This work represents an ancillary study of the prospective, single-center EdMonHG cohort (ClinicalTrials.gov identifier NCT03995537) [11]. The inclusion criteria were adult patients awaiting LT due to compensated Advanced Chronic Liver Disease (cACLD), decompensated ACLD (including acute decompensation (AD), non-acute chronic decompensation (N-AD) and acute-on-chronic liver failure (ACLF)), or acute liver failure (ALF). Exclusion criteria were ongoing immunosuppressive treatment (except corticosteroids), patients without underlying liver disease, and patients awaiting retransplantation or multiorgan transplantation.

We systematically collected recipient demographic and clinical variables (age, sex, liver disease etiology, clinical presentation stage (cACLD, N-AD, AD, ACLF, ALF), clinical severity scores routinely used to assess pre-transplant disease severity in LT candidates: Model for End-stage Liver Disease (MELD) score, Sequential Organ Failure Assessment (SOFA) score, ACLF grade and number of organ failures according to the CLIF-ACLF definition, and hospitalization status at the time of LT. We also collected donor variables (age, sex, body mass index, donation type, and cause of death), and perioperative/surgical variables (intraoperative packed red blood cell transfusion requirements, warm and cold ischemia times, and hepaticojejunostomy).

In line with the EdMonHG biobanking protocol, immune monitoring was performed every 3 months in cACLD and N-AD patients, and during admission for AD or ACLF patients. As a result, the median interval between the pre-LT sampling and transplantation was 9 days [IQR: 1–36]. T lymphocyte count and mHLA-DR expression were assessed in parallel at the same time point, using flow cytometry as previously described in the EdMonHG cohort [11].

For transcriptomic analyses, blood samples were drawn into PAXgene tubes for optimal RNA stabilization. After 2 h at room temperature, the samples were frozen at −80 °C following the manufacturer’s recommendations. We performed targeted transcriptomic analysis using RT-qPCR on a panel of 26 genes related to inflammation and immune response, the Immune Profiling Panel (IPP prototype, bioMérieux, Marcy l’Etoile, France). Of note, these 26 genes were selected based on previous work [10, 12]. This panel covered genes coding for soluble markers of the inflammatory response (pro- and anti-inflammatory), membrane markers associated with innate and adaptive responses, alarmins and genes involved in the regulation of cell cycle or metabolism (Supplementary Figure 1). The mRNA expression level was determined using the FilmArray Torch System (bioMérieux) by injecting 100 µL of the PAXgene whole blood into the IPP prototype pouch. The gene expression data were normalized using reference housekeeping genes from the original publication detailing the approach [13]. The selection of these reference genes was based on previous studies conducted by our team and other groups [14–16].

The primary outcomes were the occurrence of significant early post-LT infections (within the first month following LT) and one-year survival post-LT. Post-operative infections were defined according to the American Society of Transplantation [17] (see definitions in Supplementary Materials) and only significant infections were recorded (excluding uncomplicated cystitis, colitis and catheter colonization). Two independent experts confirmed the cases of post-LT infections a posteriori, before performing transcriptomic assays.

Post-LT outcome variables, including early allograft dysfunction, biliary and vascular complications, and prolonged postoperative corticosteroid use, were also collected.

The results were expressed as medians and interquartile ranges (IQRs) or as numbers and percentages (%). All 97 patients with available pre-transplant transcriptomic data were included; there was no missing data for the primary outcomes or transcriptomic variables.

To explore the overall structure of immune dysregulation, principal component analysis (PCA) was performed on the normalized expression values of all 26 transcriptomic markers, after centering and scaling each gene to unit variance (i.e., PCA was computed on the correlation rather than the covariance matrix), so that no individual gene’s raw variance could disproportionately drive the principal components. PCA scores were compared across disease severity stages and according to post-LT outcomes using the Kruskal-Wallis and Wilcoxon rank-sum tests, respectively. Gene loadings on PC1 were examined to identify the main contributors to immune variance. Candidate biomarkers were identified through a two-step approach. For each gene included in the transcriptomic panel, normalized expression levels were compared between groups using Wilcoxon rank-sum tests for binary outcomes (post-LT infection and one-year survival) and Kruskal–Wallis tests for comparisons across liver disease severity stages and liver disease etiologies. To account for multiple testing, p-values were adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure, and both nominal p-values and adjusted q-values were reported. In parallel, receiver operating characteristic (ROC) analyses were performed and areas under the curve (AUCs) with 95% confidence intervals were calculated for each marker. Volcano plots summarized these analyses by representing the median difference in normalized expression on the x-axis and the −log10(p-value) on the y-axis, with dot size proportional to the AUC. Genes fulfilling both q < 0.10 and AUC > 0.65 criteria were considered candidate biomarkers, highlighted in color on volcano plots, and selected for further analyses including boxplots of normalized expression, threshold-based survival analyses using Kaplan–Meier estimates with cohort median cut-offs, and calibration analyses. An AUC threshold of 0.65 was selected to enrich for markers with at least moderate discriminative ability in this exploratory study. Differences between groups in survival analyses were assessed using the log-rank test. Calibration plots for individual logistic regression models were generated to evaluate the predictive performance of candidate biomarkers. Calibration was assessed using the Hosmer–Lemeshow goodness-of-fit test and generalized additive model (GAM) smoothing, in order to assess agreement between predicted probabilities and observed outcome rates.

Pre-transplant mHLA-DR expression and lymphocyte counts were compared across disease severity stages alongside transcriptomic markers, using the Kruskal-Wallis test, for contextual reference. As their associations with post-LT outcomes have been previously reported in this cohort [11], these markers were not included in the outcome analyses. Given that this was a single-center study with a limited number of outcome events, particularly for one-year mortality (n = 7), the analysis remained exploratory and no formal multivariable analysis could be performed.

R version 4.0.2 (R Core Team 2020, Vienna, Austria) and GraphPad Prism 6.0 (GraphPad Software, La Jolla California, USA) were used for all analyses. The significance level was set at p < 0.05 or q < 0.10 for the Benjamini-Hochberg test.

This study was conducted in accordance with the principles of the Declaration of Helsinki and approved by Institutional Review Board (Comité de Protection des Personnes Ile de France XI, approval number 19039-40433). Written informed consent was obtained from all participants prior to enrollment. This study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines. This study was registered in the ClinicalTrials.gov registry (NCT03995537, date: June 20, 2019).

A total of 130 patients were initially included in the EdMonHG cohort between February 2020 and May 2023. Among these patients, 99 underwent liver transplantation (LT), and 97 of these patients had available transcriptomic data and were therefore included in the current ancillary study.

LT recipients were predominantly male (n = 78, 80%), with a median age of 56 years [48–61]. At pre-LT, 73 patients had decompensated ACLD (23 with N-AD, 21 with AD, and 29 with ACLF), 20 patients had cACLD, and 4 patients were admitted for ALF. The most common underlying liver disease among ACLD patients was alcohol-related liver disease (ALD, n = 68, 73%), followed by viral infection (n = 11, 12% - see details in Supplementary Materials), autoimmunity (n = 8, 9%), metabolic dysfunction-associated steatohepatitis (MASH) (n = 3, 3%), Wilson disease (n = 2, 2%), and progressive familial intrahepatic cholestasis (n = 1, 1%). The etiologies of ALF were acute HBV infection (n = 2), autoimmune hepatitis (n = 1) and posttraumatic ischemia (n = 1). Only three patients received pre-transplant corticosteroids. Among the ACLF patients, the median number of organ failures (OF) pre-LT was 2 [1–3], 7 (24%) patients presented with grade 1 ACLF, 16 (55%) with grade 2 ACLF and 6 patients (21%) with grade 3 ACLF. Coagulation failure (20/29), liver failure (19/29), and hemodynamic failure (10/29) were the most common organ failures. The median MELD score in the entire cohort pre-LT was 20 [14–28] and the median interval between pre-LT immune monitoring and LT was 9 days [1–36]. At the time of LT, 20 patients (20.6%) were hospitalized in the ICU, 12 (12.4%) in a hepatology ward, and 65 (67.0%) were outpatients.

Regarding donor and graft characteristics, donors were predominantly male (n = 62, 64%) with a median age of 66 years [51–72] and a median BMI of 25 kg/m² [23–28]. Causes of donor death were cerebrovascular (n = 48, 49%), traumatic (n = 17, 18%), anoxic (n = 30, 31%) and meningitis (n = 2, 2%). Of the total cohort, 17.5% (n = 17) were classified as Donation after Circulatory Death (DCD) donors. Median warm ischemia time was 37 min [27–51] and median cold ischemia time was 395 min [323–469]. Biliodigestive anastomosis was performed in 9 patients (9%).

Regarding perioperative and early post-transplant data, the median number of packed red blood cell units transfused intraoperatively was 2 [0–5]. Early allograft dysfunction (EAD) according to Olthoff criteria (see details of definition in Supplementary Materials) [18] occurred in 30 patients (31%). Biliary complications were observed in 14 patients (14%), vascular complications in 13 patients (13%), and prolonged corticosteroid therapy (>7 days) was required in 12 patients (12%).

To characterize the overall structure of immune gene expression in our cohort, we first performed PCA on the normalized expression values of all 26 transcriptomic markers. The first two principal components explained 43.5% and 9.5% of total variance, respectively (Supplementary Figure 2). PC1 loading analysis identified two gene clusters with opposing contributions. Genes contributing positively to PC1 (loadings >0.75) included markers of monocyte antigen presentation (CIITA, TAP2, CX3CR1), lymphocyte signaling (ZAP70, IL7R, CD3D), cytotoxic function (GNLY) and cell cycle regulation (CCNB1IP1). Conversely, genes contributing negatively to PC1 (loadings < -0.75) were predominantly associated with innate myeloid inflammatory and immunoregulatory responses, including the decoy interleukin-1 receptor IL1R2, a marker of suppressive myeloid activation, and the alarmin S100A9. PC2 explained a smaller proportion of total variance and was primarily driven by genes involved in innate inflammatory regulation and immune checkpoint signaling, including CD274 (PD-L1), C3AR1, and IL1RN (loadings >0.60). These genes are associated with inflammatory feedback regulation, complement-mediated innate immune activation, and counter-regulatory anti-inflammatory pathways.

On the PCA biplot, samples colored by disease severity showed a progressive spread along the PC1 axis from cACLD to ALF, with patients with more advanced disease shifting toward negative PC1 values (Figure 1A). PC1 score decreased progressively from cACLD to ALF (Kruskal-Wallis p < 0.0001, Figure 1B), indicating that the dominant axis of transcriptomic variability reflects the CAID severity continuum. Figures 1C–H illustrate representative genes from distinct immune functions, monocyte antigen presentation (CIITA), lymphocyte signaling (IL7R), cytotoxic function (GNLY), anti-inflammatory cytokines (IL10), neutrophil activation (CD177) and alarmin-related inflammation (S100A9), all showing significant progressive modulation across severity stages (all adjusted q < 0.001). Expression levels of all additional significantly modulated genes are shown in Supplementary Figure 3, alongside mHLA-DR expression and lymphocyte counts (q < 0.001 and q < 0.05 respectively). To assess whether the transcriptomic differences observed across severity stages could be explained by etiological heterogeneity, gene expression was compared across the three main underlying etiologies (ALD, viral, autoimmune). No significant differences were observed for any of the 26 genes after correction for multiple testing (all q > 0.10, Supplementary Table 1), although several genes reached nominal significance before adjustment. No etiology-specific transcriptomic differences could therefore be detected in this analysis; given the modest sample size, however, this exploratory comparison was not powered to formally exclude such differences.

Among 97 patients, 34 (35%) experienced at least one significant post-LT infection within the first month. The median time to diagnosis was 9 [6–14] days post-LT. The most frequent infections were intra-abdominal infections (n = 19), followed by pneumonia (n = 15), bacteremia (n = 2), fungal infections (n = 3) and viral infections (n = 2). Microbiological details are provided in Supplementary Materials.

Regarding recipient characteristics, infected patients presented with a greater number of pre-LT organ failures (median 1 [0–2] vs. 0 [0–1], p = 0.04), and a higher rate of ICU admission at the time of LT (38% vs. 11%, p = 0.004); no significant differences were observed in MELD or SOFA scores (Table 1). Regarding donor and perioperative factors, intraoperative packed red blood cell transfusion requirements were significantly higher in infected patients (median 4 [2–8] vs. 2 [0–5] units, p = 0.005), EAD was more frequent (44.1% vs. 23.8%, p = 0.04), and a trend toward lower DCD utilization was observed (5.9% vs. 23.8%, p = 0.05, Table 1).

On the PCA biplot, samples from infected and non-infected patients showed overlapping distributions with partially separated confidence ellipses (Figure 2A). However, PC1 score was significantly lower in infected patients compared to non-infected patients (p = 0.044, uncorrected, Figure 2B), suggesting a possibly more pronounced pre-LT immune deficit in patients who subsequently developed infections (this comparison was not adjusted for multiple testing and should be interpreted as borderline).

Volcano plot analysis identified three genes reaching adjusted significance (q < 0.10 and AUC > 0.65): GNLY and IL7R were downregulated in infected patients, whereas IL10 was upregulated (Figure 2C). These genes were retained as candidate biomarkers for downstream analyses. For all 26 transcriptomic markers, univariate Wilcoxon rank-sum p-values, Benjamini–Hochberg adjusted q-values, and ROC-derived AUCs are reported in Supplementary Table 2.

Individual boxplots confirmed differential expression of IL10, GNLY and IL7R between infected and non-infected patients (Figures 2D–F). ROC analyses demonstrated comparable discriminative performances for IL10, GNLY and IL7R, with AUCs of 0.67 [0.55–0.78], 0.67 [0.56–0.79], and 0.66 [0.54–0.79], respectively. MELD and SOFA scores did not reach significance for predicting early infection (AUC 0.61 [0.49–0.74], p = 0.07 and 0.60 [0.47–0.72], p = 0.11).

Kaplan-Meier analyses using cohort medians as thresholds demonstrated significantly reduced 30-day infection-free survival in patients with low GNLY expression (54% vs. 76%, p = 0.019, Figure 2G) and with high IL10 expression (53% vs. 77%, p = 0.013, Figure 2H). IL7R dichotomized at the median showed a borderline association with infection-free survival (56% vs. 73%, p = 0.052, Figure 2I).

Calibration analyses of univariable logistic regression models demonstrated good agreement between predicted probabilities and observed infection rates for GNLY (Hosmer–Lemeshow, p = 0.93) and IL10 (p = 0.69). In contrast, IL7R showed borderline calibration (p = 0.03), with a non-linear pattern on the calibration plot suggesting heterogeneity across predicted probability ranges (Figures 2J–L).

Seven patients (7.2%) died within 1 year of LT; six deaths were infection-related and one resulted from refractory hemodynamic failure within 24 h post-LT. Regarding clinical characteristics, non-survivors tended to be younger, more frequently admitted to the ICU at the time of LT (57% vs. 18%), and had higher MELD (28 vs. 20) and SOFA scores (6 vs. 5); none of these differences reached statistical significance (Table 2). Among perioperative factors, intraoperative transfusion requirements were significantly higher in non-survivors (median 8 [5–9] vs. 2 [0–5] units, p < 0.001) and EAD was more frequent (71% vs. 28%, p = 0.03, Table 2).

On the PCA biplot, samples from survivors and non-survivors showed largely overlapping distributions (Figure 3A), and PC1 score did not differ significantly between groups (p = 0.188, Figure 3B). This binary comparison discards timing and censoring information and should therefore be interpreted with caution; it suggests, but does not establish, that the global immune signature captured by PC1 is less discriminant for one-year mortality than for early infections. Volcano plot analysis identified four genes meeting the predefined exploratory selection criteria (q < 0.10 and AUC > 0.65): CIITA was underexpressed in non-survivors, whereas IL10, CD177 and S100A9 were overexpressed (Figure 3C). AUC values, nominal p-values and Benjamini–Hochberg adjusted q-values for all 26 transcriptomic markers are reported in Supplementary Table 3. These four genes were retained as candidate biomarkers for further exploratory analyses. Individual boxplots confirmed differential expression of CIITA, IL10, CD177 and S100A9 between survivors and non-survivors (Figures 3D–G). ROC curve analyses showed AUCs ranging from 0.79 to 0.85 for these candidate markers (Supplementary Table 3).

Given the limited number of deaths (n = 7), survival analyses were considered exploratory. Kaplan-Meier analyses using cohort median thresholds showed significantly lower one-year overall survival among patients with low CIITA expression (87.5% vs. 98.0%, p = 0.049), while similar trends were observed for high IL10, CD177 and S100A9 expression (Supplementary Figure 4).