Predicting Postoperative Liver Dysfunction Based on Blood-Derived MicroRNA Signatures

Study Population

Initially, a discovery cohort of patients was prospectively recruited starting from February 2012. Patients with hepatocellular carcinoma (HCC), cholangiocellular carcinoma (CCC), or metastatic colorectal carcinoma (mCRC) were considered eligible for inclusion. Subsequently, we validated our explorative results in a prospective cohort. Ultimately, we monitored miRNAs of interest in a subset of patients before operation as well as on postoperative day (POD) 1 and POD 5.

The present study was approved by the Institutional Ethics Committee, and written informed consent was obtained.

Definition and Classification of Postoperative LD and Complications

Patients were followed up for a postoperative period of 90 days. Postoperative LD was diagnosed following the criteria issued by the International Study Group of Liver Surgery (ISGLS).13 Accordingly, LD was defined by both abnormal values of serum bilirubin and prothrombin time on or after POD 5. Of note, in case of abnormal preoperative serum bilirubin or prothrombin time, a postoperative deviation and deterioration on 2 subsequent days after POD 5 was identified as postoperative LD. Furthermore, patients reaching normal serum bilirubin or prothrombin time values before POD 5, or who were discharged early because of good clinical performance and hence had no further blood collection, were considered as having “no LD.”

For classification of patients with postoperative morbidity, the outline given by Dindo et al.14 was applied. Death within 90 days after operation was classified as postoperative mortality.15

Measurement of Circulating miRNAs

Meticulous preparation of plasma was performed as described.16

RNA Extraction and Small RNA Sequencing

The miRNeasy Mini Kit (Qiagen, Germany) was applied to isolate total RNA, including small RNAs, from plasma. Small RNA sequencing libraries were generated using the CleanTag Small RNA Library Preparation Kit (TriLink) according to the manufacturer's recommendations. Sequencing was performed on an Illumina NextSeq 500, using single-end reads with 50 cycles (Illumina). Sequencing reads in fastq format were adapter-trimmed and quality-checked (generation of fastQC files). Quality filtered reads (phred > 30) were used for alignment against human mature miRNAs (miRBase version 21) using Bowtie 2. Mapped reads were cross-checked through genome alignments (Bowtie 2, Genome Reference Consortium human genome build 37). Mature miRNA reads were counted (isomiR sequences were summarized) and normalized as counts per million (CPM) to the total number of mapped reads. CPM values were used for statistical analysis (see below). Sequencing raw and normalized data can be accessed under the record GSE123605 from the National Center for Biotechnology Information Gene Expression Omnibus (GEO).

qPCR Analysis

qPCR analyses were performed as described.17 Robustness of RNA extraction, complementary DNA (cDNA) synthesis, and qPCR amplification was assessed using combinations of synthetic spike-in controls (Exiqon, Denmark). Hemolysis was assessed using the ratios of miRNA-23a-3p and miRNA-451a.18 Of note, no samples failed because of hemolysis or high analytical variance.

Statistical Analyses

Statistical analysis was based on nonparametric tests. Calculations were performed using SPSS (version 23.0; IBM Corp., Armonk, NY) and R (version 3.4.1; R Foundation, Vienna, Austria). P values were adjusted for multiple testing based on the false discovery rate according to the Benjamini-Hochberg method, and P values < 0.05 were considered significant. For more detail on statistical methods, refer to the Supporting Materials and Methods section.

Patients and Cohorts

A total of 48 patients with mCRC, HCC, or CCC who underwent liver resection between February 2012 and April 2016 were included in the discovery cohort. To achieve a representative cohort, 21 patients suffering from postoperative LD were matched based on basic characteristics, liver function, and extent of liver resection to 27 patients without postoperative LD (Supporting Table S1). Subsequently, an additional 98 patients served as a prospective validation cohort (Table 1). Additionally, the perioperative dynamics of miRNAs were evaluated in a group of 7 patients undergoing regular major liver resection and a group of 8 patients undergoing ALPPS (Supporting Table S2) for whom longitudinal measurements of miRNAs could be performed on the basis of repeatedly collected plasma samples.

Establishing the Predictive miRNA Panel for LD After Operation in the Discovery Cohort

For initial identification of miRNAs associated with postoperative LD, we aimed for an unbiased systematic approach using NGS. miRNAs, small nuclear RNAs, small nucleolar RNAs, transfer RNAs, and Y RNAs were identified in the data set (see Supporting Fig. S1; NGS data were submitted to GEO and can be downloaded using the accession number GSE123605). Subsequently, we focused on miRNA, given the above-mentioned benefits, and analyzed miRNA profiles in the plasma of all patients within the discovery cohort before operation to determine if miRNA profiles differ between patients who develop LD and those without delayed hepatic recovery. With an average of 3 million miRNA reads per sample, we detected a total of 554 miRNAs across all analyzed plasma samples. To identify potential biomarker candidates, cutoffs for plasma abundance (average log₂ CPM [logCPM] > 5), effect size (fold change > 1.3), and significance level (P < 0.2) were implied. As depicted in Fig. 1A, a set of 19 miRNAs remained in the analysis, of which 12 were up-regulated (red) and 7 were down-regulated (blue) in the pre-OP plasma of patients with LD. A clustered expression heat map, comparing expression of these miRNAs in patients with and without LD, is depicted in Fig. 1B.

Subsequently, we evaluated the analytical variability between NGS and qPCR-based miRNA detection and calculated the relative logarithmic differences between two miRNAs to form self-normalizing miRNA pairs, thus circumventing the need for a reference miRNA. Six of the top miRNAs, and consequently 15 miRNA pairs, were considered for this analysis. We observed high concordance between data sets (NGS vs. qPCR-based miRNAs) as illustrated in Supporting Fig. S2. The importance of miRNA pairs for achieving excellent predictive performance of negative postoperative outcomes was analyzed using random forest modelling (Fig. 2A). Two top-ranked miRNA pairs (151a-5p_192-5p, 122-5p_151a-5p) were identified with significant pre-OP differences between individuals developing LD and controls (Fig. 2B). The diagnostic performance of a multivariate model as measured by receiver operator characteristic (ROC) analysis estimated an area under the curve (AUC) of 0.66 for miRNA pair 122-5p/151a-5p (Fig. 2C), 0.75 for miRNA pair 151a-5p/192-5p (Fig. 2D), and 0.76 for a logistic regression (LR) model using the combination of both miRNA pairs (Fig. 2E).

Next, two clinically useful cutoffs were defined to predict postoperative LD. Accordingly, a low-stringency cutoff to identify patients who could undergo liver resection with very low risk (cut-off P > 0.59) as well as a stringent cutoff to identify patients who should be optimized before surgery or not undergo liver resection (cut-off P > 0.68) were defined (see Supporting Table S3). For both cutoffs, predictions were found to be significantly associated with postoperative LD (P = 0.00045 and P = 0.00054, respectively; Fig. 2F).

Validation of the Discovered miRNA Ratios in an Independent Validation Cohort

To confirm the clinical utility of our miRNA ratios, we next validated the predictive performance of the two miRNA pairs in an independent prospective validation cohort consisting of 98 patients, 90 without and 8 with postoperative LD, reflecting the natural postsurgical LD incidence of approximately 10%-20%. The diagnostic performance of a multivariate model as measured by ROC analysis showed an AUC of 0.69 for the two miRNA pairs (Fig. 3A). Further, we validated that the two cutoffs were able to predict postoperative LD for cut-off P > 0.59 with P = 0.011 (Fig. 3A) and for cut-off P > 0.68 with P = 0.032 (Fig. 3A), respectively.

Comparison of the Discovered miRNA Ratios to Other Predictors for Postoperative LD

Next, we evaluated the performance of the miRNA-based prediction model in the combined data set (N = 146) to increase the power of our analyses. First, to assess independence of the identified miRNA pairs to predict postoperative LD from other known markers of liver damage and function, as well as from clinical parameters, multivariate LR analysis was performed for both pairs independently. Strikingly, both miRNA pairs remained significant on multivariate LR analysis (Table 2).

Further, the diagnostic performance of the two cutoffs were illustrated using sensitivity (SN), SP, positive predictive value (PPV), negative predictive value (NPV), and the odds ratio (OR). The low-stringency cutoff (>0.59) yielded PPV and NPV values of 0.70 and 0.89 (Fig. 3B), respectively, whereas the stringent cutoff (>0.68) resulted in a PPV of 0.83 with an NPV of 0.85 (Fig. 3B), indicating that 83% of patients who tested positive suffered from postoperative LD, whereas 85% of patients who tested negative did not suffer from postoperative LD. The ORs for an adverse event were 18.66 (P < 0.0001) and 19.7 (P < 0.0001) for the two cutoffs, respectively. ROC curve analysis was performed for the miRNA model and compared with ROC curves of standard liver function parameters (Fig. 3B). We observed an AUC of 0.77 for the miRNA model, which exceeded the AUC of other parameters, including indocyanine green (ICG) plasma disappearance and retention rates as well as standard blood parameters like alanine transaminase (ALT), aspartate transaminase (AST), and gamma-glutamyl transpeptidase (GGT). Further, we compared SN, SP, NPV, and PPV in a subset of patients who also had preoperative volumetric analyses (N = 33). miRNA pairs were found to identify patients subsequently suffering from postoperative LD with a higher predictive potential as compared with ICG clearance, but also exceeded the predictive potential of volumetric analyses (cutoff of 30% future liver remnant), as illustrated in Supporting Table S4. Finally, ORs for other adverse postoperative outcomes were analyzed for both cut-off models. ORs for severe morbidity reached significance for both cutoffs (Fig. 3C). Although OR for mortality was found to be significant for the low-stringency cutoff, the high-stringency cutoff was not significant but showed a comparable tendency (Fig. 3D). Further, patients fulfilling our cutoffs were found to stay significantly longer on the intensive care unit (ICU) and remained hospitalized for a prolonged time (Fig. 3E,F).

miRNA Pairs Represent Significant Predictors of Postoperative Outcome in the Respective Tumor Types

As the underlying tumor entity is a known confounding factor for the analysis of postoperative outcomes in this patient cohort, further subgroup analyses were computed. Accordingly, patients were divided into the three largest subgroups—mCRC (n = 56), HCC (n = 46), and CCC (n = 41)—and were analyzed separately (Fig. 4A-C). Indeed, the preoperative miRNA signature was found to differ significantly between patients with and without postoperative LD for each tumor type. Further, ROC analysis revealed comparably good predictive potentials. Lastly, evaluation of the proposed cutoffs for prediction of LD showed a strong predictive potential, irrespective of tumor entity. (Fig. 4A-C).

Discovered miRNA Ratios Allow Dynamical Monitoring of Liver Function Recovery After Liver Resection and Might Enable Identification of an Optimized Time Point for Operation

Next, we aimed to determine dynamical changes of miRNA pairs after operation to evaluate their association with liver function. Accordingly, miRNA ratios were assessed preoperatively (pre-OP) as well as on POD 1 and POD 5 in a matched cohort of patients with regular major liver resection (n = 7) and patients undergoing the ALPPS procedure (n = 8, see details on the procedure in Fig. 5A and below). Basic characteristics are illustrated in Supporting Table S2. Although miRNA pairs as well as the combined LD probability changed significantly after liver resection (POD 1 vs. pre-OP) in all patients to a comparable level, most of them recovered until POD 5 in parallel with a regular liver function recovery.

As timing of the second stage of the ALPPS procedure is still a challenge for clinicians, the usefulness of miRNA pairs for this task was evaluated. Interestingly, patients undergoing the ALPPS procedure were found to display unaltered miRNA ratios after the second step of the procedure (regeneration grossly completed), whereas they behaved similarly to patients undergoing regular liver resection after the first step of the procedure (Fig. 5C). Of note, all patients who developed postoperative LD after the second step of the ALPPS procedure showed a worsened pattern of miRNA profile when compared with the remaining patients (P = 0.054, Fig. 5D). More importantly, the 2 patients with highest detected miRNA ratios before the second operation died after the second step of the procedure (Fig. 5D), indicating that the miRNA ratios are indeed predictive of functional liver recovery.