change estimate highly susceptible to stochastic noise (Fig. 3 D). We then evaluated whether the potency scores obtained through PRECISE-seq correlate with the in vitro TCR avidity assessed by flow cytometry. To this end, JC5 cells were transduced with CMV-specific TCRs and stimulated with serially diluted CMV pp65_495–503 peptide presented by K562-HLA-A2 cells. TCR avidity was quantified by flow cytometry for CD69 staining, measuring the peptide concentration required for a half-maximal response. All TCR clones demonstrated peptide dose-dependent CD69 upregulation, and the measured avidity (EC₅₀) was consistent with the potency scores assessed by PRECISE-seq, maintaining a similar ranking among the clones (Fig. 3 G). We noticed that the functional avidity of clone 67 was 220-fold lower than clone 4, whose potency was beyond the detection scope of PRECISE-seq. These results validate that the TCR potency scores derived from PRECISE-seq serve as a reliable proxy for functional avidity.

We utilized PRECISE-seq to explore the phenotypic landscape of CMV-specific T cells through analysis of transcriptome data. Consistent with previous encounters with the antigen, CMV-specific clonotypes were notably devoid of naïve cells, which constitute a major population in the baseline group (Fig. 4 A and Fig. S2 P). Interestingly, CMV-specific T cells were predominantly found within two effector memory T cell (T_EM) subsets, characterized by high levels of activation and cytotoxicity signatures (GZMB⁺ T_EM and GZMK⁺ T_EM) (Fig. 4, B–D; and Fig. S2, Q and R). We verified that the effector memory phenotype of CMV-specific clonotypes was not induced by the labeling process of PRECISE-seq, as these clonotypes exhibited a similar phenotypic distribution in the baseline group (Fig. S2 S). Considering that TCR potency determines the in vivo functional phenotype of antigen-specific T cells (Oliveira et al., 2021; Schmidt et al., 2023), we examined the correlation between TCR potencies and T cell phenotypes. We observed that high-potency T cells exhibited increased expression of an activation signature, suggesting a greater magnitude of T cell activation resulting from enhanced APC-T intercellular interactions (Fig. 4 E). Notably, high-potency T cells also upregulated the expression of an exhaustion signature, consistent with the prior notion that iterative antigen stimulation drives T cell exhaustion program (Klenerman and Oxenius, 2016) (Fig. 4 F). Overall, PRECISE-seq effectively delineates the antigen specificity, TCR potency, and phenotypic landscape of CMV-specific clonotypes, enabling comprehensive characterization of T cell responses to human CMV infection.

We applied PRECISE-seq to monitor the dynamics of in vivo antitumor T cell response occurring within the TME upon immunotherapy. Tumor-specific T cells experiencing chronic antigen exposure undergo TCR internalization, which reduces the availability of TCRs for binding with pMHC and limits the detection of tumor-specific T cells within the TME (Stinchcombe et al., 2023). Initially, we confirmed the ability of PRECISE-seq to capture tumor-specific T cells within tumor tissues and tumor-draining lymph nodes (tuDLNs) by transferring OT-I T cells into mice bearing B16-OVA tumors (Fig. S3 A). OT-I T cells from tumor or tuDLNs were explicitly labeled with Biotin signals by B16-OVA-SrtA, while negligible Biotin signals were detected after coculturing with B16F10-SrtA (Fig. S3, B and C). Moreover, the irrelevant P14 T cells, which are reactive to the LCMVgp_33–41 epitope and were spiked into tumor-infiltrating cells as an internal negative control, were also Biotin-negative (Fig. S3 D), reinforcing the high specificity of PRECISE-seq in detecting T cells against tumor antigens. In addition, the Biotin labeling intensity was comparable among different T cell subsets, despite their varying expression of adhesion molecules and costimulation molecules (Fig. S3, E–H), indicating that PRECISE-seq is not confounded by differences in T cell phenotypes.

Understanding the underlying mechanisms of productive antitumor T cell responses upon immunotherapy requires accurate characterization of alterations in both T cell phenotype and TCR repertoire. We hypothesized that PRECISE-seq could elucidate the in vivo antitumor T cell responsiveness by linking dynamic T cell landscapes and TCR specificities against the available antigens presented by tumor cells. To investigate this, we implanted mouse colon tumor cells MC38 into C57BL/6 mice and treated tumor-bearing mice with PD-1 blockade antibody (αPD-1 Ab) (Fig. 5 A). Tumor antigens were presented by an autologous MC38 tumor cell line (Yadav et al., 2014), with SrtA overexpressed on the surface of MC38 tumor cells. 7 days after treatment, we cocultured AP-conjugated tumor-infiltrating cells with either MC38-SrtA cells or MHCI-deficient MC38-SrtA cells (MC38 B2MKO-SrtA cells) and then sorted the Biotin⁺ CD8⁺ T cells to perform PRECISE-seq. Additionally, we isolated total tumor-infiltrating CD8⁺ T cells without AP-HA tag conjugation to establish a reference for the TCR repertoires within tumors (baseline).

We observed a notable increase in the frequency of Biotin-labeled CD8⁺ T cells in mice treated with αPD-1 Ab compared with those in untreated tumors (Fig. S3 I), correlating with a significant suppression of tumor growth under αPD-1 Ab treatment (Fig. S3, J and K). Flow cytometry analysis revealed that Biotin⁺ tumor-infiltrating CD8⁺ T cells exhibited heightened levels of effector cytokine production compared with their Biotin⁻ counterparts (Fig. S3 L). Additionally, PD-1 blockade further augmented the production of effector cytokines in Biotin⁺ CD8⁺ T cells, underscoring the role of PD-1 blockade in enhancing tumor-specific T cell responses (Fig. S3 M). Furthermore, the transfer of Biotin⁺ CD8⁺ T cells into MC38 tumor-bearing mice resulted in inhibited tumor growth, whereas tumors progressed in mice receiving Biotin⁻ CD8⁺ T cells (Fig. S3, N and O), confirming the functional responsiveness of Biotin-labeled CD8⁺ T cells to tumor antigens.

When analyzing the PRECISE-seq data, we identified 1,220 CD8⁺ T cell clonotypes in the baseline group and 719 clonotypes in the Biotin⁺ group, among which 21 were designated as high-confidence tumor-specific clonotypes (Fig. 5 B and Fig. S3 P). These tumor-specific clonotypes displayed higher baseline frequencies compared with other clonotypes, indicating a continuous expansion of tumor-specific T cells in response to tumor antigens within the TME (Fig. 5 C). Subsequently, we computed potency scores for the high-confidence tumor-specific clonotypes. Compared with CMV-specific clonotypes, tumor-specific clonotypes with available potency scores exhibited significantly lower potency (P = 0.006, two-sided Wilcoxon rank-sum test). This result supports the concept that natural T cell repertoires targeting tumor antigens tend to express TCRs with relatively low potency (Klein et al., 2014; Nikolich-Zugich et al., 2004). Following αPD-1 treatment, a modest increase in TCR potency was observed, suggesting that the increase in TCR potency may contribute to the antitumor effects mediated by αPD-1 therapy (Fig. 5 D).

To validate the accuracy of PRECISE-seq in retrieving tumor-specific clonotypes, we selected eight dominant TCRs based on their P value rankings from both the untreated and αPD-1 groups and examined their tumor reactivity (Fig. S3 Q). We transduced tumor-irrelevant P14 T cells with retroviral constructs encoding the TCR chains of these clonotypes and transferred these TCR-transduced T cells into MC38 tumor-bearing mice, using GFP-transduced P14 T cells as a tumor-irrelevant negative control. We found that the cell proportions of all eight clonotypes were significantly higher than those of the control P14 T cells in tumors 7 days after transfer (Fig. 5 E), suggesting that these TCR-transduced T cells underwent robust proliferation upon recognizing antigens on tumor cells. To confirm tumor reactivity, we re-stimulated tumor-infiltrating lymphocytes (TILs) ex vivo with MC38 cells and evaluated the reactivity by the upregulation of the TCR downstream signaling markers CD69 and CD25. Seven out of eight clones exhibited significant induction of CD69 and CD25 upon re-exposure to MC38 cells, confirming their tumor reactivity (Fig. 5 F and Fig. S3 R). To determine whether TCR potency assessed by PRECISE-seq correlates with the ex vivo functional avidity of tumor-specific TCR clonotypes, three clonotypes were selected for downstream analysis: clone 100 (high TCR potency) and clones 27 and 30 (low TCR potency), as measured by PRECISE-seq (Fig. S3 S). We compared the functional response of these three clonotypes by evaluating their capacity to produce IFN-γ. Flow cytometry analysis revealed that clone 100 exhibited significantly higher IFN-γ production than clones 27 and 30 (Fig. 5 G). This consistency with the TCR potency scores from PRECISE-seq confirms its effectiveness as a surrogate for assessing the functionality of tumor-specific T cells.

To elucidate the phenotypic dynamics of polyclonal tumor-specific T cells following immune intervention, we conducted an analysis of the transcriptome data, partitioning all CD8⁺ T cells into seven subsets: T_N or central memory T cells (T_CM), T_EM, T_EFF, tissue-resident memory cells (T_RM), precursor of exhausted T cells (T_PEX), exhausted T cells (T_EX), and T_Ly49 (Fig. 6, A and B). The comparable abundance of the AP-HA tag across different T cell subsets rules out any potential bias introduced by the conjugation process (Fig. S4, A and B). In untreated tumors, Biotin⁺ tumor-specific clonotypes mainly resided within the T_EX and T_Ly49 compartments (Fig. 6 C). We noticed that these T_Ly49 cells specifically expressed natural killer (NK)-related molecules such as FcεRIg, 2B4 (encoded by Cd244a), and the Ly49 family of inhibitory killer cell lectin-like receptors (Klra3 and Klra5), resembling a previously reported Ly49⁺ regulatory CD8⁺ T cell (Kim et al., 2015; Li et al., 2022) (Fig. 6 D). In addition, T_Ly49 from tumors shared many similarities with Ly49⁺ CD8⁺ regulatory T cells (with KIR⁺ CD8⁺ T cells as their human counterparts) identified in experimental autoimmune encephalomyelitis mice, as well as in human autoimmune and infectious diseases: a marked upregulation of cytotoxic molecules (Gzmc and Prf1), a loss of costimulatory molecules (Cd28 and Cd27), and the upregulation of Ikzf2, a transcriptional factor associated with regulatory functions in both CD4⁺ and CD8⁺ T cells (Kim et al., 2015). Flow cytometry analysis confirmed that T_Ly49 CD8⁺ T cells expressed high levels of FcεR1g, granzyme C, Helios, and immunosuppressive molecule SPP1 (OPN), but rarely produced effector cytokine IFN-γ (Fig. S4, C–E). Notably, we did not identify any cells co-expressing Trav11 (Vα14) and Traj18 in our system, thereby excluding the inclusion of invariant NKT cells within the T_Ly49 subset (Fig. S4 F).

Analysis of the TCR clonotype distributions across all CD8⁺ T cell subsets at baseline revealed that T_EFF, T_EX, and T_Ly49 cells showed the highest degrees of clonal expansion compared with naïve and memory subsets (Fig. S4 G). Notably, T_Ly49 displayed a diverse TCR repertoire, composed of multiple expanded clonotypes rather than domination by a few hyperexpanded clones (Fig. S4 G). Quantitatively, the top 10 T_Ly49 clonotypes accounted for ∼30% of the total T_Ly49 population, indicating moderate clonal expansion but not oligoclonal dominance (Fig. S4 G). While MHC class Ib–restricted CD8⁺ regulatory T cells (Tregs) exhibit a near-monoclonal TCR repertoire dominated by TRAV9N3 (Vα3.2) and TRBV12-1/12-2 (Vβ5.1/5.2) (Kim et al., 2024), ∼30% of tumor-specific T_Ly49 cells in our MC38 tumors expressed TRAV9N3—a frequency higher than in other intratumoral subsets but far below that of MHC Ib–restricted Tregs (Fig. S4 H). These data suggest that tumor-specific T_Ly49 cells likely recognize a broader and more diverse set of tumor antigens rather than being restricted to a narrow set of MHC class Ib–presented self-peptides.

Emerging evidence suggests that T_Ly49 cells suppress T cell response and play a key role in maintaining self-tolerance (Li et al., 2022). To assess the suppressive function of tumor-specific T_Ly49 cells, we cocultured tumor-specific T_Ly49 with responder T_EFF in vitro. Compared with nonsuppressive T_EM controls, T_Ly49 significantly increased cell death of T_EFF (Fig. S4, I and J) and suppressed proliferation of CTV-labeled responder T cells (Fig. S4, K–M). To examine its function in vivo, we isolated tumor-specific T_Ly49 cells (Ly49E/F⁺ 2B4⁺ PD1⁺ P14 cells) from tumor tissues and adoptively transferred them into MC38-gp33 tumor-bearing mice. Intratumoral T_EM/T_EFF cells were also isolated and transferred as a control. While T_EM/T_EFF cells inhibited tumor growth, tumor-specific T_Ly49 cells opposingly promoted tumor growth, indicating the regulatory function of T_Ly49 cells in suppressing in vivo antitumor responses (Fig. S4 N). In addition, intratumor T_Ly49 cells exhibited a gain of exhaustion signatures compared with T_EM and T_EFF subsets, implying they might be targeted by immune checkpoint blockade (Fig. S4 O).

We next examined the phenotypic dynamics of tumor-specific T cells in response to αPD-1 treatment. We observed an increase in the T_EX subset, potentially due to the conversion of T_PEX into a terminally exhausted phenotype upon αPD-1 treatment (Fig. 6 E and Fig. S5 A), consistent with the essential role of T_PEX in this process (Miller et al., 2019). We noticed that following αPD-1 Ab treatment, the frequency of tumor-specific T_Ly49 was markedly reduced, and most tumor-specific T cells acquired effector-like phenotypes, segregating into T_EFF and T_EM compartments. Such enhanced formation of effector subsets was consistent with the increased production of effector cytokines in Biotin⁺ CD8⁺ T cells after αPD-1 Ab treatment (Fig. S5 B and Fig. S3 M). This indicates that PD-1 blockade can direct tumor-specific T cells into an effector state. Given the dramatic reduction of T_Ly49 following PD-1 blockade, we next investigated the function and origin of this specific subset. The URD algorithm revealed strong directional flows originating from less-differentiated T_CM/T_EM populations, which bifurcate into either T_EX or T_Ly49 cells (Fig. S5 C). Specifically, the T_Ly49 trajectory diverged early from T_EM and followed a differentiation pathway distinct from T_EX (Fig. S5, D and E). Further experimental validation supports the differentiation analysis, demonstrating that T_EM cells serve as progenitors of T_Ly49 cells, independent of the T_EX differentiation trajectory (Fig. S5, F and G).

We next validated the effect of αPD-1 Ab on T cell fate commitment by transferring P14 cells into MC38-gp33 tumor-bearing mice followed by αPD-1 Ab treatment (Fig. 6 F). Consistent with the established notion that T_PEX expansion is a key mode of action for αPD-1 efficacy, αPD-1 Ab treatment indeed increased the absolute number of T_PEX cells within tumors (Fig. S5 H). Noticeably, αPD-1 Ab treatment led to a significant reduction in the frequency and absolute cell number of regulatory T_Ly49 cells (Fig. 6, G and H). Conversely, we observed a concomitant rise in the frequency and absolute number of CXCR3⁺ PD1⁻ T_EM cells after αPD-1 treatment (Fig. 6, I and J), indicating a shift toward antitumor effector branch. Similar results were also observed in experiments with mice bearing LLC-gp33 tumors (Fig. S5, I–L). Together, these findings demonstrate that PD-1 blockade not only expands T_PEX cells but also disrupts the balance between regulatory T_Ly49 cells and antitumor T_EM cells.

We hypothesized that an elevated ratio of T_EM/T_EFF to T_Ly49 cells could indicate an effective antitumor response following immune intervention. To explore this further, we calculated the T_EM/T_EFF-to-T_Ly49 cell ratio for individual clonotypes and categorized them into two groups: those displaying a higher T_EM/T_EFF percentage than T_Ly49, and those with a lower ratio. Treatment with αPD-1 Ab favored a T_EM/T_EFF phenotype, with a T_EM/T_EFF-to-T_Ly49 cell ratio >1 in 9 out of 10 tumor-specific clonotypes (Fig. 7 A). In summary, PRECISE-seq facilitated the monitoring of tumor-specific clonotypes within the TME, revealing a phenotypic shift from T_Ly49 to T_EM/T_EFF state in tumor-specific clonotypes following immune intervention.

To determine whether these insights about the phenotypic shift in response to PD-1 blockade can translate to the effectiveness of immunotherapies in human cancers, we collected publicly available scRNA-seq datasets from patients undergoing PD-1 blockade and assigned a cellular state for each CD8⁺ T cell based on their similarity to the CD8⁺ T cell subsets identified in PRECISE-seq data (Materials and methods). We analyzed four clinical cohorts across various cancers: (1) mismatch repair–deficient CRC (MSI CRC) (Li et al., 2023); (2) locally advanced or metastatic CRC (Chen et al., 2024); (3) patients with unresectable HBV⁺ HCC (Guo et al., 2025); and (4) metastatic melanoma (Sade-Feldman et al., 2018). In the MSI CRC cohort, immunotherapy resulted in a decrease in T_Ly49-like CD8⁺ T cells (T_Ly49/KIR) and an increase in T_EM/T_EFF cells among responders (Fig. 7 B). The T_EM/T_EFF-to-T_Ly49/KIR ratio significantly rose in these samples after treatment (Fig. 7 C). Conversely, in nonresponders, the T_EM/T_EFF-to-T_Ly49/KIR ratio remained notably lower than in responders after treatment, suggesting that a preference for the T_Ly49/KIR state was associated with resistance to immunotherapy. Such an association with response to immune checkpoint blockade was also observed in an independent CRC cohort and an HCC cohort, further supporting that the phenotypic shift from T_Ly49/KIR to T_EM/T_EFF state may play a critical role in mediating the therapeutic effects of the PD-1 blockade (Fig. S5, M–P).

Within the melanoma cohort, the T_EM/T_EFF-to-T_Ly49/KIR ratio increased following αPD-1 treatment for both responders and nonresponders (Fig. 7, D and E). However, this ratio was significantly lower in nonresponders compared with responders following treatment. Notably, even prior to treatment, responders exhibited a higher T_EM/T_EFF-to-T_Ly49/KIR ratio than nonresponders, indicating a preexisting effective antitumor response. We discovered that patients with a ratio >1 had longer survival than those with a ratio <1 (log-rank, P = 0.0056, Fig. 7 F), suggesting that the T_EM/T_EFF-to-T_Ly49/KIR ratio could serve as a predictive pretreatment marker for clinical benefits from αPD-1 treatment in melanoma patients. In summary, in line with previous observations, our findings revealed that effective antitumor responses necessitate a phenotypic shift from a regulatory T_Ly49/KIR state toward an effector state.

Recent advances in T cell–based immunotherapy have transformed cancer treatment, yet heterogeneous patient responses highlight the need to better define T cell states and behaviors following therapy. Here, we establish PRECISE-seq, which integrates SrtA-catalyzed proximity labeling with single-cell sequencing to jointly profile antigen specificity, TCR potency, and phenotypic states. PRECISE-seq sensitively detects antigen-specific T cells, including low-frequency CMV- and tumor-specific clonotypes, while preserving native cellular states. Using this approach, we identify a regulatory T_Ly49 state enriched among tumor-specific T cells and demonstrate that PD-1 blockade promotes their transition toward an effector-like state associated with improved therapeutic responsiveness. These findings provide insight into how immune checkpoint blockade reshapes T cell fate in tumors.

Existing approaches largely rely on pMHC-TCR binding measurements or activation-based functional screens (Ma et al., 2021; Newell et al., 2013; Yu et al., 2022), which either fail to capture higher order cellular context (Sibener et al., 2018) or disrupt endogenous phenotypes (Oliveira et al., 2021; Yu et al., 2022) (Fig. S5Q). Although proximity labeling strategies have been developed, nonspecific or long-range labeling limits precision (Liu et al., 2022; Liu et al., 2020; Oslund et al., 2022; Zhang et al., 2023) (Fig. S5R). PRECISE-seq addresses these limitations through short-range SrtA-mediated ligation confined to the immunological synapse (Oakley et al., 2022) and uniform chemical conjugation of labeling substrates, enabling accurate capture of contact-dependent TCR recognition (Nakandakari-Higa et al., 2024; Zhang et al., 2022). Antigen specificity is inferred directly from T cell–APC contacts, without the need for predefined epitope information or engineered receptor systems, making it uniquely suited for translational studies. Importantly, labeling intensity provides a quantitative measure of relative functional avidity in situ, representing a distinguishing feature of PRECISE-seq. Moreover, recording occurs without in vitro activation or expansion, thereby preserving endogenous phenotypic states.

Our analysis further reveals that tumor-specific T cells preferentially adopt a T_Ly49 program characterized by inhibitory lectin-like receptor expression, exhaustion features, and immunoregulatory activity distinct from previously described NK-like effector states (Beltra et al., 2023). These findings complement recent studies demonstrating exhaustion features in tumor-specific T cells within the TME (Lowery et al., 2022). Functionally, we demonstrate that these cells directly suppress antitumor immunity by inducing cell death and inhibiting proliferation of T_EFF, providing a mechanistic basis for their protumoral effects in vivo. PD-1 blockade induces a transcriptional shift from T_Ly49 toward an effector phenotype, consistent with improved clinical outcomes in immunotherapy cohorts. While our study provides a new mode of action for αPD-1 therapy, the mechanisms underlying the regulation of T_Ly49 differentiation and their functional role in tumors could be explored in follow-up studies.

By enabling concurrent mapping of antigen specificity, interaction strength, and cell state, PRECISE-seq provides a framework for dissecting T cell responsiveness in vivo and informs the mechanistic understanding of immune interventions, supporting more precise immunotherapeutic strategies. In adoptive T cell therapy, PRECISE-seq may facilitate the identification of optimal TCR candidates, as quantitative assessment of interaction strength highlights clonotypes with superior potency and favorable phenotypes, such as effector- or memory-like states. In summary, PRECISE-seq represents a significant advancement in our understanding of T cell biology, immunotherapy response, and disease pathogenesis. It opens new avenues for developing innovative therapeutic interventions and personalized treatment strategies.

C57BL/6 mice were purchased from Charles River Laboratories. C57BL/6 congenic CD45.1 mice and CD45.1 OT-I (OT-I) mice were kindly provided by Professor Xiaohuan Guo (Tsinghua University, Beijing, China). CD45.1 P14 mice were kindly provided by Professor Hai Qi (Tsinghua University, Beijing, China). Spp1-CreERT2-IRES-EGFP mice (strain no. T049832) were purchased from GemPharmatech. Rag2^−/− mice were purchased from the Jackson Laboratory. All mice were bred and housed under pathogen-free conditions at Tsinghua University, Beijing, China. Mice of 6–10 wk of age were used for all animal experiments. All mice were used in accordance with Tsinghua University Animal Ethics Committee guidelines.

Unless otherwise noted, all peptides used in this study were synthesized by SciLight Biotechnology, with a purity exceeding 95%. The chemical structure of Biotin-LPETG*G (G* refers to 2-hydroxyacetic acid) was described in the previous study (Ge et al., 2019; Williamson et al., 2012). The NHS-PEG4-maleimide was purchased from Confluore Biological Technology Co., Ltd. Sequences of all peptides are included in Table S1.

For the plasmid constructs transfected into HEK293T cells, the CD40L and SrtA proteins were cloned into the PLVX vector encoding a fluorescent reporter (tdTomato) and a P2A linker at the N terminus. The SrtA protein with an N-terminal Flag tag was directed to anchor on the cell membrane through pDisplay system, and the protein sequence was described in the previous study (Pasqual et al., 2018). This tdTomato-P2A-pDisplay SrtA plasmid was also used in the lentivirus package to construct a SrtA-expressing stable cell line. A bicistronic gene and SrtA protein were inserted into the downstream of CD40L protein to co-express these two proteins on the cell surface. The CD40 protein was cloned into a PLVX vector with a C-terminal GFP reporter fusion. The 1G4-TCR cloned into the pHAGE-IRES-RFP vector was obtained from Professor Xin Lin’s group (Tsinghua University, Beijing, China). The CMV pp65_495–503-specific TCR sequence (clone 4, clone 40, clone 52, and clone 67), synthesized by RootPath, Inc., was cloned into the pHAGE vector. The CD19 sequence was cloned between the TCRα chain and TCRβ chain via P2A self-cleaving peptide linker, allowing CD19 to serve as a co-expressed surrogate marker for monitoring TCR surface expression. The mouse tumor-reactive TCRs were cloned into a retroviral expression vector (RVKM), which encodes an EGFP and a P2A linker at the N terminus. For SrtA protein expression and purification in Escherichia coli, the SrtA was cloned into a pET-28a⁺ vector with a C-terminal 6 × His tag.

The K562-HLA-A2 cells and TCR-deficient Jurkat T cell line, a kind gift from Professor Xin Lin (Tsinghua University, Beijing, China), were cultured in RPMI 1640 complete medium (Thermo Fisher Scientific) containing 10% fetal bovine serum (FBS, Gemini), 25 mM HEPES, and 1% nonessential amino acids. B16-OVA is an OVA-transfected clone derived from the mouse melanoma cell line B16-F10 (ATCC, CRL-6475) as previously described (Han et al., 2019). LLC (CRL-1642; ATCC) is a Lewis lung carcinoma cell line. A MC38 cell line was derived from chemically induced grade III adenocarcinoma in C57BL/6 mice (Deng et al., 2014), and B2M-deficient MC38 tumor cells were obtained from Dr. Deng Pan (Tsinghua University, Beijing, China). All cell lines were cultured in an incubator at 37°C under 5% CO₂ and verified to be free of Mycoplasma contamination. PBMCs collected from HLA-A2⁺ CMV-seropositive healthy donors were obtained from Peking University Cancer Hospital with consent forms. PBMCs were isolated by density gradient centrifugation using Ficoll (Thermo Fisher Scientific) and cryopreserved in a freezing medium containing 90% FBS and 10% DMSO. HLA typing was performed by staining with anti-HLA-A2-APC (clone BB7.2; BioLegend).

Lentiviral transduction was performed as described previously (Liu et al., 2021). In brief, HEK293T cells were plated on 6-cm dishes at a density of 1 × 10⁶ cells per plate 24 h before transfection. HEK293T cells were cotransfected with 3 μg of the lentiviral plasmid, 1.5 μg of envelope vector (pMD2.G), and 1 μg of packaging vector (psPAX2). The culture medium was replaced 8 h after transfection, and the viral supernatant was collected 48 h or 72 h later. The viral supernatant was filtered through a 0.45-μm PES Syringe Filter (Thermo Fisher Scientific) and then used to infect tumor cells in the presence of 10 μg/ml polybrene (Sigma-Aldrich). Tumor cells, including B16-OVA, MC38, and K562-HLA-A2, were infected with lentivirus to overexpress SrtA protein on the cell surface. The SrtA expression was assessed by staining with an anti-Flag tag (anti-DYKDDDDK) Ab (clone L5; BioLegend), and the tumor clones with high levels of SrtA expression were sorted. Both MC38-zsGreen-LCMVgp_33–41 (MC38-gp33) and LLC-zsGreen-LCMVgp_33–41 (LLC-gp33) are single-cloned cell lines, generated by sorting zsGreen-positive cells after lentiviral transduction with a construct expressing zsGreen fused to six tandem repeats of the LCMVgp_33–41 epitope (KAVYNFATM). A TCRαβ-deficient Jurkat T cell line was transduced with the viral supernatant expressing 1G4-TCR, and TCR expression was assessed based on the co-expression of CD3 and RFP. To generate CMV-specific TCR-expressing cells, TCRαβ-deficient Jurkat cells were transduced with a bicistronic lentiviral vector encoding a CMV pp65-specific TCR coupled to a CD19-P2A selection marker. TCR-positive populations were subsequently isolated by fluorescence-activated cell sorting (FACS) based on the dual surface expression of CD3 and CD19.

1 × 10⁶ live cells were suspended in 100 μl freshly prepared NHS-PEG4-maleimide solution (50 μM in sterile PBS) and incubated for 5 min at 37°C. Then, the labeling reaction was quenched with an equivalent volume of FBS-free 1640 medium, and the supernatant was removed after centrifugation. Subsequently, the cells were resuspended in PBS containing 200 μM AP-HA probe and incubated for 5 min at 37°C. Afterward, the cells were washed with complete 1640 medium. The AP-HA tag⁺ cells were then ready for subsequent labeling experiments or analysis. To confirm chemical reaction labeling efficiency, the AP-HA tag abundance was assessed by flow cytometry through staining with an anti-HA tag Ab (clone 16B12; BioLegend). The consistent 2-log increase in AP-HA tag ΔMFI served as a standardized internal quality control for substrate availability and a benchmark for cross-batch comparisons.

The surface of HEK293T cells was conjugated with a homogeneous level of the AP-HA tag, and then, we performed the cell surface sortagging using the purified SrtA protein. The SrtA protein was expressed and purified from E. coli BL21 (DE3) cells. Sortagging reactions were performed in a total 100 μl volume at 37°C for 1 h in DMEM complete medium. 5 × 10⁵ AP-HA tag⁺ HEK293T cells were incubated with 20 μM purified SrtA protein and 100 μM Biotin-LPETGG probe. HEK293T cells after sortagging were washed three times and stained with streptavidin Ab (BioLegend), followed by flow cytometry analysis.

3 × 10⁵ MC38-SrtA tumor cells were seeded into a 96-well U-bottom plate and pulsed with serially titrated OVA_257–264 peptide ranging from 10⁻⁵ μM to 10 μM in RPMI 1640 complete medium for 2 h at 37°C. Following incubation, the supernatant was replaced with RPMI 1640 complete medium. Splenocyte suspensions were prepared by mechanical disruption of spleens from CD45.1 OT-I mice, followed by red blood cell lysis using RBC lysis solution (eBioscience). Half of the splenocytes were conjugated with the AP-HA tag, while the remaining splenocytes were left untreated. 1 × 10⁶ AP-HA–tagged or untreated splenocytes were added and cocultured with MC38-SrtA tumor cells pulsed with OVA_257–264. As a negative control, splenocytes were also co-incubated with MC38-SrtA cells in the absence of antigen stimulation. Following a 2-h incubation at 37°C, the expression of CD69 (clone FN50; BioLegend) on the CD45.1⁺ OT-I T cell was assessed by flow cytometry.

SrtA⁺ aAPCs (3 × 10⁶ cells/ml) were pulsed with 1 mg/ml NY-ESO-1_157–165 or CMV pp65_495–503 in RPMI 1640 complete medium for 2 h at 37°C. After incubation, the SrtA⁺ aAPCs were washed once to remove redundant peptides. 1G4-JC5 cells were labeled with the AP-HA tag and resuspended with RPMI 1640 complete medium. Then, 1 × 10⁵ per well AP-HA tag⁺ 1G4-JC5 cells were seeded into 96-well U-bottom plates and incubated with SrtA⁺ aAPC preloading antigen at a ratio of 1:3 for 2 h at 37°C. The Biotin-LPETGG probe was added at the beginning of the coculture with a final concentration of 100 μM. After 2 h of incubation, the reaction was quenched with FACS buffer (PBS containing 2% FBS and 1 mM EDTA), and then, cells were harvested and washed twice with FACS buffer. Subsequently, cells were stained with Abs against the HA tag, Biotin, and CD69 (clone FN50; BioLegend), followed by flow cytometry analysis.

For the coculture assay of SrtA⁺ tumor cells and OT-I T cells, splenocyte suspensions were obtained by the mechanical disruption of spleens from OT-I mice, and red blood cells were then lysed in RBC lysis solution (eBioscience). Following AP-HA tag labeling, 1 × 10⁶ splenocytes were incubated with 3 × 10⁵ B16-OVA-SrtA or B16F10-SrtA in the presence of the 100 μM Biotin-LPETGG probe in a 96-well U-bottom plate. After 2 h of incubation at 37°C, cell mixtures were then stained and analyzed by flow cytometry. For the APL assay, MC38-SrtA tumor cells (3 × 10⁶ cells/ml) were pulsed with 10 μM APL in RPMI 1640 complete medium for 2 h at 37°C. After APL peptide loading, the MC38-SrtA tumor cells were washed once to remove redundant peptides for subsequent co-incubation experiments. To evaluate the sensitivity of PRECISE labeling, CD45.1⁺ OT-I T cells were serially diluted with CD45.2⁺ polyclonal CD8⁺ T cells at ratios ranging from 1:10 to 1:10,000. The T cell mixtures were conjugated with the AP-HA tag and subsequently cocultured with B16-OVA-SrtA tumor cells in the presence of a 100 μM Biotin-LPETGG probe in a 96-well U-bottom plate. Following a 2-h incubation at 37°C, the cell mixtures were then stained and analyzed by flow cytometry.

1 × 10⁶ splenocytes derived from CD45.1 OT-I mice were incubated with freshly prepared NHS-PEG4-Mal solutions at concentrations of 0, 10, 25, and 50 μM in a total volume of 100 μl for 5 min at 37°C. After the removal of the supernatant, the cells were resuspended in PBS containing 200 μM AP-HA probe and incubated for an additional 5 min at 37°C. Subsequently, the cells were washed with RPMI 1640 complete medium, and the abundance of the AP-HA tag was assessed by flow cytometry using an anti-HA tag Ab (clone 16B12; BioLegend). To investigate the impact of AP-HA tag abundance on labeling efficiency, 1 × 10⁶ splenocytes expressing distinct levels of AP-HA tag abundance were incubated with 3 × 10⁵ B16-OVA-SrtA in the presence of 100 μM Biotin-LPETGG probe in a 96-well U-bottom plate. After a 2-h incubation at 37°C, Biotin labeling signals on CD45.1⁺ OT-I T cells were analyzed by flow cytometry.

SrtA⁺ aAPCs (3 × 10⁶ cells/ml) were pulsed with 1 mg/ml CMV pp65_495–503 peptide in RPMI 1640 complete medium for 2 h at 37°C. After incubation, the SrtA⁺ aAPCs were washed once to remove redundant peptides and resuspended in RPMI 1640 complete medium. Cryopreserved PBMCs were thawed and conjugated with the AP-HA tag as previously described. Following AP-HA conjugation, a consistent 2-log increase in AP-HA tag ΔMFI was measured by flow cytometry. This served as a standardized internal quality control for substrate availability and a benchmark for cross-batch comparisons. 1 × 10⁶ per well AP-HA tag⁺ PBMCs were seeded into 96-well U-bottom plates and incubated with 3 × 10⁵ SrtA⁺ aAPCs loading with CMV pp65_495–503 peptide in the presence of 100 μM Biotin-LPETGG probe at 37°C. After 2 h of incubation, cells were washed with FACS buffer and stained with Abs against CD8 (clone RPA-T8; BioLegend), HA tag, Flag tag, and Biotin. For single-cell transcriptomic analysis, stained cells were further incubated with DNA-barcoded sample hashtag Abs (clone LNH-94; 2M2; BioLegend) for 20 min in PBS. Prior to the cell sorting procedure, 7-aminoactinomycin D (7-AAD) (BD Bioscience) was added to exclude dead cells. The Biotin⁺ CD8⁺ T cells, as well as total CD8⁺ T cells without AP-HA tag labeling, were sorted for PRECISE-seq and bulk TCR-seq analysis. For sorting of total CD8⁺ T cells, live CD8⁺ single cells were gated. For sorting of Biotin⁺ CD8⁺ T cells, live CD8⁺HA tag⁺ Biotin⁺ single cells were gated. A stringent gating threshold was established to maximize the detection of true positives while minimizing background noise. CD8⁺ T cells with Biotin intensity exceeding 10³ were defined as the Biotin⁺ population and sorted for sequencing.

SrtA⁺ aAPCs were pulsed with 1 mg/ml CMV pp65_495–503 peptide in RPMI 1640 complete medium at a density of 3 × 10⁶ cells/ml and incubated for 2 h at 37°C. Following incubation, the SrtA⁺ aAPCs were washed once to remove excess peptides and resuspended in RPMI 1640 complete medium. PBMCs were previously isolated from HLA-A2⁺ healthy donors with CMV infection and identified by staining with HLA-A*02:01-CMV NLV-tetramer-PE (HG08T15006; Helixgen). Cryopreserved PBMCs were thawed, and half of the cells were conjugated with the AP-HA tag as previously described, while the remaining PBMCs were left untreated to serve as a baseline control. AP-HA–tagged PBMCs (1 × 10⁶ cells per well) were seeded into 96-well U-bottom plates and incubated with 3 × 10⁵ SrtA⁺ aAPCs loaded with CMV pp65_495–503 at 37°C with 5% CO₂ for 2 h. Both cultured AP-HA⁺ PBMCs and untreated PBMCs were collected and stained with anti-CD8-BV421 (301036; BioLegend). Before sorting, 7-AAD was added to discriminate between live and dead cells. Live CD8⁺ T cells were sorted and subsequently processed for bulk TCR-seq.

Antigen specificity screening of CMV-specific TCRs was conducted by RootPath, Inc., with the experimental process performed as previously described (Moravec et al., 2025). Paired full-length TCRα and TCRβ chains, separated by a P2A linker, were synthesized and cloned into AAV vectors (PackGene). Healthy donor T cells were isolated using CD3 isolation kits (STEMCELL Technologies) and activated with CD3/CD28 beads (Thermo Fisher Scientific) in ImmunoCult Human T Cell Expansion media (STEMCELL Technologies), supplemented with 200 U/ml IL-2 (PeproTech). 2 days after bead activation, T cells were electroporated (program EO-115) with Cas9 ribonucleoprotein (RNP) using P3 buffer and Nucleofector 4D Electroporation System (Lonza). RNPs consisted of SpyFi Cas9 nuclease (Aldevron) duplexed with HPLC-purified single guide RNAs (Integrated DNA Technologies) targeting the human TRAC and TRBC1/2 loci. After electroporation, TCR knockout T cells were transduced with TCR library AAV (5 × 10⁴ genome copies per cell) and cultured for an additional 2 days. Beads were then removed, and cells were maintained in culture for 6–10 days.

For TCR discovery screens, TCR library-engineered T cells were cocultured overnight in duplicate with K562-HLA-A2 cells pulsed with CMV pp65_495–503 peptide at an effector-to-target (E:T) ratio of 1:1. After coculture, reactive T cells were sorted based on the expression of T cell activation marker 4-1BB (clone 4B4-1; BioLegend). Sorted T cells were subjected to bulk TCR-seq. The quality control and alignment of sequencing reads have been described previously (Moravec et al., 2025). CMV-specific clonotypes were defined as those significantly enriched (FDR <0.05 and OR >4) in the CMV pp65_495–503-pulsed group, compared with both the mock-pulsed and baseline groups, as determined by Fisher’s exact test.

In the IFN-γ ELISpot assay, paired full-length TCRα and TCRβ chains, separated by a P2A linker, were synthesized and electroporated in TCR-deficient T cells in PBMCs. 24 h after TCR mRNA electroporation, transient TCR-T cells were costained with anti-human TCRα/β receptor Ab (IP26) and anti-human CD137 (4B4-1) Ab, followed by flow cytometry analysis for TCR and CD137 cell surface expression as quality control of electroporation.

In coculture for the IFN-γ ELISpot assay, HLA-A2 mRNA was electroporated into K562 cells. After 3 h, cells were loaded with the tested peptide for 2 h. HLA-expressing K562 cells without peptide pulsing were used as HLA-only negative control. Individual TCRs were introduced into donor T cells via mRNA electroporation as described above. 6 h after electroporation, TCR-expressing T cells were cocultured individually with different APCs for 18 h, followed by IFN-γ ELISpot analysis. TCR-T only was added as a negative control. T cells expressing TCRs PMEL or C4DLT were cocultured with HLA-A2–expressing K562 cells pulsed with peptide gp100 and WT1.

CMV-specific TCR-expressing cells were seeded at a density of 1 × 10⁵ cells per well in a 96-well U-bottom plate and cocultured with K562-HLA-A2 cells at an E:T ratio of 1:1. The CMV pp65_495–503 peptide was serially diluted in RPMI 1640 complete medium and added to the cell mixture to achieve final concentrations ranging from 10⁻⁷ μM to 100 μM. Following a 2-h incubation at 37°C with 5% CO₂, the cells were collected and stained with anti-CD69 (clone FN50; BioLegend), anti-CD19 (clone HIB19; BioLegend), and anti-CD3 (clone UCHT1; BioLegend) for subsequent flow cytometry analysis.

For single-cell suspension preparation of tumor-infiltrating immune cells, collected MC38 tumor tissues were enzymatically digested with 0.26 U/ml Liberase TL, 1 mg/ml collagenase D, and 0.1 mg/ml DNase I (Roche) at 37°C for 30 min. Cells were filtered through a 70-μm cell strainer and washed twice with FACS buffer. Both B16-OVA tumor tissues and draining lymph nodes were mechanically disrupted into single-cell suspension. Subsequently, single-cell suspensions were conjugated with the AP-HA tag as previously described and cocultured with cognate SrtA⁺ tumor cells in 96-well U-bottom plates at a ratio of 3:1 (3 × 10⁵ SrtA⁺ tumor cells per well). The Biotin-LPETGG probe was added at a final concentration of 100 μM. After 2 h of incubation, cells were washed twice with FACS buffer and stained with Abs against CD45 (clone 30-F11; BioLegend), CD8 (clone 53-6.7; BioLegend), CD44 (clone IM7; BioLegend), HA tag, Flag tag, and Biotin.

For single-cell transcriptomic analysis, stained cells were further incubated with DNA-barcoded sample hashtag Abs (clone M1/42; 30-F11; BioLegend) for 20 min in PBS, followed by 7-AAD staining. The Biotin⁺ CD8⁺ T cell populations and live non–AP-HA-tagged tumor-infiltrating CD44⁺ CD8⁺ T cells were sorted for PRECISE-seq. For sorting of Biotin⁺ CD8⁺ T cells, live CD45⁺CD8⁺CD44⁺HA tag⁺Biotin⁺ single cells were gated. The Biotin⁺ population was defined according to that on CD8⁺ T cells under incubation with B2M-deficient MC38-SrtA tumor cells.

Flow cytometry was performed following the previously published paper with some modifications (Dong et al., 2021). For surface marker staining, the single-cell suspension was incubated with Fc Block Ab (clone 2.4G2; BioXCell) for 10 min, followed by staining with Zombie fixable viability dye (BioLegend) to exclude the dead cells. After a washing step, cells were stained for cell surface markers at 4°C for 20 min in PBS. Subsequently, cells were washed twice with PBS to remove dissociative fluorescent Abs for further flow cytometry analysis. For intracellular staining, cells were fixed with fixation buffer (BioLegend) on ice for 15 min and then permeabilized twice with diluted 1× Intracellular Staining Permeabilization Wash Buffer (10×) (BioLegend). Abs against IFN-γ (clone XMG1.2; BioLegend), granzyme B (clone QA16A02; BioLegend), FcεR1g (FCABS400F; Sigma-Aldrich), and granzyme C (clone SFC1D8; BioLegend) were added and incubated for 1 h on ice. Then, the cells were washed with permeabilization buffer before flow cytometry analysis. Flow cytometry data were collected on LSR II (BD) or Symphony and/or sorted on FACSAria II or FACSAria III cell sorters (BD Biosciences). Data were analyzed by FlowJo (Tree Star).

The gating procedure began with the selection of single, live CD45⁺ leukocytes, followed by the identification of CD8⁺ T cells. Subsequently, within the Lag3⁺PD-1⁺Ly108⁻CXCR3⁻ CD8⁺ T cell subset, further subdivision was performed based on the expression of 2B4 and Ly49E/F. T_Ly49 cells were defined as 2B4⁺ Ly49E/F⁺.

6- to 8-wk-old female C57BL/6 (CD45.2) mice were subcutaneously implanted with 5 × 10⁵ B16-OVA tumor cells. For adoptive T cell transfer, CD8⁺ T cells were isolated from the lymph nodes and spleens of CD45.1 OT-I mice using EasySep Mouse CD8⁺ T Cell Isolation Kit (STEMCELL). 5 × 10⁵ purified CD45.1⁺ OT-I T cells were intravenously injected into tumor-bearing CD45.2 mice 7 days after tumor inoculation.

3 × 10⁵ MC38 tumor cells were subcutaneously inoculated into C57BL/6 female mice. The length (a) and width (b) of the tumors were measured starting on day 6 and then every 2 days thereafter, and tumor volume was calculated with the formula (ab²)/2. When the average tumor size reached about 150–200 mm³, mice were randomly divided into two groups (n = 5–7 per group), and 200 μg αPD-1 Ab (BioXCell) was intraperitoneally injected at day 7 after tumor inoculation.

5 × 10⁵ MC38-gp33 or LLC-gp33 tumor cells were subcutaneously inoculated on CD45.2 C57BL/6 mice. On day 7 after inoculation, tumor-bearing mice received 4.5 Gy total body irradiation, followed by intravenous adoptive transfer of 2 × 10⁵ activated CD45.1⁺ P14 T cells. The tumor-bearing mice were randomly divided into two groups based on tumor volume (n = 5–7 per group), and 200 μg αPD-1 Ab (BioXCell) was intraperitoneally injected at day 10 after tumor inoculation.

Biotin⁺ and Biotin⁻ tumor-infiltrating CD8⁺ T cells from MC38 tumor-bearing CD45.1 mice were sorted as previously described. 1000 Biotin⁺ or Biotin⁻ CD8⁺ T cells suspended in PBS were intravenously injected into CD45.2 Rag2^−/− mice, followed by subcutaneous inoculation of MC38 tumor cells (3 × 10⁵ per mice). Tumor curves were monitored every 2 days from day 4 after tumor inoculation.

CD45.1⁺ P14 cells were harvested from the spleen and lymph nodes of CD45.1 P14 mice and subsequently stimulated with 1 μg/ml LCMVgp33 peptide (KAVYNFATM) at a density of 1 × 10⁸ cells/ml in RPMI 1640 complete medium at 37°C for 1 h. Following the removal of excess LCMVgp33 peptide, the splenocytes were resuspended at a density of 5 × 10⁶/ml in RPMI 1640 complete medium supplemented with 100 IU/ml recombinant human IL-2 (BioLegend), 10% FBS, 25 mM HEPES, 1% nonessential amino acids, 0.1% 2-mercaptoethanol, 100 U/ml penicillin, and 100 μg/ml streptomycin. After 24 h of activation, P14 T cells were collected for retroviral transduction or adoptive transfer.

HEK293T cells were seeded in 10-cm dishes at a density of 3 × 10⁶ cells per plate 24 h before transfection. HEK293T cells were cotransfected with 10 μg of retroviral plasmids (rvkm-TCRβ-P2A-TCRα-P2A-EGFP or rvkm-EGFP), 10 μg of pCL-Eco plasmid, and 40 μl of PEI (40,000 MW). The culture medium was replaced 8 h after transfection, and the viral supernatants were harvested at 48 or 72 h thereafter. The viral supernatants were filtered through a 0.45-μm PES Syringe Filter (Thermo Fisher Scientific) and either used immediately or cryopreserved at −80°C.

Retrovirus production was performed as described previously (Kurachi et al., 2017). Activated P14 T cells were seeded into a 24-well plate at a density of 1 × 10⁶ per well in RPMI 1640 complete medium supplemented with 4 μg/ml polybrene and 100 IU/ml IL-2. 1 ml of retroviral supernatants was added to the P14 T cell suspension, followed by gentle mixing through pipetting. The P14 T cells were then subjected to centrifugal transduction with the retrovirus at 2,000 g for 2 h. After centrifugation, the P14 T cells were incubated for an additional 4 h under standard culture conditions (37°C, 5% CO₂). Subsequently, the viral supernatants was replaced with RPMI 1640 complete medium containing 100 IU/ml IL-2, and transfected P14 T cells were incubated for an additional 2 days. The surface expression of the transgenic TCR was evaluated by flow cytometry based on the expression of EGFP, and the EGFP⁺CD45.1⁺CD8⁺ T cells were then isolated using a cell sorter.

C57BL/6 mice were subcutaneously inoculated with 3 × 10⁵ MC38 and received 4.5 Gy total body irradiation 7 days after inoculation. Within 12 h after irradiation, 5 × 10⁵ fluorescence-activated cell–sorted EGFP⁺CD45.1⁺CD8⁺ T cells (purity >95%) either antigen-specific TCRs or mock-GFP were intravenously transferred into MC38 tumor-bearing mice (n = 4–8 per group). The tumor tissue and draining lymph nodes were harvested 14 days after tumor inoculations, and the absolute numbers of transferred EGFP⁺CD45.1⁺CD8⁺ T cells were quantified by counting beads (BioLegend).

To validate the antigen specificity of tumor-reactive TCR in vitro, EGFP⁺CD45.1⁺CD8⁺ T cells expressing TCRs or mock-GFP were harvested from the tuDLN of MC38 tumor-bearing mice. These T cells were seeded at a density of 5 × 10⁵ cells per well in 96-well U-bottom plates and cocultured with 3 × 10⁵ MC38 tumor cells for 18 h. The expression of activation markers on EGFP⁺CD45.1⁺CD8⁺ T cells was evaluated by flow cytometry staining with Abs against CD69 (clone H1.2F3; BioLegend), CD25 (clone PC61; BioLegend), and IFN-γ (clone XMG1.2; BioLegend).

For donor mouse preparation, CD45.2 C57BL/6 mice were subcutaneously inoculated with 5 × 10⁵ MC38-gp33 tumor cells. 7 days after inoculation, the tumor-bearing mice received 4.5 Gy total body irradiation, followed by intravenous adoptive transfer of 2 × 10⁵ activated CD45.1⁺ P14 T cells 7 days after inoculation. Tumor tissues were harvested 18 days after inoculation and enzymatically dissociated into single-cell suspensions.

For evaluation of T_Ly49 in promoting tumor growth, T_Ly49 (LAG3⁺2B4⁺Ly49E/F⁺) and T_EM-T_EFF (LAG3⁻) populations were isolated from the tumor-infiltrated CD45.1⁺ P14 T cells by FACS using stringent gating criteria. Each subset (6,000 cells) was intratumorally injected into recipient mice bearing MC38-gp33 tumor. The recipient mice that received PBS injection without T cell transfer served as negative controls. Recipient mice were prepared by subcutaneous inoculation with 5 × 10⁵ MC38-gp33 tumor cells 7 days before injection and subsequent treatment with 4.5 Gy total body irradiation. Recipient mice were randomly assigned to three groups (n = 3–6 per group), and tumor growth was monitored every 2 days starting from day 7 after tumor inoculation.

To investigate the precursor of T_Ly49, T_EM (LAG3⁻CXCR3⁺) and T_PEX (PD-1⁺Ly108⁺Tim-3⁻) were isolated from the tumor-infiltrating CD45.1⁺ P14 T cells by FACS. Each subset (6,000 cells) was intratumorally injected into recipient mice bearing MC38-gp33 tumor (n = 3 per group). The preparation of recipient mice has been previously described. Tumor growth was monitored every 2 days starting from day 7 after tumor inoculation. The phenotypic compositions of tumor-infiltrating CD45.1⁺ P14 T cells in recipient mice were assessed by flow cytometry 11 days after T cell transfer.

Cell viability was examined before single-cell library preparation to ensure the high quality of the scRNA-seq and TCR profiling data. Over 90% of cells remained viable throughout the labeling and sorting processes, which minimized the risk of RNA degradation and ensured the generation of high-fidelity sequencing data. The libraries were prepared using Chromium Next GEM Single Cell 5′ GEM, Library & Gel Bead Kit version 1.1 (Cat: 1000165) purchased from 10X Genomics according to the protocols provided by the manufacturer. The target cell recovery for each library was aimed at 8,000 cells, and the libraries were sequenced on an Illumina HiSeq X Ten platform.

Libraries for TCR-seq were prepared from 10 ng of total RNA using SMARTer Human TCRα/β Profiling Kit version 2 (Takara Bio). 18 cycles were used for each of the two semi-nested PCR amplification steps. Sequencing was performed on an Illumina NovaSeq by multiplexed paired-read run with 2 × 251 cycles.

Single-cell libraries contained samples from the same patient used in another study; Cell Hashing was employed to label cells from distinct samples (Stoeckius et al., 2018). The paired scRNA-seq and scTCR-seq data were processed with the “cellranger multi” pipeline by Cell Ranger 6.0. Briefly, scRNA-seq reads were aligned to the GRCh38 genome, while scTCR-seq reads were aligned to the human GRCh38 V(D)J reference (version 5.0.0). The intronic reads were also counted, as suggested by Cell Ranger. The “cellranger aggr” command was applied to aggregate TCR clonotype outputs from cellranger multi. The CITE-seq reads were counted using CITE-seq-Count version 1.4.5 (https://github.com/Hoohm/CITE-seq-Count). Downstream analysis was conducted using the Seurat R package, version 3.2.3 (Stuart et al., 2019). Genes detected in fewer than 3 cells were removed, while cells identified by both scRNA-seq and CITE-seq were retained. The CITE-seq data underwent normalization using the centered log-ratio transformation. The HTODemux function was employed to demultiplex samples based on the signals of hashtag oligos, which label sample origins. For the Biotin⁺ sample, predicted doublets were removed, and negative/ambiguous cells were subjected to a second round of demultiplexing to exclude definitively hashtag-negative cells. For the baseline sample, predicted singlets were retained.

After defining singlets, quality control was conducted to remove low-quality cells from the total cell population, including doublets and hashtag-negative cells. Cells were excluded if they expressed fewer than 300 genes or if the proportion of mitochondrial gene expression exceeded 15%. Subsequently, UMI counts were normalized using the NormalizeData function in Seurat, after excluding mitochondrial and ribosomal protein genes. Highly variable genes (HVGs) were defined as the top 2,500 variable genes, excluding those listed in the blacklist provided by a recent pan-cancer T cell study (Zheng et al., 2021). The effects of total UMI counts and the percentage of mitochondrial genes were regressed out. These HVGs were then utilized for principal component (PC) analysis. The optimal number of PCs was determined using the ElbowPlot function. Subsequently, the shared nearest neighbor graph was constructed with the FindNeighbors function with the selected number of top PCs, employing nn.method = “annoy” and annoy.metric = “cosine.” Subsequently, the Louvain algorithm was applied to cluster cells with a resolution setting of 0.3. Clusters expressing markers of non-T cells, such as myeloid, B, and erythroid cells, were removed. scTCR-seq data were then integrated to identify cells with detected TCRs, and clusters with a low proportion of cells with detected TCRs were subsequently excluded. For TCR profiling, clonotypes with frameshift mutations, stop codons, or > 2 TCRα or TCRβ chains (indicative of doublets) were excluded from downstream analysis. This ensured only functional and biologically relevant TCR sequences were retained. Mucosal-associated invariant T cells, known to recognize metabolites presented on MR1 and potentially skewing the analysis, were removed (Kjer-Nielsen et al., 2012).

The scRNA-seq data were integrated with a single-cell atlas of PBMCs to define conserved human CD8⁺ T cell subsets (Mogilenko et al., 2021). The public scRNA-seq dataset of CD8⁺ T cells was downloaded from the Synapse repository (https://www.synapse.org; syn22255433). The log-scaled normalized data were converted to a count matrix and then processed through the preprocessing pipeline described above, including data normalization, scaling, and regression. Next, the public dataset was integrated with CD8⁺ T cells with detected TCR from PRECISE-seq, using the integration procedure in the Seurat R package. The public dataset and baseline CD8⁺ T cells were set as references. The Seurat pipeline, as described above, was executed to identify clusters. The CD8⁺ T cell clusters were then annotated based on the known label of CD8⁺ T cells from the public dataset. Subsequently, the expression levels of gene signatures were calculated using the AddModuleScore function in the Seurat package. The curated gene signatures applied to T cells were obtained from a recent pan-cancer T cell study (Chu et al., 2023).

The TCR clonotype was defined as a unique combination of the CDR3α and CDR3β nucleotide sequences. Cells with detected paired TCRα and TCRβ were retained. To identify high-confidence CMV-specific clonotypes, we established a statistical enrichment strategy for the analysis of PRECISE-seq data. Clonotypes were considered to be enriched if they met these criteria: (1) a minimum cell number threshold of 2 in the Biotin⁺ group; (2) significant enrichment in the Biotin⁺ group compared with the baseline group, with an FDR of <0.05 by Fisher’s exact test; and (3) an OR >4.

This formulation ensures that S_i reflects the relative enrichment of a clonotype in the Biotin⁺ group, independent of its baseline abundance. By normalizing against the variability in irrelevant TCR distributions, this calculation ensures that the potency score S_i is robust and comparable across samples and experimental conditions. Additionally, to ensure accurate potency score calculations and minimize noise from ultra-low-frequency clonotypes, we restricted the calculation of potency scores to clonotypes with baseline frequencies above a specified cutoff. The cutoff was determined based on the method used to estimate baseline frequencies: 0.1% for estimates derived from scTCR-seq data and 0.01% for those derived from bulk TCR-seq data. For CMV-specific clonotypes, baseline frequencies were established through bulk TCR-seq.

TCR-seq data were processed with Cogent NGS Immune Profiler (version 1.0) obtained from Takara Bio under default parameters. TCR clonotypes with any frameshift or stop codon were excluded from downstream analysis. Subsequently, the frequencies of TRA clonotypes and TRB clonotypes were then scaled to sum to 1, respectively. To simplify the process and facilitate integrated analysis with the single-cell data, clonotypes were defined at the amino acid level, with CDR3 sequences sharing identical amino acid composition collapsed into a single clonotype for analysis.

The scRNA-seq reads were aligned to the mm 10 genome using the “cellranger count” command, while scTCR-seq reads were aligned to the mouse GRCm38 V(D)J reference (version 7.0.0) using the “cellranger vdj” command. Then, the PRECISE-seq data were processed as previously described. Cells of low quality, characterized by the expression of <300 genes or > 15% mitochondrial gene expression, were removed. Clusters exhibiting a low rate of TCR detection or the expression of non-T cell markers were also removed. HVGs were defined as the top 2,500 variable genes, excluding the TCR variable region genes. Next, cell-cycle scores were assigned based on the expression of G2/M- and S-phase markers, using the CellCycleScoring function in the Seurat package. An interferon-response score was also assigned to each cell based on the expression of the HALLMARK_INTERFERON_ALPHA_RESPONSE gene signature from the MSigDB (Liberzon et al., 2015). The unwanted effects caused by the cell cycle, total UMI counts, the percentage of mitochondrial genes, and the interferon-response signature were removed by regression. The Biotin⁺ samples and the baseline samples were integrated using the integration procedure in the Seurat R package, with baseline samples set as the reference. Subsequently, the Seurat pipeline was utilized to identify cell clusters using the Louvain algorithm. Clusters exhibiting high levels of cell-cycle scores or interferon-response scores were removed. Then, we applied the Seurat clustering pipeline again to the filtered high-quality cells.

The TCR data from PRECISE-seq were analyzed following the methods previously described, with certain modifications. Briefly, cells with only a single TCR chain detected were also retained in order to increase the number of available cells for the analysis. Clonotypes with more than two TCRα or TCRβ chains were considered artifacts resulting from doublets and were excluded from the downstream analysis. Considering that tumor-specific TCRs generally exhibit lower affinity than viral-specific TCRs (Aleksic et al., 2012), the thresholds for defining high-confidence antigen-specific clonotypes were adjusted. Specifically, high-confidence tumor-specific clonotypes in the Biotin⁺ group were defined as those meeting the following criteria: (1) a minimum cell number threshold of 2; (2) a P value <0.1; and (3) an OR >2 compared with the baseline group. The potency score for each clonotype was calculated as previously described, using the frequencies determined by scTCR-seq data.

Two distinct algorithms, URD and diffusion map, were utilized to infer the developmental trajectories of CD8⁺ T cells from the PRECISE-seq data generated using the MC38 model. To infer the differentiation origin of the T_Ly49 subset, a tree-structured trajectory was constructed using the URD R package (version 1.1.1) (Farrell et al., 2018). Three precursor subsets (T_N/T_CM, T_PEX, and T_EM) were selected, with T_N/T_CM designated as the root of the tree, while T_EX and T_Ly49 subsets were assigned as the tips. The top 30 PCs previously calculated using the Seurat package were adopted. Logistic parameters were determined using the pseudotimeDetermineLogistic function with the settings optimal.cells.forward = 20 and max.cells.back = 40. To simulate random walks from all tips, the n.per.tip parameter was set to 25,000. The divergence method for constructing the tree structure was specified as “preference.”

In the diffusion map analysis, T_PEX, T_EM, T_EX, and T_Ly49 subsets from untreated tumors were utilized (Angerer et al., 2016). The differentially expressed genes (DEGs) among these four subsets were identified using the FindAllMarkers function with min.pct = 0.2. These DEGs were then intersected with the HVGs to generate an informative gene list. The expression of these informative genes was extracted from the integrated assay to create a diffusion map with the settings n_pcs = 20, n_eigs = 20, and sigma = 3.

The public scRNA-seq datasets were analyzed with the Seurat package. For scRNA-seq dataset of MSI CRC cohort from Li et al. (2023), 27 tumor samples from 19 patients were analyzed. These patients were treated with toripalimab (αPD-1 Ab) with or without celecoxib before curative surgical resection. Responders were defined as patients who achieved a pathological complete response, while nonresponders were defined as patients who did not achieve a pathological complete response. The processed data were downloaded from the GEO database. The cell barcodes for CD8⁺ T cells were obtained from Table S2 of the original publication. Gene expression normalization was carried out as previously described. HVGs were selected for individual patients with FindVariableFeatures function in the Seurat package, and were further combined using the SelectIntegrationFeatures function. The HVGs were then scaled using the ScaleData function.

For the scRNA-seq dataset of the CRC cohort from Chen et al. (2024), 44 tumor samples from 22 patients were analyzed. The processed data were downloaded from the GEO database with the accession number GSE236581. These patients received systematic anti-PD-1 immunotherapy and the following surgery if necessary. Responders were defined as patients who achieved a complete response. For the scRNA-seq dataset of HBV⁺ HCC cohort from Guo et al. (2025), 14 tumor samples from 9 patients were analyzed. The processed data were downloaded from the GEO database with the accession number GSE235863. These patients received the combination therapy of anti-PD-1 and lenvatinib without prior systemic treatment. Responders were defined in the original publication as including both complete and partial responses. For the scRNA-seq dataset of the melanoma cohort from Sade-Feldman et al. (2018), 45 tumor samples from 32 patients were analyzed (one sample was excluded due to a low number of CD8⁺ T cells available for the analysis). We analyzed the patients who were treated with αPD-1 Ab with or without αCTLA-4 Ab. Two samples from patients who only received αCTLA-4 Ab treatment were excluded. The normalized gene expression matrix was downloaded from the GEO database. The cell names for CD8⁺ T cells were obtained from Table S2 of the original publication. We selected and scaled the HVGs and then performed the clustering procedure to exclude a small subset of B cells.

To assign a cellular state for each CD8⁺ T cell in the public datasets, we first construct a reference profile for each CD8⁺ T cell subset identified in the PRECISE-seq data of the mouse tumor model. To enhance discrimination, the T_PEX subset was merged into the T_EX subset. Then, the marker genes for each subset were identified from PRECISE-seq data using the FindAllMarkers function. The informative genes were defined as the intersection of marker genes and HVGs. Subsequently, the scaled data of informative genes, stored in the “integrated” assay, were utilized to calculate the average profile of each CD8⁺ T cell subset. The mouse gene symbols of informative genes were then converted to their human one-to-one orthologs. The converted informative genes were further filtered based on their intersection with the HVGs of each query dataset. The Pearson correlations between each CD8⁺ T cell from public datasets and the reference profiles were calculated using the scaled data of these genes. The cellular state for each CD8⁺ T cell was assigned based on the most similar reference profile. The T_EM/T_EFF-to-T_Ly49/KIR cell ratio based on the assigned cellular state was used to split samples into two groups (ratio > 1 and <1). The Kaplan–Meier survival analysis was subsequently employed to assess the association between these groups and survival rates, with the survminer R package utilized for visualization.

Statistical analysis was performed using GraphPad Prism 8 (GraphPad Software). Unless otherwise noted in the figure legends, each dot represents a biological replicate. Data are presented as the mean ± SEM. Comparisons of two groups were performed by using two-tailed paired Student’s t test unless otherwise indicated, ns, P > 0.05; *P < 0.05; **P < 0.01; ***P < 0.001; or ****P < 0.0001. Fisher’s exact test was performed with Fisher’s test using MATLAB R202la. Information on specific statistical tests is listed in the figure legends and/or method details.

Fig. S1 shows the development of PRECISE‑seq and its capacity to label cell–cell contacts and antigen‑specific interactions. Fig. S2 shows the characterization of CMV‑specific human CD8⁺ TCR repertoires and phenotypes captured by PRECISE‑seq. Fig. S3 shows in vivo monitoring of tumor‑specific T cells and the identification/validation of tumor‑reactive TCRs. Fig. S4 shows phenotypic and clonotypic features of T_Ly49 cells compared with other CD8⁺ T cell states. Fig. S5 shows how tumor‑specific T cells skew toward a regulatory T_Ly49 program and how αPD‑1 therapy reshapes these states. Tables S1 and S2 list peptide sequences and hashtag Abs used for single‑cell sequencing.

The PRECISE-seq datasets reported in this paper have been deposited in the Genome Sequence Archive (GSA) (Chen et al., 2021) in the National Genomics Data Center (CNCB-NGDC Members and Partners, 2022), China National Center for Bioinformation/Beijing Institute of Genomics, Chinese Academy of Sciences. The human dataset has been deposited in the GSA-Human database with an accession number HRA006819 that is publicly accessible at https://ngdc.cncb.ac.cn/gsa-human. The mouse dataset is deposited in the GSA database with an accession number CRA015233 and can be found at https://ngdc.cncb.ac.cn/gsa. All other materials will be made available upon request.

We extend our gratitude to Xin Lin (Tsinghua University, Beijing, China) for generously providing the 1G4-TCR plasmid, K562-HLA-A2 cell line, and the TCR-deficient Jurkat T cell line, as well as to Deng Pan (Tsinghua University, Beijing, China) for supplying the B2M-deficient MC38 cell line used in this study. We also acknowledge Dr. Zai Chang from the Animal Facility at Tsinghua University for his invaluable support in animal care. Special thanks to Zhenghang Wang (Peking University Cancer Hospital & Institute) for providing PBMC samples from HLA-A2⁺ healthy donors.

This work was funded by the Beijing Natural Science Foundation (Z200023 to M.M. Xu; Z200010 to P.R. Chen), the National Natural Science Foundation of China (81922054 to M.M. Xu; 21937001, 22137001, and 91753000 to P.R. Chen), the Beijing Natural Science Foundation Tsinghua University Dushi Program (20251080034 to M.M.X), and the Ministry of Science and Technology (2021YFA1302603 to P.R. Chen). Additionally, funding was provided through the New Cornerstone Science Foundation via the New Cornerstone Investigator Program to P.R. Chen. This work was funded by Beijing Natural Science Foundation Tsinghua University Dushi Program (20251080034 to M.M. Xu).

Author contributions: Shibo Liu: data curation, formal analysis, methodology, and writing—original draft. Guanghao Liang: data curation, formal analysis, methodology, project administration, software, visualization, and writing—original draft, review, and editing. Yayun Yang: formal analysis, investigation, methodology, validation, visualization, and writing—original draft, review, and editing. Yanyang Shi: formal analysis, investigation, methodology, validation, visualization, and writing—original draft, review, and editing. Lihui Dong: investigation, project administration, and writing—original draft. Yue Zhao: investigation, methodology, and resources. Boyuan Mei: writing—review and editing. Jun Wang: resources. Feng Lin: resources and writing—review and editing. Yilin Li: methodology and resources. Wenxin Dong: methodology and supervision. Chengyang Liu: investigation. Yuhui Cao: visualization. Dali Han: formal analysis and supervision. Peng R. Chen: conceptualization, funding acquisition, supervision, and writing—review and editing. Meng Michelle Xu: conceptualization, funding acquisition, methodology, project administration, supervision, and writing—original draft, review, and editing.