2.1 Baseline demographic and clinical characteristics of the studied population

A total of 293 patients with HCC who underwent LT (hereafter referred to as recipients) and had complete clinical information and available pretransplant blood samples from two independent transplant centers were enrolled in this study. The median follow-up time in the training and validation cohorts was 3.0 and 3.9 years, respectively. The demographic and clinical characteristics of recipients in the training and validation cohorts are shown in Table 1. Both the training (14 females and 176 males) and validation cohorts (11 females and 92 males, p = 0.333) had more than 60% of patients who were over 50 years old (p = 0.700). A total of 96.8 and 98.1% of recipients in the training and validation cohorts had liver cirrhosis (p = 0.542), and hepatitis B virus (HBV) infection rates were 86.8 and 91.3% (p = 0.259), respectively. Recipients with model for end-stage liver disease (MELD) scores over 26 accounted for 57.4 and 8.7% in the training and validation cohorts (p < 0.001). The tumor features did not differ between the training and validation cohorts: 47.4 versus 53.4% recipients had multifocal lesions (p = 0.324), 25.3 versus 31.1% recipients had the max tumor lesion larger than 5 cm (p = 0.287), 26.8 versus 35.0% recipients had the total tumor lesion over 10 cm (p = 0.147). 53.7 and 60.2% of recipients received pretransplant down-stage treatment in the training and validation cohort (p = 0.284). The pretransplant AFP, SII, alanine aminotransferase (ALT), and aspartate aminotransferase (AST) were consistent between the training and validation cohorts, as shown in Table 1. At the same time, the NLR and γ-glutamyltranspeptidase (GGT) were different between the two cohorts. There were 45 and 54 recipients who suffered HCC relapse after LT in the training and validation cohort, respectively.

2.2 Recurrence-related inflammatory biomarkers and clinical variables

Thirty-eight cytokines and six commonly used clinical systemic inflammatory indexes and tumoral biomarkers, including ALT, AST, GGT, SII, NLR and AFP, were detected for the whole study population. Before the COX regression analysis, we tested the proportional hazards (PH) assumption for a Cox regression model fit. The results showed that all the forty-four variables fitted the PHs assumption (p > 0.050; Figure S1). The above results supported our further conducting Cox proportional hazards regression analysis to identify independent risk factors for posttransplant recurrence. We applied univariate Cox proportional hazards regression analysis to screen out recurrence-related inflammatory biomarkers, and the results are shown in Table 2. To enroll all the recurrence-related biomarkers, variables with a p value less than 0.15 were selected for subsequent analysis. Thirteen variables were identified as recurrence-related factors including lnAFP (Hazard Rario [HR]: 1.14, 1.05–1.25; p = 0.003), lnNLR (HR: 1.39, 1.06–1.83; p = 0.018), lnSII (HR: 1.62, 1.29–2.04; p < 0.001), lnGGT (HR: 1.58, 1.20–2.07; p = 0.001), lngp130 (HR: 0.70, 0.46–1.08; p = 0.110), lnIFN-β (HR: 0.69, 0.43–1.11; p = 0.125), lnsIL6Rα (HR: 0.72, 0.46–1.11; p = 0.135), lnIL26 (HR: 1.52, 1.05–2.20; p = 0.025), lnIL32 (HR: 2.13, 1.03–4.40; p = 0.042), lnOsteocalcin (HR: 0.59, 0.47–0.75; p < 0.001), lnsTNFR2 (HR: 1.95, 1.03–3.68; p = 0.040), lnTSLP (HR: 2.08, 0.93–4.67; p = 0.076), lnIL6 (HR: 1.93, 1.40–2.66; p < 0.001). We established a forest plot to present these results (Figure 1B). To gain deeper insight into those recurrence-related biomarkers (n = 10), we mapped the protein–protein interaction network based on the String database. We found that IL6 was the core molecule of the network (Figure 1C), which was in line with our previous findings (unpublished data).

2.3 Inflammatory fingerprint incorporated four biomarkers and predicted HCC recurrence after LT in the training cohort

Multivariate Cox PHs regression analysis identified four independent risk factors including lnAFP (HR: 1.14, 1.04–1.24; p = 0.005), lnSII (HR: 1.40, 1.08–1.81; p = 0.009), lnIL6 (HR: 2.17, 1.51–3.11; p < 0.001), lnOsteocalcin (HR: 0.68, 0.51–0.90; p = 0.013) (Table 2). Their area under the receiver operating characteristic curve (AUROC) was 0.600, 0.686, 0.617, and 0.704, respectively (Figure S2A–D). The Kaplan–Meier (KM) plot presented the survival analysis results of the four inflammatory biomarkers based on the optimal cut-off values (Figure S2E–H). The optimal cut-off values of lnAFP, lnSII, lnIL6, and lnOsteocalcin were 5.7, 5.5, 2.5, and 7.4, respectively. It should be pointed out that all four biomarkers had stable expression levels in the training and validation cohorts (Figure S3A–D).

A digital prognostic model named inflammatory fingerprint (IFP) was developed based on the coefficient value of the inflammatory biomarkers, and the value of the IFP was calculated using the formula: 0.76 × lnIL6 + 0.35 × lnSII + 0.13 × lnAFP − 0.37 × lnOsteocalcin. The optimal cut-off value of IFP determined by receiver operating characteristic (ROC) analysis was 1.7, and the AUROC was 0.792 (Figure 2A); the time-dependent ROC curve showed that the AUROC of the IFP increased over time (Figure 2B); the 1, 2, and 3-year AUROC was demonstrated in Figure S3E separately. Then, we established a predictive nomogram based on the Cox PHs regression analysis (Figure 2C). The overall index of concordance (C-index) was 0.736, indicating satisfactory predictive discrimination of the IFP. Figure 2D showed the calibration curve of the IFP in predicting 3-year RFS in the training cohort, and the Brier score was 0.178 (95% CI, 0.147–0.209) (Figure S4A), indicating good predictive accuracy of the IFP. Subsequently, we divided recipients in the training cohort into the high IFP group (n = 67) and the low IFP group (n = 123). KM analysis indicated that recipients in the high IFP group and low IFP group had a significant difference in RFS (HR: 7.43, 95% CI, 4.28–12.92, p < 0.001; Figure 2E); the 3-year RFS rate was 37.9 and 86.9%, respectively. The results were identical in the analysis using a competing-risk model (Figure S4C). According to our study, the traditional candidate selection criteria, including the Milan criteria, UCSF criteria, and Tokyo criteria, performed well in predicting posttransplant HCC recurrence for recipients in the training cohort. Recipients who fulfilled the criteria had significantly better RFS than those who exceeded the criteria (p < 0.050; Figure S5A,C,E). The 3-year RFS rate was 81.8 versus 53.8%, 77.3 versus 54.1%, and 81.2 versus 47.4% for patients who fulfilled or exceeded the three criteria (Milan criteria, UCSF criteria, and Tokyo criteria), respectively. To further explore the superiority of the IFP, we compared it with the three traditional criteria through the net reclassification improvement (NRI) and the integrated discrimination index (IDI). The NRI was 27.0, 27.0, and 25.3% compared with the Milan, UCSF, and Tokyo criteria. The IDI also supported the superiority of the IFP as it was 10.3, 10.3, and 10.2% when compared with the Milan, UCSF, and Tokyo criteria, respectively (Figure S5B,D,F). The NRI and IDI indicated that the IFP had better predictive capacity in predicting posttransplant HCC recurrence in the training cohort.

Combined analysis showed that the IFP could optimize the recurrence risk stratification based on the traditional criteria in the training cohort. We created new subgroups by combining the criteria with the IFP as follows (Figure 3A): (1) “MC+IFP” ① Group 1: exceeding MC with high IFP (n = 44); ② Group 2: exceeding MC with low IFP (n = 41); ③ Group 3: fulfilling MC with high IFP (n = 23); ④ Group 4: fulfilling MC with low IFP (n = 82). (2) “UCSF+IFP” ① Group 1: exceeding UCSF with high IFP (n = 35); ② Group 2: exceeding UCSF with low IFP (n = 31); ③ Group 3: fulfilling UCSF with high IFP (n = 32); ④ Group 4: fulfilling UCSF with low IFP (n = 92). (3) “Tokyo+IFP” ① Group 1: exceeding Tokyo with high IFP (n = 41); ② Group 2: exceeding Tokyo with low IFP (n = 26); ③ Group 3: fulfilling Tokyo with high IFP (n = 26); ④ Group 4: fulfilling Tokyo with low IFP (n = 97). First of all, we found that nearly half of recipients who exceeded the criteria had low IFP and better RFS, which was close to that of recipients fulfilling the criteria with low IFP. Specifically, recipients in groups 2 and 4 of the “MC+IFP” combination had similar RFS (p > 0.050) and the 3-year RFS rates were 85.4 versus 87.7%, respectively. And recipients in groups 2 and 4 of the “UCSF+IFP” combination also had similar RFS (p > 0.050), and the 3-year RFS rates were 87.1 versus 86.9%, respectively. Likewise, recipients in groups 2 and 4 of the “Tokyo+IFP” combination had similar RFS (p > 0.050), and the 3-year RFS rates were 84.6 versus 87.6%, respectively. To our delight, the IFP could also screen out recipients with high recurrence risk even if they met the criteria. The KM analysis showed that recipients who fulfilled the criteria with a high IFP had poorer RFS, which was close to that of recipients who exceeded the criteria with a high IFP. For instance, recipients in groups 1 and 3 of the “MC+IFP” combination had worse RFS, and the 3-year RFS rates were 26.3 and 60.6%, respectively. And recipients in groups 1 and 3 of the “UCSF+IFP” combination also had worse RFS, and the 3-year RFS rates were 27.5 and 49.8%, respectively. Likewise, recipients in groups 1 and 3 of the “Tokyo+IFP” combination had worse RFS, and the 3-year RFS rates were 25.7 and 57.4% (Figure 3B–D). These results indicated that the IFP could refine the recurrence risk stratification based on the traditional candidate selection criteria.

Additionally, we found that the IFP also performed well in predicting posttransplant overall survival (OS) of patients in the training cohort, with AUROC at 0.761 (Figure S6A). Recipients in the low IFP group had better OS than recipients in the high IFP group (HR: 6.40, 95% CI: 3.50–11.68, p < 0.001; Figure S6B); the 3-year OS rate was 87.9 and 53.8%, respectively.

2.4 Recipients in the high IFP group were characterized by impaired peripheral T cell function

We carried out flow cytometry (FC) to profile the peripheral T cell status in 101 out of the 190 studied recipients in the training cohort (Table S1). Thirty-two subsets of peripheral blood T cells were monitored, and their absolute expression levels (count/µL) were used for further analysis. The consensus cluster analysis identified two clusters characterized by different T cell function statuses (cluster 1: n = 43, cluster 2: n = 58; Figure 4A). Recipients in cluster 2 had higher expression levels of ten T cell subsets, including total CD8⁺ T (log2FC = 1.45, p < 0.001), activated CD8⁺ T (log2FC = 1.17, p < 0.001), effector CD8⁺ T (log2FC = 1.54, p < 0.001), effector memory CD8⁺ T (log2FC = 1.59, p < 0.001), central memory CD8⁺ T (log2FC = 1.57, p < 0.001), activated CD4⁺ T (log2FC = 1.09, p < 0.001), central memory CD4⁺ T (log2FC = 1.47, p < 0.001), primary CD4⁺ T (log2FC = 1.53, p < 0.001), effector memory double negative T (DNT) (log2FC = 1.19, p < 0.001), and primary DNTs (log2FC = 1.07, p < 0.001) than recipients in cluster 1. The results were shown by volcano plot (Figure 4B). As most of the above T cell subsets had activated functional status, we named cluster 1 the “resting” classification and cluster 2 the “activated” classification. Recipients in the activated classification had significantly better RFS than those recipients in the resting classification (HR: 1.96, 95% CI: 1.01–3.85, p = 0.041; Figure 4C). The 3-year RFS rate was 71.3 and 53.5%, respectively. Further analysis found that recipients with high IFP were mainly in the resting classification and provided a primary explanation for the predictive capacity of the IFP (Figure 4D).

2.5 External validation of the IFP in predicting HCC recurrence after LT

A total of 103 patients were included in the validation cohort. The AUROC of the IFP in predicting posttransplant HCC recurrence was 0.807 (Figure 5A), and the time-dependent ROC curve showed that the predictive capacity of the IFP increased over time (Figure 5B). The 1, 2, and 3-year AUROC was shown in Figure S3F separately. The overall C-index was 0.681, indicating satisfactory predictive discrimination of the IFP. Figure 5C showed the calibration curve of the IFP in predicting 3-year RFS in the validation cohort, and the Brier score was 0.198 (95% CI, 0.169–0.227) (Figure S4B), indicating good predictive accuracy of the IFP. Recipients in the validation cohort were divided into the low IFP group (n = 58) and high IFP group (n = 45); survival analysis indicated that recipients in the high IFP group had worse RFS than recipients in the low IFP group (HR:3.61, 95% CI, 2.10–6.30, p < 0.001; Figure 5D); the 3-year recurrence rate was 23.0 and 64.4%, respectively. The combination of the IFP and the aforementioned traditional candidate selection criteria performed well in the recurrence risk stratification of recipients in the validation cohort. Due to the limited sample size of the validation cohort, we divided the population into three subgroups based on the criteria and the IFP as follows (Figure 6A): (1) “MC+IFP” ① Group 1: exceeding MC with high IFP (n = 33); ② Group 2: exceeding MC with low IFP (n = 27); ③ Group 3: fulfilling MC (n = 43). (2) “UCSF+IFP” ① Group 1: exceeding UCSF (n = 47); ② Group 2: fulfilling UCSF with high IFP (n = 16);③ Group 3: fulfilling UCSF with low IFP (n = 40). (3) “Tokyo+IFP” ① Group 1: exceeding Tokyo (n = 46); ② Group 2: fulfilling Tokyo with high IFP (n = 15); ③ Group 3: fulfilling Tokyo with low IFP (n = 42). Nearly half of recipients who exceeded the Milan criteria had low IFP and better RFS, close to that of patients who fulfilled the criteria with low IFP (Figure 6B); the 3-year RFS rates were 55.6 and 67.4%. Recipients fulfilled the UCSF criteria, and the Tokyo criteria with high IFP had poorer RFS, close to that of recipients who exceeded the criteria (Figure 6C,D). The 3-year RFS rate was 37.5 versus 27.7% and 40.0 versus 28.3%, respectively. These results indicated that the IFP could also refine the recurrence risk stratification for recipients in the validation cohort.

In accordance with our previous findings in the training cohort, the IFP also performed well in predicting the posttransplant OS of recipients in the validation cohort with AUROC at 0.750 (Figure S6C). Recipients with low IFP (n = 58) had better OS than recipients with high IFP (n = 45, HR: 3.06, 95% CI: 1.74–5.40, p < 0.001; Figure S6D); the 3-year OS rate was 72.4 versus 40.0%. In addition, we applied propensity score matching (PSM) to the low IFP (n = 182) and high IFP (n = 111) groups in the whole studied population (n = 293). Before matching, the tumor characteristics, including number, single and total diameter, pretransplant AFP levels, and liver functions, were significantly different in the two groups (Table S2). There were 81 recipients in each group after matching, and the characteristics with significant differences were balanced (Table S2). In the 162 matched recipients, the AUROC of the IFP in predicting posttransplant HCC recurrence was 0.812 (Figure S7A), and the C-index was 0.734, indicating satisfactory predictive discrimination of the IFP. Besides, the calibration curve of the IFP in predicting 3-year RFS was shown in Figure S7B, and the Brier score was 0.191 (95% CI, 0.163–0.218; Figure S7C), indicating good predictive accuracy. Recipients in the validation cohort were divided into the high IFP group (n = 81) and low IFP group (n = 81); survival analysis indicated that recipients in the high IFP group had worse RFS than recipients in the low IFP group (HR:5.19, 95% CI, 3.25–8.28, p < 0.001; Figure S7D); the 3-year recurrence rate was 31.9 and 80.1%, respectively.

4.1 Study population

A total of 293 patients underwent LT from donation after cardiac death in two independent medical centers, the First Affiliated Hospital of Zhejiang University School of Medicine and Shulan (Hangzhou) Hospital, from January 2015 to December 2019 were enrolled in this retrospective study. The operations were performed by surgeons who had received identical surgical training and were certified. The inclusion criteria were as follows: (1) a complete preoperative ultrasound, computed tomography (CT), or magnetic resonance imaging (MRI) scan examination of the liver imaging data, (2) a posttransplant histopathology report of HCC on the explant liver, and (3) recipients who were older than 18 years. The exclusion criteria were (1) other types of liver cancer instead of HCC proven by posttransplant pathology, (2) tumor thrombus in the main truck of portal vein reported by pretransplant imaging, (3) OS less than 30 days, (4) patients who underwent secondary LT or multiorgan transplantation, and (5) patients with incomplete medical records or unavailable blood sample. In the current study, two batches of plasma collected from blood samples were sent for cytokine profiling. In detail, the plasma samples collected from January 1, 2018 to December 31, 2019 were set into batch 1 (the training cohort, n = 190) to ensure the data quality. The plasma samples collected from January 1, 2015 to December 31, 2017 were set into batch 2 (the validation cohort, n = 103) to testify the predictive capacity of our prediction model in samples stored for a long period.

Figure 1A showed the schematic illustration of our research. The study was designed in compliance with the STROBE and TRIPOD guidelines.^{50, 51}

4.2 Data collection and follow-up

Clinical data were collected from the CLTR, strictly according to the Regulations on Human Organ Transplant and national legal requirements, including age, gender, body mass index (BMI), MELD score, HBV infection, cirrhotic status, pretransplant downstaging therapy or liver resection (DS/LR), tumor number, the maximum and total diameter of tumor, pretransplant AFP level, blood routine test results, and follow-up information. Among them, cirrhosis status, tumor number, and the maximum and total diameter of the tumor were derived from pretransplant CT or MRI, and the pathological diagnosis of HCC was confirmed by explant histopathology. The grouping of age, BMI, MELD score, and tumor features was in reference to previous studies.^{52, 53} According to previously reported literature, similar pretransplant and posttransplant management strategies were utilized in the two centers.⁵² Patients were divided into different groups by three traditional candidate selection criteria: Milan criteria, UCSF criteria, and Tokyo criteria, as reported previously. Patients were regularly followed up in outpatient clinics, and serum AFP, ultrasound, and CT (every 3 months for the first year, every 6 months for the next year, and then annually) were the main items of the recurrence monitoring. The primary endpoint for this study was death due to any cause or the last follow-up (loss to follow-up).

4.3 Cytokine and T-cell profiling of pretransplant blood sample

4.4 Statistics

For baseline data analysis and presentation, the Kolmogorov–Smirnov test (K–S test) was applied to test the normality of continuous variables, and the variable did not fit a normal distribution if the p value < 0.050. Normally distributed variables were presented as a mean ± standard deviation, and non-normally distributed variables were presented as the medians (interquartile range), and categorical variables were expressed as numbers (percentages). To balance important patient characteristics between groups and reduce selection bias, we implemented stratified and PSM, and the caliper value was set to 0.2. RFS was defined as the time from LT to the tumor return, death led by all causes, or the last follow-up conducted. For differential analysis between the training and validation cohorts, Student's t-test was utilized to analyze normally distributed continuous variables, and the Mann–Whitney test or Wilcoxon rank sum test was used for non-normally distributed continuous variables; Pearson's chi-square test or Fisher's exact test was applied to analyze categorical variables. For subsequent analyses, all concentration data were natural log-transformed and standardized. We carried out univariate and multivariate Cox proportional hazards regression analysis to identify independent prognostic factors in a stepwise mode. The risk model for predicting tumor recurrence after LT was developed based on the coefficients of independent prognostic factors derived from the multivariate Cox proportional hazards regression analysis. The survival curve was analyzed using the KM method, and comparisons between groups were performed using the log-rank test. The ROC curve and concordance index (C-index) were used to evaluate the predictive discrimination of the model. The calibration curve and Brier score were used to present and evaluate the model's predictive accuracy. To ensure the rigor of the method, the PH assumption of the variables was checked before the performance of the multivariate Cox regression analysis. The NRI and the IDI were calculated to compare the model's predictive capacity with traditional candidate selection criteria, including Milan criteria, UCSF criteria, and Tokyo criteria.^{8, 54} Competing risk analysis was performed using the Fine-Gray model to calculate the hazard of recurrence while accounting for the competing outcome of nonrecurrence mortality. The consensus clustering analysis was used to identify function subtypes of T cells.

Statistical analysis was performed using IBM SPSS Statistics 25.0 (SPSS Inc., Chicago, IL), R-project (version 3.6.1, https://www.r-project.org/), GraphPad Prism Software (GraphPad Prism version 9.1 for Windows; GraphPad Software, www.graphpad.com), and Sangerbox tools, a free online platform for data analysis (http://vip.sangerbox.com/). Cluster analysis was performed using the ConsensusClusterPlus package of R, and the optimal number of clusters was determined using the empirical cumulative distribution function plot. The schematic diagram was created with BioRender, a scientific graphic software (www.biorender.com). . A two-sided p value < 0.050 was considered statistically significant.