Exploratory identification of an ACLF-associated six-gene liver-tissue transcriptomic signature using bioinformatics analysis

Background Acute-on-chronic liver failure (ACLF) is a life-threatening disease with high morbidity and mortality, yet its pathophysiological mechanisms remain poorly understood. Methods Datasets GSE14668 and GSE38941 were quality-controlled, normalized, and merged into a training dataset. Differentially expressed genes (DEGs) were intersected with MR-prioritized candidate genes for integration with the liver-tissue transcriptomic data. An exploratory six-gene liver-tissue transcriptomic signature was derived using multiple machine-learning algorithms and evaluated in three human liver transcriptomic datasets (GSE139602, GSE298435, and GSE62029) and an exploratory ACLF rat model dataset containing three liver specimens per group. Label-permutation analysis (1,000 iterations) and random six-gene panel comparison (1,000 panels) were performed as negative controls. Model output was interpreted using Shapley additive explanations. Results 2,714 DEGs were identified, including 1,479 upregulated and 1,235 downregulated genes. Intersection of the DEGs with MR-prioritized candidate genes yielded 32 upstream candidate genes. After cross-cohort expression filtering, 28 of these genes were available in the machine-learning input matrices. The Lasso + glmBoost model subsequently identified 6 core genes (ALDH2, C1QB, C1QC, DHODH, ITGAL, and RIPK4) from these 28 features, all of which contributed to the model’s prediction outcomes. In exploratory specimen-level analyses, the AUC was 1.000 in the training cohort, 1.000 in GSE139602, 0.972 in GSE298435, and 1.000 in GSE62029. The rat dataset contained three specimens per group and also yielded an AUC of 1.000; however, neither the label-permutation comparison (P = 0.101) nor the random six-gene panel comparison (P = 0.216) was statistically significant. This result was therefore interpreted as exploratory cross-species expression evidence. In label-permutation analyses, the observed AUC exceeded the null distribution in the training cohort and all three human evaluation datasets (all P ≤ 0.006), but not in the rat dataset ( P = 0.101). Random six-gene panel comparisons supported an advantage of the selected panel in the training cohort, GSE298435, and GSE62029 (all P ≤ 0.011), but not in GSE139602 or the rat dataset, indicating cohort-dependent gene-panel specificity. The six core genes represented diverse ACLF-related biological annotations, including mitochondrial/metabolic, complement, and leukocyte adhesion-related features. Conclusion A liver tissue-based six-gene signature captured a non-random separation between the ACLF group and the Control group across the evaluated human datasets, whereas its advantage over random six-gene panels was cohort dependent. The signature may reflect heterogeneous ACLF-associated liver transcriptomic changes involving metabolic, mitochondrial, complement, and leukocyte adhesion-related signals. The rat analysis provided exploratory cross-species expression evidence only, and no functional inference was made in the absence of gene perturbation experiments. Discrimination between ACLF and other chronic liver disease states was not evaluated.