The transcriptomic data and clinical information shown in the Supplemental Figures were obtained from the TCGA repository at GDC (https://portal.gdc.cancer.gov/). mRNA expression screening of OSCC patients was performed according to NCCN Guidelines Version 2.2021. The expression of CD3 and TCF7 transcripts, survival time and survival status were included in the downloaded survival data. To obtain reliable and ideal grouping, each gene was divided into high- and low-expression groups using the R packages “survminer” and “maxstat”; p values were calculated using the log-rank test.

The sample size followed the principle of 20 EPV (Event Per Variable) (12), and 167 paraffin tissue specimens from surgical resection at the Department of Oral and Maxillofacial Surgery, Sun Yat-sen Memorial Hospital, Sun Yat-sen University were collected from January 2015 to December 2019. All the patients met the following criteria: (a) a first pathologic diagnosis of OSCC; (b) no radiotherapy or chemotherapy before surgery; (c) preserved postoperative wax blocks; (d) available complete electronic medical records, histological pathology reports and follow-up records. Clinicopathological factors, including age, sex, alcohol consumption, tobacco factor, primary site location, T stage, lymph node metastasis, TNM stage, histological grade, CD3⁺ T cell expression, and CD3⁺TCF7⁺ T cell expression, were collected. The clinical stages and differentiation degrees of the tumor were classified based on TNM classification of the AJCC 7th edition and WHO grade classification, respectively. The 167 patients were followed up postoperatively every 3 months for 1 year and every 6 months from the second year for 5 years. Written informed consent was obtained from each patient participating in the study.

All the paraffin tissue specimens were fixed with 4% formaldehyde, embedded in conventional paraffin, serially sectioned at approximately 4 μm thickness, and dewaxed. EDTA alkaline antigen retrieval was performed, and the samples were blocked with goat serum. After incubation with the primary antibody, the slides were washed with TBST for 3 min, incubated with the secondary antibody (goat anti-mouse HRB, 10 min), washed with TBST for 3 min and incubated with a fluorescent dye (Panovue). The above process was repeated for additional markers, after which DAPI was added dropwise; after staining, an anti-quenching agent was added dropwise, and coverslips were sealed. The incubation conditions were as follows: TCF7 (cst, rabbit monoclonal antibody, 1:200, first cycle), CD3 (Abcam, rabbit monoclonal antibody, 1:100, second cycle), and DAPI (1:200). The tissue sections were scanned using the Akoya Vectra Polaris system (1.0.10; PerkinElmer, Inc), and images were analyzed using inForm software (2.4.6; PerkinElmer, Inc). The results were evaluated as follows: four fields of view were selected arbitrarily for each patient’s section, the percentage of CD3⁺ TCF7⁺ T cells among total CD3⁺ T cells in each field of view was obtained, and the sum was averaged as the percentage of TCF7⁺ T cells in the tumor. The optimal cutoff percentage values of TCF7⁺ T cells were divided in X-tile 3.6.1 (Yale University, New Haven, Connecticut, USA) software (13) based on OS and RFS.

Statistical analysis was performed using SPSS 25.0 (SPSS, Inc., Chicago, IL, USA) and R (R Foundation for Statistical Computing, Beijing, China). In the baseline table, the measures were expressed as means ± standard deviation, and χ² test was used for comparisons between groups. Cox regression was employed for both univariate and multivariate analyses. The p value was determined by the Benjamini & Hochberg (BH) method for multiple testing correction. p<0.05 was considered a statistically significant difference. The statistically significant variables in univariate analysis were subjected to multivariate analysis, and the variables ultimately selected were used as variables in the nomogram. Nomograms and calibration curves were plotted using R software (rms package, survivor package). Regarding the performance of the predictive model, the C-index was applied to evaluate the discrimination of different models, where the C-index was compared using the compareC function. Calibration curves were used to assess calibration of the model under 1000 resamplings (B=1000). We used decision curve analysis to compare clinical utility among the models at 1, 3 and 5 years.

We first estimated the prognostic value of CD3⁺TCF7⁺ T cells in OSCC patients at the transcriptional level. The transcriptome data for 212 OSCC patients and their survival information were downloaded from TCGA. The patients were divided into a CD3^hi TCF7^hi (DH) group, CD3^low TCF7^low (DL) group, CD3^hi TCF7^low (CD3^hi) group, and CD3^low TCF7^hi (TCF7^hi) group based on transcript levels. No significant difference (p=0.160) was found in the OS rates among the four groups ( Supplementary Figure 1A ). However, statistical significance (p=0.041) was shown for the DH group and DL group ( Supplementary Figure 1B ), indicating that the DH group had a better prognosis based on transcriptional data.

To test whether TCF7⁺ T cells also influence the prognosis of OSCC patients, the number of TCF7⁺ T cells was assessed by multiplex immunofluorescence analysis in 167 OSCC patients ( Figure 1 ). The proportion of TCF7⁺ T cells to CD3+ T cells was used as a continuous variable. The 167 patients were divided into a TCF7⁺ T cell low expression group (hereafter referred to as the TCF7^low group) ( Figures 1E, F ) and a TCF7⁺ T cell^hi expression group (hereafter referred to as the TCF7^hi group) according to the results calculated by X-tile software ( Figures 1A – D, G, H ). The cutoff value in the OS cohort was 0.39, with 81 patients assigned to the TCF7^low group ( Figure 1B ); the cutoff value in the RFS cohort was 0.41, with 88 patients assigned to the TCF7^low group ( Figure 1D ). Based on both OS and RFS survival curves, the TCF7^hi group had a more favorable prognosis than the TCF7^low group ( Figures 1A, C ). Associations between the TCF7^hi and TCF7^low groups in the OS cohort and RFS cohort and clinicopathological characteristics are summarized in Table 1 . TCF7⁺ T cells were associated with T stage, lymph node metastasis and TNM stage, with statistically significant differences (p<0.05), but not with sex, age, tobacco, alcohol, histological grade, or CD3 density (p> 0.05). Clinicopathological baseline tables for the TCF7^hi and TCF7^low groups were established ( Table 1 ).

We performed univariate and multivariate analyses to further assess whether the TCF7⁺ T cell proportion is an independent predictor of OS and RFS. Univariate analysis showed ( Table 2 ) that the TCF7⁺ T cell percentage as a continuous variable was a risk factor for OS and RFS (OS: HR=0.013, 95% CI=0.002 to 0.074, p<0.001; RFS: HR=0.041, 95% CI=0.009 to 0.185, p<0.001) and that T stage (T2-OS: HR=2.187, 95% CI=0.814 to 5.872, p=0.185; T2-RFS: HR=1.056, 95% CI=0.572 to 1.950, p=0.919; T3-OS: HR=8.621, 95% CI=3.709 to 20.039, p<0.001; T3-RFS: HR=2.645, 95% CI=1.592 to 4.397), lymph node metastasis (N1-OS: HR=4.510, 95% CI=2.077 to 9.792, p<0.001; N1-RFS: HR=3.548, 95% CI=1.972 to 6.382, p<0.001; N2a-OS: HR=7.789, 95% CI=2.718 to 22.327, p<0.001; N2a-RFS: HR=5.077, 95% CI=2.064 to 12.490, p<0.001; N2b-OS: HR=7.770, 95% CI=4.065 to 14.854, p< 0.001; N2b-RFS: HR=4.770, 95% CI=2.881 to 7.900, p<0.001; N2c-OS: HR=9.796, 95% CI=1.261 to 76.104, p=0.0517, N2c-RFS: HR<0.001, 95% CI=0 to infinity, p=0.996), TNM stage (II-OS: HR=1.695, 95% CI=0.401 to 7.155, p=0.541; II-RFS: HR=0.822, 95% CI=0.364 to 1.857, p=0.113), histological grade (OS: HR=0.471, 95% CI=0.286 to 0.776, p=0.006; RFS: HR=0.564, 95% CI=0.367 to 0.868, p=0.018) were also risk factors for OS and RFS. Additionally, the percentage of TCF7⁺ T cells was transformed into a dichotomous variable and then included in multivariate analysis ( Table 3 ). Thus, the TCF7⁺ T cell expression level was a risk factor for OS and RFS (OS: HR=0.235, 95% CI=0.133 to 0.414, p<0.001; RFS: HR=2.661, 95% CI=1.675 to 4.227, p<0.001). In multivariate analysis, TNM stage (OS: HR=8.539, 95% CI=2.144 to 34.013, p=0.004; RFS: HR=4.992, 95% CI=1.054 to 23.649, p=0.086) and histological grade (OS: HR=0.408, 95% CI=0.246 to 0.679, p=0.002; RFS: HR=0.469, 95% CI=0.303 to 0.724, p=0.001) were risk factors for OS and RFS.

Patients were divided into two clinical risk levels by TNM stage: TNM (I+II) and TNM (III+IV) ( Figure 2 ). Differences in the percentage of TCF7⁺ T cells regarding OS and RFS were analyzed separately. Although differences in TNM (I+II) between the TCF7^hi group and TCF7^low group were not statistically significant for OS or RFS (OS: p=0.130, Figure 2A ; RFS: p=0.780, Figure 2B ), differences between the TCF7^hi group and TCF7^low group for TNM (III+IV) were statistically significant (OS: p<0.001, Figure 2A ; RFS: p<0.001, Figure 2B ). Similarly, histological grade classified patients into two pathological risk levels: a well-differentiated group (well) and a poorly differentiated group (poor). In the well-differentiated group, differences in OS and RFS between the TCF7^hi and TCF7^low groups were not statistically significant (Well-OS: p=0.006, Figure 2C ; Well-RFS: p=0.290, Figure 2D ). However, in the poorly differentiated group, differences in OS and RFS were both statistically significant (Poor-OS: p<0.001, Figure 2C ; Poor-RFS: p<0.001, Figure 2D ).

In a previous analysis, TCF7⁺ T cell expression could be used as an independent predictor of postoperative OSCC; this variable alone was included in the Cox risk regression model, which calculated a C-index=0.701 (95% CI=0.654 to 0.748) for TCF7⁺ T cell expression for OS and a C-index=0.654 (95% CI=0.610 to 0.698) for RFS, ( Table 4 ), suggesting that the model has good predictive accuracy. Next, significantly different variables (TNM stage, histological grade, and TCF7⁺ T-cell expression) were obtained from the results of multivariate analysis. Different Cox risk regression models using the C-index as a measure were established to investigate whether inclusion of the TCF7⁺ T cell percentage enhances the predictive accuracy of the TNM stage and histological grade. Inclusion of the TCF7⁺ T cell percentage improved the prediction accuracy of the TNM stage and histological grade ( Table 4 ). The variable grade also improved the C-index from 0.594 (0.539–0.649) to 0.746 (0.690–0.802) for OS (p<0.001) and from 0.564 (0.511–0.617) to 0.686 (0.631–0.740) for RFS (p<0.001) after including the TCF7⁺ T cell percentage (p<0.001). The variable TNM stage increased the C-index from 0.732 (0.685–0.779) to 0.798 (0.752–0.838) for OS (p<0.001) and from 0.690 (0.646–0.734) to 0.740 (0.695–0.785) (p<0.001) for PFS. In TNM+Grade multivariate analysis, the C-index increased from 0.765 (0.715–0.815) to 0.820 (0.778–0.862) for OS (p<0.001) and from 0.728 (0.678–0.777) to 0.758 (0.711–0.805) for RFS after including the TCF7⁺ T cell percentage (p = 0.004).

We incorporated the TNM stage, histological grade, and TCF7⁺ T cell percentage status into the Cox risk regression model to construct two nomograms to predict OS and RFS in OSCC patients at 1, 3, and 5 years after surgery ( Figures 3A, E ). The C-index displayed good predictive accuracy for both the OS and RFS nomograms (OS: C-index=0.728, 95% CI=0.678 to 0.777); C-index=0.758, 95% CI=0.711 to 0.858) ( Table 4 ). We also constructed calibration curves for predicting OS and RFS at 1, 3 and 5 years by internal resampling 1000 times ( Figures 3B – D , F – H ), and the calibration curves showed good consistency in predicting and the actual event occurrence of patients.

We further compared the clinical utility of models (“TNM+Grade+TCF7” vs. “TNM+Grade”) using decision curves within 1, 3 and 5 years ( Figure 4 ). First, for both the “TNM+Grade+TCF7” and “TNM+Grade” prediction models, the DCA showed that the net benefit range was greater than that of the prediction model with all variables included (“All”). Second, the area under the decision curve of “TNM+Grade+TCF7” was larger than that of “TNM+Grade” at 1, 3, and 5 years for OS ( Figures 4A – C ) and RFS ( Figures 4D, E ) but smaller than the area under the decision curve at year 5 ( Figure 4F ). This finding also indicated that adding the TCF7⁺ T cell expression level improved the clinical utility of the existing model, suggesting that TCF7⁺ T cells warrant further research.