Patients were enrolled from January 2015 to June 2017 at the Zhongshan Hospital Fudan University (Shanghai, China). HCC was diagnosed and treated according to the American Association for the Study of Liver Diseases guidelines (5). Inclusion criteria were 1) definitive pathological diagnosis of HCC based on the WHO criteria; 2) curative resection, defined as complete macroscopic removal of the tumor (29); 3) no prior anticancer treatment, including chemotherapy, radiotherapy, targeted therapy, or immune therapy; and 4) no history of other types of malignancies. A total of 58 eligible patients were recruited into this study. Tumor samples, adjacent non-tumor tissue samples, or peripheral blood samples were collected for high-throughput TCR sequencing. Patients’ liver function was evaluated using the Child–Pugh score and albumin–bilirubin (ALBI) score (30). The tumor stage was determined according to the Barcelona Clinic Liver Cancer (BCLC) and China Liver Cancer (CNLC) staging system (31, 32). All peripheral blood samples were obtained before surgery, and all surgically resected tissue samples were confirmed independently by 2 experienced pathologists.

All patients were prospectively monitored by liver function test, serum AFP, and abdominal ultrasonography at each follow-up visit. MRI or CT of the abdomen was performed every 6 months. If intrahepatic recurrence or distal metastasis was clinically suspected on the basis of symptoms or unexplained elevation of tumor marker levels, MRI, CT, or bone scan was performed immediately. Every patient was followed up every 2 to 3 months in the first 2 years after surgery and then every 3 to 6 months thereafter. Overtime survival was the interval from the date of surgery to either the date of death or the last follow‐up visit. Time to recurrence (TTR) was defined as the interval between the date of surgery and the date when intrahepatic recurrence or extrahepatic metastasis was first diagnosed (33).

The study protocol was approved by the Institutional Ethics Committee of the Zhongshan Hospital, Fudan University, with approval number B2019-060(2) and was conducted according to the ethics guidelines of the 1975 Declaration of Helsinki. Written informed consent was granted by each of the recruited patients.

The rearranged TCR repertoires were amplified by the multiplex PCR (MPCR) method described in previously published papers (34). In short, genomic DNA for each sample was extracted using a Genomic DNA Extraction Kit (TIANGEN, Beijing, China, Cat# DP304-03#) and then amplified using the Multiplex PCR Kit (QIAGEN, Valencia, CA, USA, Cat# 206143#) with 32 forward V primers and 13 reverse J primers. The cycling conditions were as follows: 95°C for 15 min, 30 cycles at 94°C for 30 s, 60°C for 90 s, and 72°C for 30 s, plus a final extension at 72°C for 5 min. Before the beginning of the study, the synthesized various TCRβ templates were used to evaluate the PCR bias of multiple primers and then supported us with the optimized primer mix to minimize the PCR bias (35). At last, the PCR product was purified by running on a 2% agarose gel reaction, and then paired-end 100-bp reads were outputted on the BGISEQ-500 sequencer.

IMonitor (36) pipeline was used to analyze TCR sequencing data. First, low-quality reads were discarded, and clean paired-end reads were merged by their overlapping nucleotides. Second, the merge reads were assigned to VDJ germline segments and alleles from IMGT (http://www.imgt.org/) database by BLAST alignment and further re-alignment. Third, the sequencing error was corrected. Then the abundant and Shannon’s entropy of each TCR VDJ segment was calculated.

Principal component analysis (PCA) of peripheral blood from 439 healthy donors (37) and 43 HCC patients was performed by the scaled abundance of overlapped 624 TCRβ VJ pairings. The prcomp () function in R was used to performed PCA, and principal component (PC) 1 and PC2 were used for visualization.

Discriminative TCRβ VJ pairings between peripheral blood from healthy donors (37) and 43 HCC patients were identified using the “FindAllMarkers” function, implemented in the Seurat package; discriminative TCRβ VJ pairings with log-scaled fold change ≥0.25 and area under the curve (AUC) >0.7 between two groups of samples were identified using the receiver operating characteristic (ROC) test.

LASSO regression analysis was performed by glmnet package to construct a model to predict HCC patients and healthy people based on TCRβ VJ pairing usage from peripheral blood. By using random partitioning of the dataset, the 80% samples of 43 HCC (35/43) were randomly selected and set as the training group, and the rest of the 20% (8/43) of samples were used as the test group. In the training group, the number of HCC samples and normal controls (37) was sampled equally to build the classification model. Finally, 35 samples were randomly selected from both HCC patients and healthy people as the training group (70 samples). And the rest of the samples were selected from both HCC patients and healthy people as the testing group (412 samples). The response type of model output is set “binomial.” The lambda that gives minimum mean cross-validated error was determined by the “cv.glmnet” function. Given the lambda, the selected features and coefficient of these features were determined. Then the prediction result was calculated by the regression model using the “predict” function. The AUC was calculated by the ROC R package. To exclude just a chance occurrence and demonstrate the robustness of the LASSO model, we repeated the above process 100 times. The model coefficients for features, accuracy, recall sensitivity, precision, and specificity from 100 repeats are recorded.

Paired t-test was used to compare the abundance and diversity of TCRβ CDR3 clonotypes among three tissues. The comparisons of multiple groups and the abundance of TCRβ VJ among tumors, adjacent normal tissues, and peripheral blood were performed using the paired Kruskal–Wallis test. The unpaired Wilcoxon test was used to compare the proportion of shared clones of each two tissues, while the paired Wilcoxon test was used to compare the proportion of shared clones/clonotypes among three tissues. The p-values of multiple comparisons were adjusted by the Bonferroni correction. A chi-square test was used to test the association between these clonotypes with patient clinical parameters. All statistical analyses and presentations were performed using R.

To explore the heterogeneity of TCR features among different tissue types in HCC patients, we performed TCR sequencing of 40 tumor samples (T), 37 adjacent non-tumor tissue samples (N), and 43 peripheral blood samples (B) collected from 58 patients diagnosed with HCC. Some of the 58 patients collected only one tissue of T, N, and B; some patients collected two of them, while some patients collected three of them. The specific tissues collection strategy is shown in Figure 1 and Supplementary Table 1 . The baseline characteristics of the HCC cohort are shown in Table 1 . For a total of 120 samples from 58 patients, we acquired an average of 11,037,670 ± 5,581,145 clean sequencing reads, 47,565 ± 25,114 distinct CDR3 nucleotide (nt) clonotypes, and 30,246 ± 15,807 distinct CDR3 amino acid (aa) clonotypes ( Supplementary Figure 1 ). For each sample, a total of 624 VJ gene combinations were identified, including 48 distinct Vβ gene segments and 13 distinct Jβ gene segments. The detailed sequencing data information is shown in Supplementary Table 1 .

As part of the TCR variable chain, CDR3 is crucial for T-cell recognition of antigens, and it directly reflects the T-cell immune response status. To investigate differences of TCRβ repertoires across tumors, adjacent non-tumor tissues, and peripheral blood, we compared the number ( Figure 2A ) and the diversity ( Figure 2B ) of CDR3 aa among these tissues. There was no significant difference in the number of CDR3 among three types of tissues (T vs. N, p = 0.05, T vs. B, p = 0.22, N vs. B, p = 0.52; paired t-test; Figure 2A ). Shannon’s entropy was used to measure the CDR3 diversity (38). A higher value of Shannon’s entropy indicates a higher diversity, which is also negatively correlated with the abundance of top 100 CDR3 (r = −0.869, p < 2.2e−16; t-test; Figure 2C ). The CDR3 diversity was significantly higher in peripheral blood than in the other two types of tissues, whereas it was similar between tumor and non-tumor adjacent tissues (T vs. N, p = 0.25, T vs. B, p = 1.3e−05, N vs. B, p = 2.1e−09; paired t-test; Figure 2B ). The result implied a higher TCR diversity in peripheral blood. When combined with the results of Figures 2A, B , we can speculate the existence of expanded TCR clones in both non-tumor and tumor tissues.

To investigate the intra- and inter-patient heterogeneity of the TCRβ CDR3 repertoire, the abundance of shared clonotypes was calculated between any two types of tissues from one patient or between any two patients. We observed that the abundance of shared clonotypes abundance between any two tissues within a patient was significantly higher than that between different patients ( Figure 3A ), revealing higher heterogeneity among patients. The tumor and adjacent non-tumor tissues shared higher fraction of clonotypes abundance than either of them shared with peripheral blood (Kruskal–Wallis test, p < 2.2e−16; the mean percentage of TN/T vs. TB/T = 0.608 ± 0.203 vs. 0.157 ± 0.117, p = 4.0e−08; TN/N vs. NB/N = 0.634 ± 0.162 vs. 0.274 ± 0.161, p = 6.5e−07; Wilcoxon test, p-value adjusted by Bonferroni correction; Figure 3A , left graph). This suggested that the majority of tumor-infiltrating T cells were recruited from adjacent non-tumor tissues rather than directly from peripheral blood. To some extent, the results calculated by TCR sequencing data in previous tumor environment studies (39, 40) were similar to our results can also support our point ( Supplementary Figures 2A, B ). In addition, the clones shared between peripheral blood and adjacent non-tumor tissues were not significantly different from those shared between peripheral blood and tumors (Kruskal–Wallis test, p < 2.2e−16; the mean percentage of NB/N vs. TB/T = 0.274 ± 0.161 vs. 0.157 ± 0.117, p = 0.089; Wilcoxon test, p-value adjusted by Bonferroni correction; Figure 3A , left graph).

We further compared the proportion of shared clones and clonotypes among three tissues (TNB) in a single tissue. The proportion of shared clones means the abundance of shared clonotypes in each tissue, which is the ratio of reads of shared clonotypes to the total reads. The proportion of shared clonotypes refers to the ratio of the number of shared clonotypes to the number of all clone types. The results showed that the proportion of shared clones was the highest in adjacent non-tumor tissues, lower in the tumor, and the lowest in peripheral blood (the mean proportion of TNB/N = 0.266 ± 0.130, TNB/T = 0.135 ± 0.134, TNB/B = 3.68e−03 ± 3.00e−03, TNB/N vs. TNB/T, p = 0.032; TNB/T vs. TNB/B, p = 3.7e−04; paired Wilcoxon test, p-value adjusted by Bonferroni correction; Figure 3B , left graph). However, for the proportion of shared clonotypes, there was no difference between tumor and adjacent non-tumor tissues, while both tumor and adjacent non-tumor tissues had a higher proportion than peripheral blood (the mean proportion of TNB/T = 3.37e−03 ± 3.72 e−03, TNB/N = 2.11e−03 ± 2.81e−03, TNB/B = 6.47e−04 ± 5.11e−04, TNB/T vs. TNB/N, p = 0.46; TNB/N vs. TNB/B, p = 1.1e−03; TNB/T vs. TNB/B, p = 3.7e−04; paired Wilcoxon test; Figure 3B , right graph). The degree of clonal expansion in adjacent non-tumor tissues was significantly higher than that in tumoral tissues (340.03 vs. 76.94, p = 2.3e−03; paired Wilcoxon test) ( Figure 3C ), which implicated that the clonal expansion of tumor-infiltrating T cells was compromised in the tumoral immunosuppressive microenvironment.

Despite the obvious heterogeneity of CDR3 clonotypes among patients, we found three shared clones in all 43 peripheral blood samples ( Figure 3D ). The shared clonotypes are TRBV20-1_AHPEDSSFYICSAGSQGKQETQYFGPGTRLLVLx_TRBJ2-5, TRBV9_xLYFCASSGSNSPLHFGNGTRLTVTGMGAPLFTx_TRBJ1-6, and TRBV4-3_EDSALYLCASSHIGRRREQFFGPGTRLTVLGKKx_TRBJ2-1, with an average fraction of 0.558%, 0.407%, and 0.168%, respectively ( Figure 3D ). However, we did not find any annotation for peptides or antigens recognized by these TCRs from databases or published papers. Considering significant clonotypes are not necessarily present simultaneously in all peripheral blood samples, we selected clonotypes that present in most peripheral blood samples and grouped them into with or without an identical clonotype. Then we used the chi-square test to test the association between these clonotypes with patient clinical parameters. The result shows that several clonotypes are significantly associated with HCC prognosis, HBV DNA/antigen presence, cirrhosis, etc. ( Figure 3E ).

To profile the landscape of TCRβ V and J genes, we calculated the usage abundance of TCRβ V and J gene segments for every sample in all three tissue types. In abundance heatmap result ( Supplementary Figure 3A ), we can find that the usage pattern of Vβ genes in peripheral blood such as TRBV18, etc., are seemingly distinguished from tumor and adjacent non-tumor tissues. In correlation coefficient analysis ( Supplementary Figure 3B ), the majority of correlation coefficients higher than 0.6 show the overall similarity across different tissues. However, we can also find that the usage pattern of Vβ genes and VJ pairings in peripheral blood is seemingly distinguished from tumor and adjacent non-tumor tissues. In general, usage patterns of Vβ genes in peripheral blood were a little distinguished from those in tumor and adjacent non-tumor tissues, while the usage patterns of Jβ genes in different tissues were similar. In the respective five most abundant Vβ genes of three tissues, we found that there were four overlapped ones, including TRBV6-6 (average ratio in T, N, B: from 9.07% ± 4.94% to 14.25% ± 6.48%), TRBV30 (average ratio in T, N, B: from 5.46% ± 5.25% to 18.99% ± 9.63%), TRBV28 (average ratio in T, N, B: from 4.94% ± 4.83% to 8.75% ± 11.01%), and TRBV20-1 (average ratio in T, N, B: from 4.75% ± 3.87% to 6.89% ± 2.97%). In the respective five most abundant Jβ genes of three tissues, the Jβ gene TRBJ2-7 (average ratio in T, N, B: from 14.95% ± 6.15% to 20.34% ± 10.14%) ranked the highest in all tissues. Other overlapped Jβ genes were TRBJ1-2 (average ratio in T, N, B: from 12.87% ± 6.77% to 19.33% ± 6.42%) and TRBJ2-1 (average ratio in T, N, B: from 10.21% ± 4.40% to 11.90% ± 4.85%; Figures 4A, B ). TRBV20-1, TRBJ2-7, and TRBJ2-1 have also been reported as the most common abundant gene segments in previous studies of HBV-associated HCC, pancreatic cancer, and nasopharyngeal carcinoma, while TRBV6-6 has also been found as the frequent gene segment in another study of HCC (23, 26, 28, 41).

Some V and J gene segments demonstrated significantly different usage abundances in different tissues (all p < 0.05, Kruskal–Wallis test) ( Figure 4C ). For example, TRBV30, TRBV18, and TRBV5-1 exhibited significantly higher usage abundances in peripheral blood than those in tumoral and peritumoral tissues (all adjusted p < 0.05, Wilcoxon test, p-value adjusted by Bonferroni correction). It is reported that TRBV30 has low usage abundance in healthy blood donors and some diseases, including acute myocardial infarction (AMI) or colorectal cancer (CRC) (42–45) but exhibited high usage abundances in the peripheral blood of patients infected with H5N6 influenza virus and patients with systemic lupus erythematosus (37, 46). TRBV18 was detected more abundantly in tumor tissues than in adjacent non-tumor tissues of lung cancer (47). TRBJ1-2 was used more abundantly in tumors than in adjacent non-tumor tissues (adjusted p = 3e−05, Wilcoxon test, p-value adjusted by the Bonferroni correction). The tissue distribution pattern of TRBJ1-2 was consistent with that in the previous study of HBV-associated HCC (27). In addition, TRBV6-1/2/3/4, TRBV28, and TRBJ2-7 exhibited significantly higher abundances in tumor and adjacent non-tumor tissues than in peripheral blood (all adjusted p < 0.05, Wilcoxon test, p-value adjusted by the Bonferroni correction). Moreover, TRBV6-6 showed the highest usage in tumor tissues, lower in adjacent non-tumor tissues, and the lowest in peripheral blood, which implied that the expansion of this TCRβ V gene-related clones may be due to stimulation of tumor antigens (adjusted p < 0.05, Wilcoxon test, p-value adjusted by the Bonferroni correction).

We also compared the usage pattern of TCRβ VJ pairings among three tissues. There were 26 dominant VJ pairings (mean abundance among three tissues >1%) showing significant differences in their usage abundances among three tissues (p < 0.01, Kruskal–Wallis test; Figure 4D ). Among them, VJ pairings such as TRBV20-1_TRBJ1-2, TRBV28_TRBJ1-2, and TRBV6-6_TRBJ1-2 were significantly enriched in the tumor, whereas other VJ pairings such as TRBV30_TRBJ1-1, TRBV30_TRBJ1-2, and TRBV30_TRBJ2-7 exhibited higher abundance in peripheral blood.

Later on, we assessed the association between the TCRβ CDR3 features and HCC clinical characteristics, including tumor size, microvascular invasion (MVI), and the AFP concentration in the blood. TCR diversities of three tissues showed no significant association with any clinical features ( Supplementary Figure 4A ). However, the proportion of shared clones among different tissue types showed association with some clinicopathological characteristics ( Supplementary Figures 4B, C ). Patients with a lower risk of postoperative recurrence tended to have a higher proportion of blood-peritumoral tissues shared clones in blood TCR repertoire (NB/B) (p < 0.05, Wilcoxon test; Supplementary Figure 4B ). A higher proportion of blood-peritumoral tissues shared clones in peritumoral TCR repertoire (NB/N) was correlated with non-MVI, tumor size <5 cm, and negative HBV-DNA load (all p < 0.05, Wilcoxon test; Supplementary Figure 4B ). Moreover, patients without MVI exhibited a higher proportion of tumor–peritumoral tissues shared clones in peritumoral TCR repertoire (TN/N) than those with MVI (p < 0.05, Wilcoxon test; Supplementary Figure 4B ). Furthermore, non-relapse patients tended to have a higher proportion of three tissues shared clones in tumoral TCR repertoire (TNB/T) (p < 0.05, Wilcoxon test; Supplementary Figure 4C ).

Considering that the TCR repertoire in peripheral blood is a systemic reflection of the immune response of tumors (48), we postulated that the features of TCRβ VJ pairings in the peripheral blood may potentially serve as a diagnostic biomarker. We first analyzed the data of peripheral TCR repertoires from 439 healthy controls from Liu’s published dataset (37) and 43 HCC patients in the cohort. PCA based on TCRβ VJ pairing usage profile showed that HCC patients were distinctly different from healthy people, revealing a unique TCR repertoire in HCC patients ( Figure 5A ). The differential analysis identified 90 TCRβ VJ pairings showing significantly different abundances between HCC patients and healthy donors. Among them, 65 TCRβ VJ pairings were more abundant in healthy donors, including TRBV3-1_TRBJ2-1 (healthy vs. HCC = 0.314 ± 0.202 vs. 0.015 ± 0.013) and TRBV12-3_TRBJ2-7 (healthy vs. HCC = 0.270 ± 0.173 vs. 0.019 ± 0.018). On the other hand, 25 TCRβ VJ pairings were enriched in HCC patients, including TRBV6-6_TRBJ2-1 (HCC vs. healthy = 0.755 ± 0.314 vs. 0.048 ± 0.047) and TRBV27_TRBJ1-6 (HCC vs. healthy = 0.325 ± 0.509 vs. 0.085 ± 0.010) ( Figure 5B ).

To establish a robust model for using the peripheral TCRβ VJ pairings as a biomarker for HCC diagnosis, peripheral TCRβ VJ pairings of 43 HCC samples and 439 healthy controls were acquired as the training cohort (80%) and testing cohort (20%) by using random partitioning for HCC diagnosis model construction (experimental section and Figure 5C ). The baseline characteristics of the training cohort and test cohort of HCC patients are shown in Table 1 . We used the LASSO regression for TCRβ VJ pairing for logistic regression to develop a model for HCC rapid diagnosis. HCC patients can be differentiated from healthy people with an AUC of 0.9746 ± 0.0481 (sensitivity = 0.9675 ± 0.0603, specificity = 0.9998 ± 0.0007, average of 100 repeats) in the testing cohort ( Figures 5D , E ). The results above support that TCRβ VJ pairing usage in peripheral blood has strong discriminability to identify HCC patients who were selected in each repeat and may be a highly promising biomarker for HCC diagnosis.