Demographic Profile
The demographic profile of both the training and test cohorts were comparable with significant increase in liver function parameters (AST, INR, Bilirubin, ALP and ALT) in ALF more so in the Non-survivors (ALF-NS) respectively. Severity indices such as MELD was high in ALF-NS as compared to ALF-S (p<0.05). Infection at baseline found to be higher in ALF-NS (>40%) as compare ALF-S. During ICU stay ALF patients develop different complications, such as hepatic encephalopathy, acute kidney injury, sepsis, necrosis and multi organ failure are also considered as death due to acute liver failure. One month mortality (30 days) was ~70% in our study cohort (Table 1).
Baseline Plasma Metabolome Profile of ALF patients are indicative of poor outcomes
Baseline plasma samples of ALF patients; Survival ;(n=8), non-survival;(n=32) and Healthy controls;(n=5) were subjected to an untargeted metabolomic evaluation. A total of 9814 features were identified both in the positive and negative ionization mode. Of those, 548 features were annotated (mass, mass+rt, mass+rt+spectral databases) and compared between the study groups. A total of 321 (153 up- and 168 downregulated) differentially expressed metabolites (DEM’s; FC>1.5, p<0.05) were identified in ALF as compared to HC. Further, 233 (114 up- and 119 downregulated) DEM’s were identified when ALF-NS were compared to ALF-S respectively (Figure 1A). The partial least square discriminant analysis (PLS-DA) along with hierarchical clustering analysis clear segregated ALF-NS from ALF-S or HC (Figure 1B, 1C Supplementary Figure 3 and 4, Supplementary table 1). ALF-NS showed significant increase in metabolic pathways associated with Tryptophan metabolism, Tyrosine metabolism, Histidine metabolism and others. Similarly, pathways linked to lysine degradation, nicotinate and nicotinamide metabolism, purine metabolism, glycerophospholipid metabolism and others were downregulated (Figure 1D). Amongst the DEMs, mean decrease accuracy was highest for chenodeoxycholic acid, L-Tyrosine, 4-(2-Aminophenyl)-2,4-dioxobutanoate and others, making them the most important metabolites which can segregate ALF-NS from ALF-S and other groups (Figure 1E and Supplementary Figure 5). Probability Of Diseases (POD); Non survival calculated using top 5 metabolites showed highest AUROC and a POD>80% (Figure 1F). Together, our results suggest that ALF-NS has a unique baseline metabotype primarily linked with increase in inflammation and oxidative stress. The identified the top 5 metabolites could be used to segregate ALF-NS and warrants validation in larger cohorts.
Baseline plasma meta-proteome profile of ALF patients are novel and are indicative of gut dysbiosis
Microbiome dysregulation is important factor associated with ALF severity leading to inflammation and dysregulation of various metabolic factors. Meta-proteome analysis of baseline ALF plasma samples identified bacterial peptides >15 phylum and >25 genera and were dysregulated in ALF-NS with significant increase in the Firmicutes, Cyanobacteria and tenericutes and others as compared to ALF-S or healthy control (p<0.05; Figure 2A, Supplementary Figure 6). The richness (alpha-diversity) was highest in ALF-NS (p<0.05) and Principal coordinate analysis (Beta diversity) analysis clearly segregate the ALF-NS microbiome from ALF-S or HC (Figure 2B, Supplementary Figure 7, Supplementary table 2). The partial least square discriminant analysis (PLS-DA) clear segregated ALF-NS from ALF-S or HC (Supplementary Figure 8). Linear discriminant analysis showed clear segregation of ALF-NS and ALF-S (Supplementary Figure 9). Random forest and MDA identify top 15 bacteria peptides (Supplementary Figure 10 and 11). Circulatory microenvironment is significantly more perturbed in the ALF-NS as compared to ALF-S or HC. We observed functional increase in microbial metabolism linked to defence and transport, xenobiotic degradation, energy and amino acid metabolism and others in ALF-NS (Figure 2C). Interestingly, increase in Actinobacteria, Cyanobacteria, and Proteobacteria associated peptides in ALF-NS directly correlated with clinical parameters like liver functionality, kidney functionality and severity indices in ALF suggesting stark association of bacterial peptides with the clinical phenotype of ALF-NS (Figure 2D). Additionally, ALF-NS specific bacterial peptides showed prominent positive and negative correlation with the circulating DEM’s (233 metabolites) suggesting prominent association of met-proteins with the circulatory metabotype in ALF-NS (Figure 2E). Together our results show that ALF-NS have high diversity and any change in the ALF-NS specific bacterial peptide would bring a concordant change in the circulating metabolome and the clinical phenotype of ALF patients.
Integrome analysis of baseline plasma metabolites and metabolome in ALF
Plasma metabolome and meta-proteome analysis was subjected to integration using the Weighted Metabolite Correlation Network Analysis (WMCNA) and Weighted Meta-proteins Correlation Network Analysis (WMpCNA). ALF-NS specific correlation clusters were identified and subjected to pathway analysis. The W[M/Mp] CNA performs unsupervised clustering of biomolecules and group them into modules having similar expression patterns or possibly share common biological functions.
WMCNA: A total of 548 metabolites identified 8 modules (soft threshold>9, scale-free topology fit index>0.85; p<0.05; Figure 3A). Metabolites in the red, yellow and turquoise module were ALF-Non-Survivor specific modules (Figure 3A).
WMpCNA: A total of 579 bacterial peptides identified 5 modules (soft threshold>9, scale-free topology fit index>0.85; p<0.05; Figure 3B). The module trait relationship analysis identified Mp_Grey and Mp_Turquoise modules specific to ALF-NS (Supplementary Figure 12) and associated with LCA; Streptomyces, Actinomycetes, Streptococcus pneumoniae and others functionally linked with dysfunction of energy metabolism, lipid metabolism, amino acids-metabolism and others (Supplementary Figure 13).
Multi-modular Correlation Network: Mean module intensity of metabolites modules (WMCNA) and the meta-proteome module (WMpCNA) were cross-correlated to obtain multi-modular correlation network (MMCN). Correlation and network analysis performed in the study showed a striking association between meta-proteome modules and metabolic pathway (p<0.05, r2>0.5) (Figure 3C). Interestingly, bacterial phylum:LCA present in the Mp_turquoise, Mp_yellow and Mp_grey module showed direct correlation (R2>0.3, p<0.05) with the metabolic pathways histidine, arginine, tryptophan, bile acid and others (metabolic modules) (Figure 3D, Supplementary table 3). Additionally, Linear regression analysis was performed between metabolome and meta-proteome cluster along with metabolic pathway and individual meta-proteome phylum with the clinical complications occur in the patients during the course of hospitalization. Which showed significant association of meta-proteome cluster, metabolic pathway and meta-proteome phylum in the prediction of infection at baseline, sepsis, necrosis and multi-organ failure whereas the metabolome module showed striking association with AKI, and hepatic encephalopathy seen in ALF patients (Figure 3E, Figure F, Figure G, Supplementary Figure 14). Expression of the metabolic pathway’s members were significantly increased in ALF-NS as compared to Survivors (Supplementary figure 15). Finally, ALF-NS specific meta-proteomic and metabolic modules along with the top 5 metabolites showed direct correlation with the clinical profiles specifically with the severity of ALF (Figure 3F) suggesting that this increase in the pathogenic bacteria associated with metabolic pathway could be directly linked to the development of ALF severity. Together our results highlight an outstanding association of metaprotein phylum and metabolic pathway in the development of severity in ALF.
Validation of circulating indicator for early mortality using HRMS and machine learning
On the basis of FC, MDA, p-value and AUROC values top 5 meta-proteomic LCA (POD>80%) and metabolites were selected (Figure 4A, Supplementary Figure 16). Interestingly, linear regression of the top 5 metabolites and bacterial peptides with the clinical complication of ALF showed comparable results though top 5 metabolites showed significant increase in prediction of infection at baseline, necrosis and development of hepatic encephalopathy (Figure 4B). Of the selected top 5 meta-proteomic LCA and metabolite indicators, we decided to validated the metabolite indicators (easy to validate and requires very less per sample time). HRMS quantitation and validation in the test cohort (plasma and one drop blood samples; 160 ALF samples (ALF-S;53, ALF-NS;107)) showed significant increase in the top 5 metabolite indicators in ALF-NS as compared to ALF-S (Figure 4C). Interestingly, AUROC analysis along with the COX univariate and multivariate analysis showed significant association of Chenodeoxycholic acid (C02528, HR=1.7, p<0.05) with early mortality in ALF patients (Supplementary Figure 17). Interestingly, ALF patients with more than 10 FC plasma value for Chenodeoxycholic acid showed significant lower survival (Figure 4D). Next, top 5 metabolites: (4-(2-Aminophenyl)-2,4-dioxobutanoate, Carnosine, Chenodeoxycholic acid, Alanyl tyrosine and L-Tyrosine) were validated in the plasma and one drop blood samples of 160 ALF samples (ALF-S;53, ALF-NS;107) using machine learning approach. We used five ML algorithms (Linear Discriminant Analysis (LDA), k-Nearest Neighbours (kNN), Support Vector Machines (SVM) with a linear kernel, Classification and Regression Trees (CART) and Random Forest (RF) with which accuracy, sensitivity, specificity and p-value were calculated. In total, 30 trained and tested ML models were generated using 5 ML algorithms across 5 metabolites combination. Fourfold (outer) nested repeated (five times) tenfold (inner) cross-validation (With randomized stratified splitting) was used to train and test ML models and the hyperparameters of each algorithm were optimized. Accuracy and the kappa of model development (training cohort) were significant for all metabolites and comparison for the test cohort (plasma and paired one drop blood) (Figure 4E). Prediction capability of the identified top 5 metabolites in combination showed highest accuracy, sensitivity, specificity and p-value in plasma(100µl) but not so good in one drop blood (Figure-4F). Compared to individual metabolites, making the identified panel candidate indicators and the random forest model as the preferred ML algorithm for ALF severity detection in plasma and one drop blood, respectively.
In this pilot study we for the first time found out heterogenous changes in baseline metabolite and meta-proteome in Acute liver failure patients specially who are predisposed to early mortality. We found that there is a high circulation of chenodeoxycholic acid in ALF-NS. Chenodeoxycholic acid is major hydrophobic primary bile acid usually involved in the liver injury by activating NLRP3 inflammasome. Chenodeoxycholic acid also induce the Reactive oxygen species formation by various pathway. Specific increase of the chenodeoxycholic acid in baseline ALF patients might cause highest hazard particularly patients predisposed to early mortality (Figure 5A). From this pilot study we suggested that, targeting chenodeoxycholic acid might be ameliorate the acute liver injury and can hinder the inflammasome activation which ultimately improves the inflammation in liver.