A total of 471 patients with hepatobiliary cancers were screened from two clinical trials (NCT03895970, NCT03892577) at the Peking Union Medical College Hospital. Eligible patients with hepatobiliary cancers who met the inclusion and criteria were recruited to the present study. The main inclusion criteria for patients with hepatobiliary cancer were as follows: (1) pathologically confirmed as hepatobiliary cancer or confirmed by imaging as HCC (by the American Association for the Study of Liver Diseases or standard for the diagnosis and treatment of primary liver cancer 2017 in China)23 24; (2) Eastern Cooperative Oncology Group performance status (ECOG PS) score 0–2 and (3) life expectancy of at least 3 months. Two hundred eighty-four patients were excluded based on the following criteria: did not collect blood sample (n=116), blood unmet cfDNA quality (n=41), inadequate organ function (n=34), lacked follow-up data (n=33), organ transplantation status (n=11), active autoimmune disease (n=8), no measurable disease (n=7), died (n=6), withdrew consent (n=13) and excluded for other reasons (n=15). A total of 187 patients with hepatobiliary cancers were divided into three cohorts according to therapy. ICI cohort 1 consisted of 43 patients with hepatobiliary cancers who received combination therapy of PD-1 inhibitor with lenvatinib. ICI cohort 2 included 108 patients with hepatobiliary cancers who received ICI-based therapy. The non-ICI cohort included 36 patients with hepatobiliary cancers who received non-ICI therapy (figure 1).

The objective response was measured according to the Response Evaluation Criteria in Solid Tumors V.1.1 guidelines,25 and durable clinical benefit (DCB) was defined as complete response (CR), partial response (PR) or stable disease (SD) for ≥24 weeks,26 which were evaluated by professional radiologists at our center who were blinded to the therapeutic outcomes and clinicopathological features. For cohort 1 and cohort 2, PFS was defined as the time from the start of anti-PD-1/PD-L1 treatment to the first documented disease progression or death from any cause. OS was defined as the time between the start of anti-PD-1/PD-L1 treatment and death due to any cause. For non-ICI cohort 3, PFS and OS were evaluated as the time from the start of the systematic therapy.

The plasma cfDNA and blood cell DNA mutation profiles were assessed by a cancer gene-targeted next-generation sequencing (NGS) panel. Methods of DNA extraction, target capture and NGS are provided in the online supplemental material. Algorithm of genomic feature including tumor mutation burden (TMB), copy number instability (CNI) and molecular mutation burden (MMB) are provided in the online supplemental material.

The tumor burden score (TBS) was calculated by the maximum tumor size and number of tumors27 28 using the following formula:

TBS2=Maximuntumordiameter2+Numberoftumors2

A total of 43 patients were randomly allocated to the discovery set (n=30) and validation set (n=13). Univariate Cox regression analysis was performed to analyze the gene copy number value significantly associated with OS. Least absolute shrinkage and selection operator (LASSO) Cox regression analysis was used to determine the coefficient for each feature and estimate the likelihood deviance. The coefficients and partial likelihood deviance were calculated by the ‘glmnet’ package in R. The copy number variation (CNV) risk score was calculated by the following formula:

CNVriskscore=∑i=1nCoefi∗xi

where Coef_i is the risk coefficient of each factor calculated by the LASSO Cox model, and x_i is the copy number value of each factor.

The time-dependent receiver operating characteristic (tROC) curve to detect the predictive power of the risk score was constructed by the ‘survival ROC’ package, and Kaplan-Meier analysis with the log-rank test was carried out by the ‘survival’ package in R. The optimal cut-off value of the CNV risk score was determined by the Youden Index.

The clinicopathological features of the discovery and validation cohorts were compared using the χ² test (two-tailed). OS and PFS were analyzed by multivariate Cox regression and the Kaplan-Meier method, and the log-rank test was used to detect the significant differences between different groups via the ‘survival’ and ‘survminer’ packages in R. The data are expressed as the median and IQR. The Wilcoxon rank-sum test was applied to assess the differences between two groups. Fisher’s exact test was used to analyze the correction between CNV risk score with clinicopathological characteristics. P<0.05 was considered significant for two-sided tests. Statistical analyses were performed using R V.3.6.2.

The detailed clinical characteristics are shown in table 1 and online supplemental table S1. Four hundred seventy-one patients with hepatobiliary cancers were screened, and 284 patients were excluded. Then, a total of 187 patients with hepatobiliary cancers were divided into three cohorts according to therapy. ICI cohort 1 consisted of 43 patients with hepatobiliary cancers who received combination therapy of PD-1 inhibitor with lenvatinib. ICI cohort 1 included 12 patients with HCC, 19 patients with intrahepatic cholangiocarcinoma (ICC), 4 patients with extrahepatic cholangiocarcinoma (ECC), 5 patients with GBC and 3 patients with combined hepatocellular-cholangiocarcinoma (CHCC). The median follow-up was 7.87 months, and 30.2% of 43 patients reached DCB from PD-1 inhibitor plus lenvatinib therapy during the follow-up period. ICI cohort 2 included 108 patients with hepatobiliary cancers who received ICI therapy. ICI cohort 2 included 44 patients with HCC, 39 patients with ICC, 16 patients with ECC, 8 patients with GBC and 1 patient with CHCC. The median follow-up was 11.05 months, and 50% of 108 patients reached DCB from ICI therapy. In the non-ICI cohort, the median follow-up was 5.83 months, and 30.6% of 36 patients reached DCB. Sex, histological type and histological grade were comparable among the three cohorts (p>0.05).

Samples from 187 patients with hepatobiliary cancers were collected. Plasma cfDNA and white blood cell DNA were sequenced by a cancer gene-targeted NGS panel. The mutation profiles of 10 canonical pathways, including the cell cycle, Hippo, Myc, Notch, Nrf2, PI3 kinase/Akt, RTK-RAS, TGFβ signaling, p53 and β-catenin/WNT pathways, were assessed.29 The TP53, RICTOR, NF1, CDKN2A, RB1, FBXW7, NFE2L2, PIK3CA, STK11, ERBB4, KRAS and NRAS genes were frequently mutated in hepatobiliary cancers pretreatment (online supplemental figure S1). The landscape of the CNVs in plasma cfDNA is shown in online supplemental figure S2.

Next, we assessed whether the CNVs in plasma cfDNA were predictive of the response to combination therapy of PD-1 inhibitor with lenvatinib. Forty-three patients were randomly divided into two groups: the discovery cohort (HCC=7, ICC=13, ECC=3 and others=7) and the validation cohort (HCC=5, ICC=6, ECC=1 and GBC=1). Using LASSO Cox regression methods, the genes with CNVs in plasma cfDNA were selected to build the CNV score model to predict survival after combination therapy in the discovery cohort. The formula for the CNV score was based on the copy number of eight genes, including CALR, NR4A3, IDH2, IGF1R, ETV6, STAT3, NF2 and CTCF. The formula for the CNV score was as follows: CNV risk score=(−1.4267831)×CALR +0.5515164×STAT3+1.5124620×IDH2+1.5372432×ETV6+4.1835445×IGF1R+(−1.3164812)×NR4A3+1.1780366×NF2+1.5359307×CTCF (online supplemental figure S3).

To validate the CNV risk score of plasma cfDNA to predict survival after combination therapy, R packages ‘survminer’ were used to generate the optimum cut-off value of the CNV score. In the training cohort, we included 30 patients with a CNV risk score of baseline plasma cfDNA higher than 15.68 (high-risk group) with shorter survival times after combination therapy of PD-1 inhibitor with lenvatinib and those with a CNV risk score of baseline plasma cfDNA lower than 15.68 (low-risk group) with longer survival times after combination therapy (figures 2A,C and 3B). We analyzed the CNV risk score and OS of 13 patients from cohort 1 that were not used as part of the training set. The result showed that the patients with high CNV risk score had median OS of 10.37 months and patients with low CNV risk score did not reach (p=0.11) (online supplemental figure 5A). In addition, we compared the CNV risk score between the DCB and no durable benefit (NDB) groups and the CNV risk score among the PR, PD and SD groups, and the results showed the CNV risk score did not significantly change among those groups of 13 patients (online supplemental figure 5B,C). In the validation cohort, the optimum cut-off of the CNV score was the same as that in the discovery cohort, and the results were similar to those in the discovery cohort (figure 2E,F,G). When the distribution of the CNV risk score and survival status were assessed in the training and validation cohorts, the results showed that patients with lower CNV risk scores had better survival than those with higher CNV risk scores (figure 2B,F). The tROC curves of the CNV risk score in the discovery and validation cohorts are shown in figure 4. In the discovery cohort, the CNV risk score had an area under the curve (AUC) of 0.908 at 3 months, 0.847 at 5 months, 0.850 at 10 months and 0.881 at 15 months (figure 2D). In the validation cohort, the CNV risk score had an AUC of 0.867 at 5 months, 0.851 at 10 months and 0.907 at 15 months (figure 2H).

In ICI cohort 1, Kaplan-Meier curves showed that all patients with hepatobiliary cancer with low CNV risk scores (n=20) had longer OS and PFS than patients with hepatobiliary cancer with high CNV risk scores (n=23) (figure 3A, low vs high CNV risk score, PFS: HR=0.39, log-rank p=0.0095; figure 3B, OS: HR=0.045, log-rank p<0.0001). During the follow-up period, the median OS of patients with hepatobiliary cancer with low CNV risk scores was not reached, and the median OS of patients with hepatobiliary cancer with high CNV risk scores was 6.5 months. The median PFS of patients with hepatobiliary cancer with low and high CNV risk scores was 6.17 months and 2.60 months, respectively.

To validate this finding, we further analyzed the 108 patients with hepatobiliary cancer from independent ICI cohort 2 from our clinical trials. The results were similar to those of ICI cohort 1. Favorable OS and PFS were observed in patients with hepatobiliary cancer with low CNV risk scores (figure 3C, low vs high CNV risk score: median PFS, 6.2 months vs 3.03 months, HR=0.61, PFS log-rank p=0.021; figure 3D, median OS, 20.90 months vs 9.80 months, HR=0.50, OS log-rank p=0.015). In the non-ICI cohort, OS and PFS were not significantly different between patients with low and high CNV risk scores (figure 3E,F, PFS log-rank p=0.32, OS log-rank p=0.62).

To further confirm the effect of the CNV risk score of plasma cfDNA at baseline on the OS of patients with hepatobiliary cancer who received combination therapy, multiple Cox regression was used to assess the effect of multiple factors, including cancer type, PD-L1 expression, ECOG score, Child-Pugh score, alpha-fetoprotein (AFP), TBS, macrovascular invasion, etiology, TNM stage and CNV risk score, on OS. The results showed that the CNV risk score of plasma cfDNA was an independent factor for predicting the OS of patients with hepatobiliary cancer who received combination therapy of PD-1 inhibitor with lenvatinib (figure 4, global log-rank p=0.0068; C-index=0.85; CNV risk score: HR=0.021, p=0.009). Using Fisher’s exact test in ICI cohorts 1 and 2, we found that the CNV risk score (high/low) was associated with histological type (online supplemental table S2, p<0.05), TBS (online supplemental table S2, p<0.05) and maximum tumor diameter (online supplemental table S2, p<0.05), but not with sex, histological grade or macrovascular invasion (online supplemental table S2, p>0.05).

In ICI cohort 1, the CNV risk score of plasma cfDNA was lower in patients showing DCB than in those showing NDB (figure 5A, p=0.042). When the optimum cut-off of the CNV score was used, the result from ICI cohort 1 showed that 53% of patients in the low CNV risk group achieved DCB after combination therapy of PD-1 inhibitor with lenvatinib. In contrast, 12% of patients in the high CNV risk group achieved DCB, and 88% of patients could not benefit from combination therapy of PD-1 inhibitor with lenvatinib (figure 5B, DCB: 53% vs 12%, NDB: 47% vs 88%, p=0.004). In ICI cohort 1, there was a lower tendency of PD and a higher tendency of SD and PR in the low CNV risk group than in the high CNV risk group (figure 5D, PD: 26% vs 52%, SD: 53% vs 38%, PR: 21% vs 10%, p=0.218; disease control rate (DCR), 74% vs 48%, p=0.093). Similar results were observed in ICI cohort 2. The CNV risk score of plasma cfDNA was lower in patients showing DCB than in those showing NDB (figure 5E, p=0.0023). Compared with that in the high CNV risk group, a higher percentage of patients in the low CNV risk group achieved DCB after ICI therapy (figure 5F; DCB: 64% vs 37%, NDB: 36% vs 63%, p=0.005). The CNV risk score was higher in the PD group than in the SD and PR groups in both ICI cohort 1 and ICI cohort 2 (figure 5C,G; ICI cohort 1: PD vs SD, p=0.039, PD vs PR, not significant; ICI cohort 2: PD vs SD, p=0.0067, PD vs PR, p=0.027). In ICI cohort 2, the percentage of PD was significantly lower and the percentages of SD and PR were significantly higher in the low CNV risk group than in the high CNV risk group (figure 5H, PR: 15% vs 20%, SD: 69% vs 47%, PD: 15% vs 33%, p=0.046; DCR: 85% vs 63%, p=0.011).

The molecular features of hepatobiliary cancers with high and low CNV risk scores were investigated. Hepatobiliary cancers with high CNV risk scores were characterized by high TMB and MMB. Moreover, higher CNI scores were observed in hepatobiliary cancers with high CNV risk scores than in hepatobiliary cancers with low CNV risk scores (online supplemental figure S4, p<0.05).