25 liver cirrhotic patients (LC group) and 122 HCC patients including 44 AFP negative HCC patients (NEG group), 78 HCC patients with AFP over 20 ng/ml (POS group) in the First Affiliated Hospital of Zhejiang University School of Medicine from April 2012 to December 2016 were enrolled in the study. All Patients in the LC and HCC group underwent liver transplantation and were diagnosed according to post-transplant pathological examination. The exclusion criteria included patients younger than 18 years, undergoing multiorgan transplantation or re-transplantation, or with missing essential data for analysis. Another cohort of 52 healthy control samples (HC group) collected from the same batch of individuals who underwent healthy examination. We collected the data including demographics, body mass index (BMI), pre- operative AFP level, alanine transaminase (ALT) level, aspartate transaminase (AST) level, morphological features (tumor number and largest tumor size), skeletal muscle index [SMI, to define sarcopenia (12)], neutrophil-lymphocyte ratio (NLR), post-transplant recurrence, and patients’ survival for analysis. Informed consent was obtained from all the participants, and the study protocol was approved by the Human Ethics Committee of the hospital.

Peripheral blood samples (EDTA-K2 anticoagulant) were collected from fasted patients or healthy volunteers in the morning of LT or healthy examination, and centrifuged at 3000 rpm for 10 min, then stored the plasma at −80°C, until use. The plasma samples were thawed at 4°C, and the quality control (QC) samples were prepared by pooling aliquots (10 μl) of each sample. Acetonitrile (800 μl) was added to the plasma (200 μl) sample and vortexed for 1 min. We then incubated the mixture at room temperature for 1 min and centrifuged it at 14000 rpm for 10 min at 4°C. The acquired clear supernatant was transferred to UPLC vials, and was then stored at 4°C until detection. The pretreatment of the QC samples was the same as that for the test samples.

We performed reversed-phase analysis on a Waters ACQUITY Ultra Performance LC system using an ACQUITY UPLC BEH C18 analytical column (i.d., 2.1 mm × 100 mm; particle size 1.7 mm; pore size, 130 Å). We then used water/formic acid (99.9:0.1 v/v) as mobile phase A and acetonitrile/formic acid (99.9:0.1 v/v) as mobile phase B. A linear gradient LC system (Waters, Milford MA) was optimized as follows: the composition of mobile phase B was changed from 3% to 80% in 7 min, reached 98% in 8 min and held for 5 min, and then reached 100% in 1 min and held for 3 min. The sample manager was kept at 4°C, with an injection volume of 2 μl for each analysis. The QC samples were injected at regular intervals (every 14 samples) throughout the analytical run. These inserted QC samples were used to evaluate the repeatability of sample pretreatment and monitor the stability of the LC–MS system during sequence analysis.

We used a Waters Q-TOF Premier mass spectrometer to perform the mass spectrometry in positive ion electrospray mode. The instrumental parameters were set as follows: The mass scan range was 50 m/z–1000 m/z using an accumulation time of 0.2 s per spectrum; the MS acquisition rate was set to 0.3 s with a 0.02 s inter scan delay; high-purity nitrogen was used as nebulizer and drying gas. The nitrogen drying gas was at a constant flow rate of 600 L/h, and the source temperature was set at 120°C. For the positive mode, the capillary voltage was set at 3.0 kV and the sampling cone voltage was set at 45.0 V. Argon was used as collision gas. MS/MS analysis was performed on the mass spectrometer set at different collision energies of 10 eV–50 eV according to the stability of each metabolite. The time of flight analyzer was used in V mode and tuned for maximum resolution (>10,000 resolving power at m/z 556.2771). The instrument was previously calibrated with sodium formate; the lock mass spray for precise mass determination was set by leucine enkephalin at 556.2771 m/z with concentration of 0.5 ng/L in the positive ion mode. All analyses were acquired using the lock spray to ensure accuracy and reproducibility.

We referred to our previously published metabolomic data (13). The dataset was generated based on the retention time, m/z, and normalized signal intensity of the peaks. The preprocessed data obtained by MassLynx were exported and analyzed using SIMCA-P 14.1 (Umetrics AB, Sweden). Firstly, principal component analysis (PCA) was introduced to evaluate the reliability of the resulting dataset (including QC samples). Secondly, supervised orthogonal partial least squares discriminant analysis (OPLS-DA) was performed to better distinguishing the two groups. Potential biomarkers of differentiating AFP negative HCC patients from LC and HC groups were selected according to the Variable Importance in the Projection (VIP) values, fold change (FC), and Wilcoxon Test. Statistical analysis including logistic regression and cox regression was performed using SPSS version 25.0 statistical software (SPSS inc. Chicago, IL, USA) and GraphPad Prism version 9 (GraphPad, La Jolla, CA, USA). Random forest analysis and nomogram construction were performed by R Version 3.6.1. Area under the time-dependent receiver operating characteristic curve (AUROC) were used to evaluate discriminative ability. The AUROC difference is performed using DeLong’s test. The Hosmer-Lemeshow (HL) goodness-of-fit test was used to assess the calibration of the model. Mann-Whitney U test were used to compare the Metabolite-Score between different groups. Kaplan-Meier analysis and Breslow test were used to compare the survival between groups. P < 0.05 was considered statistically significant throughout the study.

147 patients included in this study underwent LT for HCC or cirrhosis treatment, and 52 healthy volunteers were also enrolled. Of all the patients with different liver diseases, 132 were male (89.8%) and 15 were female (10.2%), while 14 were male (26.9%) and 38 were female (73.1%) in healthy controls. The mean age in LC group and NEG group was 47.9 ± 9.8 and 53.2 ± 8.8 years, respectively (P=0.021). This could be explained by the fact that cirrhosis is an intermediate process of chronic hepatic disease developing to HCC. The AFP level in these two groups was 79.9 ± 275.4 and 8.4 ± 5.5 ng/mL, respectively (P=0.170). The difference of liver functions (including ALT and AST) was not significant between the two groups. Baseline features of all the study subjects including HC group and POS group were listed in Table 1 , and particular features of tumor patients are listed in Supplementary Table S1 .

Metabolomic profiling was performed on the plasma of 52 healthy volunteers and 147 patients with HCC or cirrhosis, and general workflow of this study is listed as Figure 1A . The total ion chromatograms of a single sample from each group were acquired by the UPLC-MS platform. Using MZmine ver. 2.0 software, this pre-treatment revealed 1242 integral peaks following extraction ion chromatography detection in all samples, which was reported in our previous work (13). PCA plot (R2X=0.631, Q2 = 0.421) showed that QC sample cluster together, indicating the high stability and reproducibility of the instrument ( Supplementary Figure S1 ). Besides, HC group showed an obvious separation from NEG HCC group and LC group, while NEG HCC group was roughly separated from LC group ( Figure 1B ).

In order to identify metabolomic markers for AFP negative HCC, pair-wise comparisons were performed among HC, LC and NEG group based on OPLS-DA models ( Figures 1C–E ) and Wilcoxon Test. Validation of the OPLS-DA model of LC and NEG group was obtained from 200 permutation tests ( Supplementary Figures S2A–C ). The validation plot demonstrated that the original model was valid: the Q2 regression line had a negative intercept, and the intercepts of R2 were lower than the original point to the right. S-plots of these OPLS-DA models were further investigated to acquire the correlation value of metabolites ( Supplementary Figures S2D–F ). Ions with a variable importance value (VIP) >1, fold change (FC) >1.5 and P<0.01 were selected. Thus, 116 overlapping ions were selected for further identification ( Figure 2A ). By excluding those ions with over one third cases of ‘0’ value, 15 metabolites including MG (monoacylglyceride), PC (phosphatidylcholine), DG (diglyceride) and SM (sphingomyelin) were finally selected ( Table 2 and Supplementary Table S2 ), and their VIP values and correlation values were also listed. In comparison to LC group, 4 metabolites (Chenodeoxycholic acid glycine conjugate, MG(18:2/0:0/0:0), 1-Oleoylglycerophosphoserine, PC(16:0/16:0)) were significantly decreased in NEG group, whereas 11 metabolites (DG(9M5/9M5/0:0), PC(22:6/16:0), SM(d18:1/18:1), LysoPC(17:0), LysoPC(16:0), PC(22:6/18:2), PC(18:2/18:2), 3-Methoxybenzenepropanoic acid, PC(18:2/20:4), 3-Carboxy-4-methyl-5-propyl-2-furanpropionic acid, PC(14:0/20:4)) were significantly elevated ( Figure 2B ).

The biological pathways involved in the metabolism of these 15 differential metabolites were determined by enrichment analysis using MetaboAnalyst. All matched pathways were shown according to p values from the pathway enrichment analysis (y-axis) and pathway impact values from pathway topology analysis (x-axis) (14), with the most impacted pathways colored in red. One pathway was considered specifically related to AFP negative HCC, that is, glycerophospholipid metabolism ( Figure 2C ).

Random forest (RF) analysis was further used to discriminate AFP negative HCC patient from liver cirrhosis patients based on 15-metabolites panel, which showed relatively low error rate of 25.49% in the training set. Moreover, the prediction of validation data based on training set RF models also yielded satisfactory results with error rate of 14.28% for LC vs. NEG. In order to identify potential biomarkers for AFP negative HCC, the top 7 ranked differential metabolites in the respective models were selected according to the mean decrease accuracy (MDA), which denoted the percent decrease in accuracy when the trial was performed in the absence of the metabolite ( Figure 3A ). The PCoA plot also showed these two groups of samples could almost cluster separately ( Figure 3B ). Subsequent Logistic regression analysis showed that PC(16:0/16:0), PC(18:2/18:2) and SM(d18:1/18:1) were independent risk factors distinguishing AFP negative HCC from liver cirrhosis patients ( Supplementary Table S3 ). Thus, a three-marker model was constructed: Metabolites-Score = -0.071* PC(16:0/16:0) + 0.038* PC(18:2/18:2) + 0.293* SM(d18:1/18:1)-0.553.The ROC curve for the three-marker model was then constructed and a nomogram was then established as well ( Figures 3C, D ). The model showed good discrimination (AUROC=0.913, 95%CI 0.848-0.977, P<0.001) and calibration (HL P=0.739). According to the model, AFP negative patients but with Metabolites-Score more than 1.2895 could be regarded as having a high risk of HCC. The sensitivity and specificity for the model were 0.727 and 0.92, respectively.

We further validated our three-marker model in all patients with HCC or cirrhosis. The diagnostic value of this model was assessed, showing a good discrimination (AUROC=0.912, 95%CI 0.857-0.967, P<0.001) and calibration (HL P=0.645). The cut-off value of Metabolites-Score was also set at 1.2895 with a sensitivity of 0.713 and a specificity of 0.92. To compare the diagnostic performance between our model and AFP, we also performed ROC analysis for AFP and the AUROC was 0.812 (95%CI 0.716-0.909, P<0.001). When the cut-off value was set at 3.7 ng/ml, the sensitivity and specificity were 0.91 and 0.6, respectively. Though the AUROC of our three-marker model was higher than that of AFP, the difference was not significant between them (ΔAUROC=0.1, P=0.13). By combining AFP with three-marker model, we are able to achieve a higher accuracy for diagnosis with an AUROC of 0.951 (95%CI 0.917-0.986, P<0.001, Figure 4 ) and a HL P value of 0.216. The diagnostic performance of the combination model was significantly better than three-marker model (ΔAUROC=0.039, P=0.006) or AFP along (ΔAUROC=0.139, P=0.014), with a positive predictive value of 0.981 and a negative predictive value of 0.575 ( Supplementary Table S4 ).

We further explore the relationship between Metabolites-Score and clinical parameters, and 122 HCC patients were enrolled. We stratified all HCC patients into two groups according to AFP level (≤400 ng/mL and >400ng/mL), tumor number (single and multiple) and largest tumor size (≤5cm and >5cm), though no statistically significant difference in Metabolites-Score were found between any two groups (P>0.05, Figures 5A–C ). In addition, we analyze the relationship between Metabolites-Score and body nutrition status in all HCC patients. In overweight patients group (BMI≥24kg/m²), the Metabolites-Score was higher than that in normal weight patients group (BMI<24kg/m², 3.81 ± 3.13 vs. 2.99 ± 2.13, P=0.243, Figure 5D ). In sarcopenic patient group, the Metabolites-Score was lower than that in non-sarcopenic patient group (2.56 ± 2.11 vs. 3.46 ± 2.60, P=0.155, Figure 5E ). NLR, which represents patient immune status, was also included in the study. Patients with a NLR over 5 had significantly lower Metabolites-Score than patients with a NLR below 5 (2.14 ± 1.88 vs. 3.56 ± 2.60, P=0.012, Figure 5F ).

After excluding the patients who died within two months, 42 AFP negative HCC patients were enrolled for prognostic analysis. 17 patients died during follow-up, with 1-, 3-, and 5-year overall survival (OS) rates of 92.9%, 62.8% and 52.2%, respectively. 18 patients were diagnosed with tumor recurrence during follow-up, with 1-, 3-, and 5-year tumor-free survival (TFS) rates of 66.5%, 58.9% and 54.7%, respectively. According to the univariable Cox regression analysis, MG(18:2/0:0/0:0) was the only metabolite having a moderate prediction capability for TFS (HR=1.160, 95%CI 1.012-1.330, P=0.033, Table 3 ). Based on the normalized peak intensity of MG(18:2/0:0/0:0), the patients were divided into low risk group (n=30) and high risk group (n=12). TFS and OS was significantly different between the two groups (P<0.05, Figures 6A, B ), especially in early survival. We also validated the prognostic value of MG(18:2/0:0/0:0) in AFP positive HCC patients, but it showed no difference between low risk group (n=55) and high risk group (n=19, Supplementary Figures S3A, B ).