Lactate-related biomarkers in severe acute pancreatitis: insights from machine learning, molecular dynamics, and experimental validation.

Figure  1 shows the workflow of this study. To study the molecular differences between SAP and NSAP, we carried out Gene Set Enrichment Analysis (GSEA) on the transcriptomic data from both groups. The results showed that the SAP group was enriched in metabolism-related pathways, such as thermogenesis, PPAR signaling, and oxidative phosphorylation (Fig.  2 A–C). The NSAP group was enriched in immune-related pathways, such as Th1/Th2/Th17 differentiation and antigen processing and presentation (Fig.  2 D–F). Fig. 1 Study workflow Study workflow Fig. 2 GSEA analysis reveals distinct pathway activation patterns between SAP and NSAP. ( A – C ) Pathways upregulated in SAP. ( D – F ) Pathways upregulated in NSAP GSEA analysis reveals distinct pathway activation patterns between SAP and NSAP. ( A – C ) Pathways upregulated in SAP. ( D – F ) Pathways upregulated in NSAP To identify LRGs associated with AP progression, we intersected DEGs between SAP and NSAP with the predefined LRG set, yielding 24 candidate genes (Fig.  3 A–B). A PPI network constructed via the STRING database revealed potential interactions among these proteins (Fig.  3 C). GeneMANIA analysis further confirmed their co-expression, co-localization, and functional clustering, showing enrichment in ATP synthesis–coupled electron transport, mitochondrial ATP synthesis–coupled electron transport, and positive regulation of the respiratory electron transport chain (Fig.  3 D). Functional enrichment using Metascape indicated that these genes were mainly involved in the aerobic electron transport chain, NADH dehydrogenase complex assembly, and DNA double-strand break repair (Fig.  3 E). Fig. 3 Identification of hub genes associated with lactate-related dysregulation in SAP and functional enrichment analysis. ( A ) Volcano plot displaying DEGs between SAP and NSAP. ( B ) Venn diagram illustrating the intersection between DEGs and predefined LRGs, yielding 24 candidate genes. ( C ) PPI network constructed from the intersected genes using STRING, highlighting key interaction clusters. ( D ) GeneMANIA functional interaction network predicting gene–gene associations, co-expression patterns, and shared pathway features among the intersected genes. ( E ) Functional enrichment analysis of the intersected genes based on Metascape Identification of hub genes associated with lactate-related dysregulation in SAP and functional enrichment analysis. ( A ) Volcano plot displaying DEGs between SAP and NSAP. ( B ) Venn diagram illustrating the intersection between DEGs and predefined LRGs, yielding 24 candidate genes. ( C ) PPI network constructed from the intersected genes using STRING, highlighting key interaction clusters. ( D ) GeneMANIA functional interaction network predicting gene–gene associations, co-expression patterns, and shared pathway features among the intersected genes. ( E ) Functional enrichment analysis of the intersected genes based on Metascape To establish a molecular diagnostic model for SAP, we applied LASSO regression, SVM-RFE, and RF to the differentially expressed LRGs. LASSO regression identified eight potential marker genes (Supplementary Fig.  1 A), SVM-RFE selected ten candidate genes (Supplementary Fig. 1B), and RF analysis identified seven genes with importance scores above the threshold (Supplementary Fig.  1 C–D). We compared the results of three machine learning algorithms and identified two key genes, CCNA2 and H2BC5, as candidates for further investigation (Fig.  4 A). Their chromosomal locations are shown in Fig.  4 B. Later expression analysis showed that both genes were clearly increased in SAP (Fig.  4 C). This finding suggests that they may take part in the progression from AP to severe disease. The two-gene SAP model showed good diagnostic ability (AUC = 0.884). The AUC for CCNA2 was 0.827, and the AUC for H2BC5 was 0.816 (Fig.  4 D–E). Furthermore, we compared their expression levels across different severities of AP as well as healthy controls. CCNA2 showed no significant difference between normal controls and NSAP, whereas H2BC5 expression progressively increased with the onset and aggravation of AP (Fig.  4 F–G). The in vitro experimental results showed a significant increase in the inflammatory cytokines IL-1β and IL-6, confirming the successful establishment of the AP in vitro model. Subsequently, it was further demonstrated that the model genes were significantly upregulated in AP cell model (Fig.  4 H). To further validate the diagnostic value of model genes in clinical samples, we collected peripheral blood from clinical patients. Results showed that CCNA2 and H2BC5 were significantly upregulated in SAP patients (Fig.  4 I). Fig. 4 Diagnostic performance and experimental validation of the CCNA2–H2BC5 two-gene model for SAP. ( A ) Venn diagram illustrating the intersection of genes selected by multiple machine-learning algorithms, identifying CCNA2 and H2BC5 as the final model genes. ( B ) Chromosomal localization of CCNA2 and H2BC5. ( C ) Boxplots of CCNA2 and H2BC5 expression between SAP and NSAP patients. ( D – E ) ROC curves for the two-gene model and individual genes. ( F – G ) The expression of CCNA2 and H2BC5 in different groups. ( H ) Expression levels of CCNA2, H2BC5, and key inflammatory cytokines in the STC-induced in vitro AP model, confirming gene upregulation in injury conditions ( n  = 3). ( I ) Expression levels of model genes in peripheral blood of clinical samples ( n  = 8) Diagnostic performance and experimental validation of the CCNA2–H2BC5 two-gene model for SAP. ( A ) Venn diagram illustrating the intersection of genes selected by multiple machine-learning algorithms, identifying CCNA2 and H2BC5 as the final model genes. ( B ) Chromosomal localization of CCNA2 and H2BC5. ( C ) Boxplots of CCNA2 and H2BC5 expression between SAP and NSAP patients. ( D – E ) ROC curves for the two-gene model and individual genes. ( F – G ) The expression of CCNA2 and H2BC5 in different groups. ( H ) Expression levels of CCNA2, H2BC5, and key inflammatory cytokines in the STC-induced in vitro AP model, confirming gene upregulation in injury conditions ( n  = 3). ( I ) Expression levels of model genes in peripheral blood of clinical samples ( n  = 8) To investigate the functional relevance of lactate in regulating the identified hub genes, we first treated H6C7 cells with exogenous lactate. Lactate stimulation (20mM) markedly increased the mRNA expression of CCNA2 and H2BC5, accompanied by significant upregulation of the inflammatory cytokines IL-1β and IL-6, indicating that lactate exacerbates inflammatory signaling in pancreatic cells (Fig.  5 A). To further explore the biological significance of these genes, we designed three independent siRNA sequences for each target. Among them, si-CCNA2-3 and si-H2BC5-1 achieved the highest knockdown efficiency, and were therefore selected for subsequent experiments (Fig.  5 B-C). The knockdown of CCNA2 or H2BC5 significantly reduced IL-1β and IL-6 expression under both STC and lactate-stimulated conditions. These findings suggest that CCNA2 and H2BC5 act as downstream effectors of lactate-driven immunometabolic activation, contributing to the amplification of inflammatory responses in vitro (Fig.  5 D). Fig. 5 Lactate stimulation upregulates CCNA2/H2BC5 and inflammatory cytokines, while siRNA-mediated knockdown attenuates lactate-induced inflammation. ( A ) mRNA expression levels of IL-1β, IL-6, CCNA2, and H2BC5 in H6C7 cells following treatment with exogenous lactate (20mM) ( n  = 3). ( B ) Screening of three siRNA sequences for CCNA2 to determine knockdown efficiency ( n  = 3). ( C ) Screening of three siRNA sequences for H2BC5 to determine knockdown efficiency ( n  = 3). ( D ) Effects of CCNA2 or H2BC5 knockdown on inflammatory cytokine production ( n  = 3) Lactate stimulation upregulates CCNA2/H2BC5 and inflammatory cytokines, while siRNA-mediated knockdown attenuates lactate-induced inflammation. ( A ) mRNA expression levels of IL-1β, IL-6, CCNA2, and H2BC5 in H6C7 cells following treatment with exogenous lactate (20mM) ( n  = 3). ( B ) Screening of three siRNA sequences for CCNA2 to determine knockdown efficiency ( n  = 3). ( C ) Screening of three siRNA sequences for H2BC5 to determine knockdown efficiency ( n  = 3). ( D ) Effects of CCNA2 or H2BC5 knockdown on inflammatory cytokine production ( n  = 3) GSVA was conducted on all AP samples to investigate the potential functional roles of CCNA2 and H2BC5. CCNA2 showed enrichment in several KEGG pathways, such as the p53 signaling pathway, oocyte meiosis, and steroid biosynthesis. GO terms further suggested that CCNA2 participates in processes including centromere remodeling, spindle microtubule attachment, chromosome organization, and cellular responses to insulin-like growth factor (Fig.  6 A–B). For H2BC5, KEGG analysis indicated enrichment in complement and coagulation cascades and the NOD-like receptor signaling pathway. Corresponding GO annotations pointed to functions related to nucleosome assembly, chromatin distribution, and the regulation of Th2 cell differentiation (Fig.  6 C–D). In addition, immune infiltration profiling using CIBERSORT demonstrated distinct correlations between the model genes and various immune cell subsets. CCNA2 expression correlated positively with resting mast cells and negatively with resting CD4 memory T cells and CD8 T cells (Fig.  6 E). Similarly, H2BC5 expression was negatively correlated with resting CD4 memory T cells and CD8 T cells, but positively correlated with neutrophils (Fig.  6 F). A comprehensive immune interaction network (Fig.  6 G) illustrated the associations between model genes and immune cells, as well as correlations among immune cell subsets. Collectively, these findings suggest that CCNA2 and H2BC5 play distinct regulatory roles within the AP immune microenvironment. Fig. 6 Functional annotation and immune landscape of the model genes CCNA2 and H2BC5. ( A – B ) KEGG and GO enrichment of CCNA2. ( C – D ) KEGG and GO enrichment of H2BC5. ( E – F ) Correlation heatmaps showing relationships between CCNA2/H2BC5 expression and infiltration of major immune cell subsets. ( G ) Immune interaction network showing relationships between model genes and infiltrating immune cells Functional annotation and immune landscape of the model genes CCNA2 and H2BC5. ( A – B ) KEGG and GO enrichment of CCNA2. ( C – D ) KEGG and GO enrichment of H2BC5. ( E – F ) Correlation heatmaps showing relationships between CCNA2/H2BC5 expression and infiltration of major immune cell subsets. ( G ) Immune interaction network showing relationships between model genes and infiltrating immune cells To clarify the expression sources and regulatory mechanisms of the model genes, we further performed single-cell localization and immunofluorescence analyses. The results showed that CCNA2 was primarily expressed in pancreatic B cells (Fig.  7 A), and was detected in nucleoplasm and cytosol (Fig.  7 B). H2BC5 expression was more enriched in pancreatic α cells (Fig.  7 C), and was also detected in nucleoplasm and cytosol (Fig.  7 D). Subsequently, to characterize the transcription factors (TFs) of these genes, we integrated four databases and identified several TFs with high regulatory potential. Specifically, eight TFs were associated with CCNA2 (ZNF610, KLF12, SP1, ZNF341, ATF2, ATF7, CREB1, and HOXA4), while two TFs were associated with H2BC5 (MEF2B and MEF2C) (Fig.  7 E–F). These findings provide a basis for further investigation of their transcriptional regulatory networks. Fig. 7 Single-cell localization analysis, immunofluorescence, and transcription factor prediction of model genes. ( A – B ) Single-cell localization and immunofluorescence of CCNA2. ( C – D ) Single-cell localization and immunofluorescence of H2BC5. ( E – F ) TF prediction for CCNA2 and H2BC5 based on four databases Single-cell localization analysis, immunofluorescence, and transcription factor prediction of model genes. ( A – B ) Single-cell localization and immunofluorescence of CCNA2. ( C – D ) Single-cell localization and immunofluorescence of H2BC5. ( E – F ) TF prediction for CCNA2 and H2BC5 based on four databases To identify potential therapeutic agents, we imported the DEGs into the CMap database and obtained the top 10 candidate molecules. We then used the CB-Dock2 online platform to perform molecular docking between these drugs and the model genes (Supplementary Table 1 ). Among them, the binding energy of CCNA2 with AMG-232 was − 12.8 kcal/mol, and that of H2BC5 with Torin-1 was − 8.7 kcal/mol, suggesting strong binding affinities. Structural analysis revealed that AMG-232 formed seven stable hydrogen bonds with CCNA2 (Fig.  8 A), whereas Torin-1 did not form stable hydrogen bonds with H2BC5, but instead engaged in 11 hydrophobic interactions and two π–π stackings (Fig.  8 B). Next, 200-ns MD simulations were conducted to evaluate the stability of the complexes. The RMSD curves indicated that both complexes quickly reached equilibrium during the early stage of simulation and remained relatively stable throughout (Fig. 8 C–D). RMSF analysis showed that both complexes had only small atomic movements (Fig. 8 E–F). In the same way, SASA and Rg results showed that both complexes stayed tightly bound and structurally stable during the whole simulation (Fig. 8 G–J). To check stability further, FELs made from RMSD and Rg data showed clear low-energy basins for each system (Fig. 8 K–L). PCA gave more information and showed that most coordinated motions were found in loop regions (Fig. 8 M–N). When the FEL was plotted on the first two principal components (PC1 and PC2), the AMG-232–CCNA2 complex showed two main low-energy states (Fig. 8 M). The H2BC5–Torin-1 complex showed more energy minima (Fig. 8 N). This suggests that AMG-232 may bind to a more focused and specific site. Hydrogen-bond analysis supported these results. The AMG-232–CCNA2 complex kept 1–8 hydrogen bonds, and the H2BC5–Torin-1 complex kept 1–4 hydrogen bonds, showing stable interactions in both systems (Fig. 8 O–P). Considering that Torin-1 has been reported to possess anti-inflammatory properties [ 42 – 44 ] and exhibited favorable docking affinities with CCNA2 and H2BC5 (− 9.2 kcal/mol and − 8.7 kcal/mol, respectively), we further evaluated its therapeutic potential in AP using an in vitro model. In the STC-induced AP model, Torin-1 markedly attenuated the elevation of inflammatory cytokines and the production of intracellular ROS (Fig. 8 Q–R). Moreover, Torin-1 significantly increased intracellular GSH levels, further supporting its role in alleviating oxidative stress (Fig. 8 S). Fig. 8 Molecular docking and MD simulation of model genes with predicted drugs. ( A ) Molecular docking structure of CCNA2–AMG-232, with blue dashed lines indicating hydrogen bonds; ( B ) Molecular docking structure of H2BC5–Torin-1, with gray dashed lines indicating hydrophobic interactions and green dashed lines indicating π–π stacking; ( C – D ) RMSD curves of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( E – F ) RMSF curves of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( G – H ) SASA of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( I – J ) Rg of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; (K–L) Two-dimensional and three-dimensional FEL of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; (M–N) PCA covariance matrix heatmaps and FELs based on PC1 and PC2 for CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( O – P ) Hydrogen bond variations in CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( Q – S ) In vitro evaluation of Torin-1 in STC-induced AP model: ( Q ) inflammatory cytokine levels, ( R ) intracellular ROS production, and ( S ) intracellular GSH levels ( n  = 3) Molecular docking and MD simulation of model genes with predicted drugs. ( A ) Molecular docking structure of CCNA2–AMG-232, with blue dashed lines indicating hydrogen bonds; ( B ) Molecular docking structure of H2BC5–Torin-1, with gray dashed lines indicating hydrophobic interactions and green dashed lines indicating π–π stacking; ( C – D ) RMSD curves of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( E – F ) RMSF curves of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( G – H ) SASA of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( I – J ) Rg of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; (K–L) Two-dimensional and three-dimensional FEL of CCNA2–AMG-232 and H2BC5–Torin-1 complexes; (M–N) PCA covariance matrix heatmaps and FELs based on PC1 and PC2 for CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( O – P ) Hydrogen bond variations in CCNA2–AMG-232 and H2BC5–Torin-1 complexes; ( Q – S ) In vitro evaluation of Torin-1 in STC-induced AP model: ( Q ) inflammatory cytokine levels, ( R ) intracellular ROS production, and ( S ) intracellular GSH levels ( n  = 3)

Transcriptomic data of AP patients were retrieved from the GEO database ( https://www.ncbi.nlm.nih.gov/geo/ ), using the dataset GSE194331 [ 11 ], which includes 87 AP patients with varying disease severity. Based on clinical information, patients were stratified into the SAP group and the NSAP group. In addition, a total of 587 LRGs were obtained from previously published studies [ 12 , 13 ]. Differentially expressed genes (DEGs) between SAP and NSAP samples were identified using the “limma” R package (version 3.40.6), which applies a linear model–based approach [ 14 ]. Genes with more than 50% zero expression values were excluded. Expression data were transformed with the voom function, fitted to a multiple linear regression model using lmFit, and further processed with eBayes to compute moderated t-statistics, moderated F-statistics, and log-odds ratios through empirical Bayes moderation. Gene Set Enrichment Analysis (GSEA) was subsequently performed to evaluate enriched pathways across the global transcriptome, identifying upregulated and downregulated functional categories [ 15 ]. LRGs were then intersected with DEGs using a Venn diagram to identify SAP-associated LRGs. To explore protein–protein interactions (PPIs) among SAP-associated LRGs, we constructed a PPI network using the STRING database ( https://cn.string-db.org/ ), applying a minimum interaction confidence score of 0.15 [ 16 ]. Gene–gene interaction networks were further generated using the GeneMANIA platform ( https://genemania.org/ ), integrating information from co-expression, co-localization, and shared signaling pathways [ 17 ]. Functional annotation and pathway enrichment of intersected genes were performed using the Metascape database ( https://metascape.org/ ) [ 18 ]. To construct a molecular model discriminating SAP from NSAP, three ML algorithms were applied to the differential LRGs: LASSO regression: implemented via the “glmnet” package, with 10-fold cross-validation to determine the optimal λ and select key feature genes [ 19 , 20 ]. SVM-RFE: performed using the “e1071” package, with recursive feature elimination and 10-fold cross-validation to identify the optimal feature subset [ 21 ]. Random Forest (RF): trained with 500 trees using the “randomForest” package, and genes with importance scores > 1 were retained [ 22 ]. The intersected results from all three methods were defined as the core model genes. Predictive performance of the model genes was evaluated by receiver operating characteristic (ROC) curve analysis using the “pROC” package, with the area under the curve (AUC) as the diagnostic metric [ 23 ]. To validate the expression of model genes, peripheral blood samples from 24 participants were collected from the First Affiliated Hospital of Dalian Medical University, including healthy control group ( n  = 8), MAP group ( n  = 8), and SAP group ( n  = 8). The study was approved by the Ethics Committee of the First Affiliated Hospital of Dalian Medical University. All enrolled patients and healthy volunteers agreed to the use of their blood samples for the study by signing a written consent form. All methods were performed in accordance with the relevant guidelines and regulations. RNA was isolated from peripheral whole blood using the QIAamp RNA Blood Mini Kit, which extracts RNA from all nucleated blood cells. Complementary DNA was made with a reverse transcription kit. The expression levels of target genes were measured by quantitative real-time PCR (qPCR). To make an in vitro model of acute pancreatitis, H6C7 cells were treated with 250 µM sodium taurocholate (STC) for 24 h, based on earlier reports [ 24 , 25 ]. After the model was set up, the cells were treated with 10 nM Torin-1 (Beyotime, Shanghai, China) to test its effects. The success of the model and the response to Torin-1 were measured by checking IL-1β and IL-6 levels. Changes in gene expression were measured by qPCR. To study how Torin-1 affected oxidative stress, intracellular ROS and glutathione (GSH) levels were measured with fluorescence-based assay kits (Nanjing Jiancheng Bioengineering Institute, Nanjing, China) following the manufacturer’s instructions. Total RNA was extracted with TRIzol, converted to cDNA, and analyzed by qPCR using SYBR Green I. Relative transcript levels were calculated with the 2 − ΔΔCt method. To test if lactate changes the expression of model-related genes and inflammatory mediators, H6C7 cells were treated with 20 mM exogenous lactate under normal culture conditions [ 26 ]. After stimulation, total RNA was collected and analyzed by qPCR to measure CCNA2, H2BC5, and selected inflammatory cytokines. To study the roles of CCNA2 and H2BC5, gene knockdown experiments were done in H6C7 cells using specific siRNAs. This allowed testing how reduced gene expression affects lactate-induced inflammatory responses. Three siRNAs targeting H2BC5 (si-H2BC5-1, si-H2BC5-2, si-H2BC5-3) and three targeting CCNA2 (si-CCNA2-1, si-CCNA2-2, si-CCNA2-3), along with negative controls, were bought from Hema Biology (Huzhou, China). After transfection, the cells received the same stimulation as control groups, and RNA was collected to measure changes in inflammatory cytokine expression. Gene Set Variation Analysis (GSVA) was used to study how the model genes take part in metabolic and inflammatory pathways in all AP samples [ 27 ]. To examine immune patterns, the CIBERSORT algorithm was used to estimate the relative levels of 22 immune cell types in each sample. The samples were then grouped by high or low gene expression, and differences in immune infiltration were compared [ 28 ]. Correlations between immune cell levels and model gene expression were calculated and shown with the corrplot package [ 29 ]. Single-cell expression patterns and immunofluorescence localization of the model genes in pancreatic tissues were examined using data from the Human Protein Atlas (HPA; https://www.proteinatlas.org/ ) [ 30 ]. To find possible upstream transcription factors (TFs), the TFTF online platform was used [ 31 ]. It integrates four tools: FIMO_JASPAR [ 32 ], ChIP_Atlas [ 33 ], cor_GTEx [ 34 ], and PWMEnrich_JASPAR [ 35 ]. To screen for candidate therapeutic compounds that may target the model genes, we queried the Connectivity Map (CMap) database ( https://clue.io/ ) [ 36 ]. Molecular docking was conducted to evaluate binding energy, binding sites, and structural stability, with protein 3D structures obtained from the PDB database ( https://www.rcsb.org/ ) [ 37 ] or predicted by the AlphaFold Protein Structure Database ( https://alphafold.ebi.ac.uk/ ) [ 38 ]. Docking simulations were performed using the CB-Dock2 platform ( https://cadd.labshare.cn/cb-dock2/index.php ) [ 39 ]. Subsequently, MD simulations were carried out with GROMACS 2021.3 [ 40 ] under the CHARMM36 force field [ 41 ]. Structural stability and dynamic behavior of drug–target complexes were assessed through: Root mean square deviation (RMSD): overall structural stability over time. Root mean square fluctuation (RMSF): residue-level flexibility. Solvent-accessible surface area (SASA): solvation dynamics and conformational compactness. Radius of gyration (Rg): global structural compactness. Hydrogen bond variation: stability of ligand–receptor interactions. Free energy landscape (FEL): thermodynamic stability and distribution of low-energy conformations. Principal component analysis (PCA): dominant conformational motions, with covariance matrix heatmaps illustrating correlated structural dynamics. Statistical analyses were performed using R software (version 4.1.3) and GraphPad Prism (version 10.1.2). All experiments were independently repeated at least three times, and data were expressed as mean ± SD. The normality of data distribution was assessed using the Shapiro-Wilk test. The Student’s t-test was used for comparisons between two groups. For comparisons involving more than two groups, one-way ANOVA was performed, followed by Tukey’s post-hoc test to adjust for multiple comparisons. For transcriptomic analysis, P-values were adjusted using the Benjamini-Hochberg method to control the false discovery rate (FDR). Correlation analysis was performed using Spearman’s correlation coefficients. A P-value < 0.05 was considered statistically significant.

This study, based on transcriptomic comparisons between SAP and NSAP, systematically delineated the molecular characteristics of LRGs in AP severity and identified CCNA2 and H2BC5 as two hub biomarkers using multiple ML algorithms. Together with drug prediction and molecular simulation results, this study highlights AMG-232 and Torin-1 as potential therapeutic leads. Our work provides molecular clues for the early prediction and stratification of SAP and offers a theoretical framework for targeted therapeutic exploration of lactate metabolism–immune interactions, rather than definitive mechanistic evidence.

AP is a common acute abdominal disorder that can rapidly progress from mild, self-limiting inflammation to SAP characterized by MODS, posing a serious threat to patient survival [ 45 ]. Early identification and intervention in the progression to severe disease are critical to improving prognosis [ 46 ]. Lactate, as a byproduct of anaerobic glycolysis and a metabolic signaling molecule, is markedly elevated under hypoxia, inflammation, and hypermetabolic stress [ 47 ]. It has long served as an important marker for risk assessment and prognostic evaluation in critically ill patients [ 48 , 49 ]. More recent evidence indicates that lactate is not merely a by-product of altered perfusion or metabolic stress; it can also shape inflammatory responses by influencing immune cell activity and driving epigenetic processes such as lysine lactylation [ 50 ]. These findings open a new way to study the metabolic basis of AP severity [ 47 , 51 ]. In this study, we focused on lactate metabolism and lactylation to identify LRGs linked to AP severity using transcriptomic data. We built a diagnostic model for SAP with three machine learning algorithms, and two hub genes, CCNA2 and H2BC5, were found. After that, through functional enrichment analysis, immune infiltration analysis, single-cell localization, drug prediction, and MD simulations, we clarified how these genes take part in immunometabolic imbalance and why they may have clinical value. GSEA results showed that SAP patients were enriched in thermogenesis, PPAR signaling, and oxidative phosphorylation pathways, reflecting mitochondrial dysfunction and metabolic reprogramming. These findings match the main features of severe inflammation, including high ROS levels, poor oxidative phosphorylation, and energy imbalance. In NSAP, patients showed enrichment in immune pathways, such as Th1/Th2/Th17 differentiation and antigen presentation. This shows that the immune response is more active at the early stage of the disease. These results suggest that metabolic imbalance together with immune suppression drives AP severity. Lactate acts both as a result and a cause of this imbalance. We also found that the 24 differentially expressed LRGs were mainly related to aerobic electron transport, NADH dehydrogenase complex assembly, and DNA double-strand break repair. This supports the link between SAP, mitochondrial injury, and high oxidative stress. Recent studies show that mitochondrial dysfunction lowers ATP production and raises lactate levels. This creates a harmful metabolic cycle and changes the lactate/pyruvate ratio. These changes then affect immune cell polarization [ 52 , 53 ]. CCNA2 encodes cyclin A2, which controls G1/S and G2/M transitions and is linked to DNA repair and p53 signaling [ 54 – 56 ]. We found that CCNA2 was clearly increased in SAP and enriched in processes such as chromosome segregation, spindle activity, and insulin-like growth factor responses. Immune infiltration analysis showed that CCNA2 had a positive correlation with resting mast cells and a negative correlation with CD4 + memory T cells and CD8 + T cells. This suggests that it may support stress-related cell growth while reducing cytotoxic T-cell activity, which may lead to immune suppression in SAP. H2BC5 (formerly HIST1H2BD) is a histone H2B family gene involved in chromatin assembly and transcription control [ 57 ]. It was also increased in SAP and enriched in complement–coagulation cascades, NOD-like receptor signaling, and Th2-related pathways. Its positive correlation with neutrophils suggests that it may help promote inflammatory gene expression and neutrophil extracellular trap formation by changing chromatin accessibility, which may further increase inflammation. From an immunometabolic view, lactate and its downstream change, lactylation, act as key points that can be used to adjust immune activity [ 47 , 58 ]. Changing lactate metabolism can shift macrophage polarization, T-cell activity, and chromatin-based inflammatory programs [ 59 ]. The finding of lactate-related CCNA2 and H2BC5 supports the idea that molecular features from MD simulations can help design small-molecule drugs that aim to restore immunometabolic balance. Drug prediction and molecular docking showed that CCNA2 forms a stable complex with AMG-232, and H2BC5 forms a stable complex with Torin-1. These results were confirmed by MD simulations. AMG-232 is a strong MDM2 inhibitor that can restore p53 function, stop cell-cycle progression, and improve metabolic problems in many tumor models and clinical studies [ 60 , 61 ]. Torin-1 is an ATP-competitive mTOR inhibitor that blocks both mTORC1 and mTORC2 signaling [ 62 ]. Torin-1 also reduces inflammation, lowers oxidative stress, and adjusts immune-cell activity in several disease models [ 62 , 63 ]. For example, Torin-1 can reduce inflammation-related apoptosis, restore autophagy, and improve metabolic balance by supporting mitochondrial function and redox control. In AP, these actions are very important, because high ROS levels, damaged mitochondria, and excessive inflammatory signaling are major features of severe disease. In our in vitro work, Torin-1 clearly lowered inflammatory cytokine release, reduced ROS, and increased GSH levels. These results show its combined anti-inflammatory and antioxidative effects. The choice of AMG-232 and Torin-1 was based on their predicted stable docking with CCNA2 and H2BC5 and their effect on cell cycle and metabolic pathways in SAP. In this study, MD simulations were used to get quantitative data, such as binding free energy and hydrogen-bond occupancy. This data can help in virtual drug screening. The MD results support ranking and improving potential drugs and combine computational prediction with experimental tests. This method is an early example of an engineering-guided drug discovery approach, where modeling, simulation, and experiments together guide drug design. It is important to note that AMG-232–CCNA2 and Torin-1–H2BC5 show predicted structural fit but do not confirm biological activity. Docking and MD results show possible binding stability but do not prove clinical use. More biochemical, pharmacological, and in vivo studies are needed to confirm real therapeutic effects. Compared to earlier SAP biomarker studies [ 11 , 64 , 65 ], this study has several strengths. It focuses on lactate-related molecular networks, linking metabolism changes with immune dysfunction to give deeper mechanistic insight. Using three machine learning methods allowed consensus-based feature selection, which improves reliability. Molecular docking combined with 200-ns MD simulations helped screen potential drugs. The diagnostic model compared NSAP and SAP, helping early detection of disease severity. Experiments in cells and clinical blood samples confirmed the increase of CCNA2 and H2BC5 and their link to inflammatory responses, showing translational relevance. However, this study has some limitations. First, it used public transcriptomic datasets, and sample size and population differences may limit general use. Second, the links between CCNA2/H2BC5 and lactate metabolism are mostly from correlative bioinformatics analyses, not direct mechanistic proof. Finally, while our study combines robust clinical transcriptomic data with in vitro functional assays to support the roles of CCNA2 and H2BC5, specific animal models were not established in this phase. Generating knockout animal models is necessary to fully elucidate their systemic effects in SAP. This will be the primary focus of our future investigations to further substantiate the molecular mechanisms proposed here.

Acute pancreatitis (AP) is an inflammatory disorder of the pancreas, and its clinical signs can vary widely. Most patients have a mild and self-limiting illness, but some can quickly develop severe acute pancreatitis (SAP), which often comes with multiple organ dysfunction syndrome (MODS) [ 1 ]. AP affects more than 2.7 million people worldwide each year, and the incidence is about 32.8 per 100,000 population. The age-standardized incidence has slightly decreased, but the total number of cases is still increasing [ 2 ]. The 2012 revised Atlanta classification divides AP into mild (MAP), moderately severe (MSAP), and SAP [ 3 ]. About 10–20% of AP patients progress to SAP, and the mortality rate is around 20–30% [ 4 , 5 ]. Therefore, early biomarkers are greatly needed for clinical management. Lactate was once known only as the end product of glycolysis and as a marker of poor tissue perfusion or metabolic stress. It is often used to judge disease severity in critically ill patients [ 6 ]. Now, more studies show that lactate has wider roles. It acts as a signaling molecule that reflects cellular metabolic status in low-oxygen or high-metabolic conditions, and it can directly shape immune and inflammatory responses [ 7 , 8 ]. Lactate can push macrophages toward an M2 phenotype and reduce the activity of CD8⁺ T cells. In this way, it creates an immunosuppressive state that keeps inflammation going and worsens tissue damage [ 9 ]. Lactate also takes part in epigenetic regulation through lysine lactylation, a newly identified post-translational modification that controls the transcription of inflammation-related genes and links metabolism with chromatin remodeling [ 10 ]. However, the exact roles of lactate metabolism and lactylation in the progression of AP are still not clear. MAP and MSAP usually show milder clinical patterns, with reversible organ dysfunction and low mortality. Because of this, we grouped MAP and MSAP together as non-severe AP (NSAP) in this study. Our aim was to explore pathways related to lactate metabolism and lactylation, find lactate-related genes (LRGs), and compare their expression in SAP and NSAP. To do this, we used machine-learning methods to identify lactate-centered molecular markers that may help predict SAP. Our goal was to build a useful framework for early risk evaluation in AP and to point out possible therapeutic targets. This study also builds a clear computational-to-experimental workflow with value for bioengineering research. By integrating transcriptomics, machine learning (ML), molecular docking, molecular dynamics (MD), and in vitro validation, we designed a modular and reproducible framework that can be engineered into predictive diagnostic algorithms, decision-support systems, and structure-based virtual screening tools for precision medicine.

Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1