Abstract

Introduction

Metabolic dysfunction-related hepatic condition (MDHC), previously identified as a Non-alcoholic fatty liver ailment (NAFLA), impacts approximately 25% of adults worldwide and emerged as a prevalent chronic hepatic ailment on a global scale [1, 2]. It is strongly associated with metabolic abnormalities like overweightness, cardiovascular risk, and type-2-diabetes mellitus (T2DM) [3, 4]. Its pathogenesis involves a multifaceted complex phenotypic and microscopic spectrum from basic fatty liver, inflammatory steatohepatitis, and fatty liver with necroinflammatory lesions, non-alcoholic steatohepatitis (NASH) the harbinger of the progressive fibrosis and cirrhotic liver and ultimately hepatocellular carcinoma (HCC), increasing the likelihoods of liver transplantation [5, 6]. With no specific pharmacotherapy approved yet directed to MDHC /NASH [7].

The intricate molecular mechanisms underlying the development of NASH continue to be largely elusive [8]. However, the widely embraced "multiple hit" hypothesis provides valuable insights [9]. This hypothesis suggests that NASH pathology involves a combination of various destructive events, such as liver fat buildup, insulin resistance, inflammation, impaired adipose tissue, altered gut microbiota, and genetic or epigenetic factors [10]. Inflammation is a key driver of NASH [11]. It is currently believed that innate immune responses are essential in maintaining hepatic inflammation in NASH [12]. Moreover, the incidence of hepatic fibrosis emerges as an indicator of NASH-related mortality [13].

The absence of a decisive therapy for NASH poses a significant challenge [14]. While lifestyle interventions, such as weight loss and managing comorbidities, remain the gold standard for NASH treatment in individuals with cirrhosis and obesity or overweightness, there is a pressing need for more effective solutions [15]. Various drugs have been tested or are being developed, targeting different aspects within and outside the liver, with the goal of improving steatohepatitis and fibrosis [16, 17]. The ongoing efforts in therapy development for MDHC and NASH primarily revolve around addressing metabolic imbalances, oxidative stress, and innate immunity [18].

The involvement of the gut microbiota in developing MDHC/NASH has been an active research entity [19]. shifts in gut microbiota composition, disruption of the gut barrier, and liver inflammation are key factors in the pathogenesis [20]. Thus, targeting the gut microbiota has been a potential approach for pharmacological treatments of MDHC [21]. The administration of probiotics has demonstrated promising results in reducing intra-hepatic triglyceride content and aspartate aminotransferase levels in NASH patients [22]. Furthermore, fecal microbiota transplantation has shown promising outcomes in alleviating steatohepatitis in animal models resulting in a significant reduction in the proinflammatory cytokines, intrahepatic lipid accumulation, as well as the NAS (MDHC Activity Score) [23–25].

Numerous drugs have undergone clinical trials for NASH treatment, but most studies have been limited in size and duration, making repeated biopsies impractical for trial participants [26, 27]. As a result, alternative methods for assessing drugs’ potential efficacy have become essential. Recent advancements have been made in utilizing non-invasive biomarkers in the initial phases of drug development [28] encompassing a range of metabolic parameters like lipid metabolism, insulin resistance, systemic inflammation, and hepatic lipogenesis, that are employed for confirming the therapeutic effects of NASH drugs in initial trials [29]. By integrating liver enzymes, ultrasound, and molecular biomarkers, researchers can effectively monitor treatment response in NASH patients, sparing them from the discomfort and potential complications associated with repeated invasive procedures [30].

The use of animal models in biological research and drug development is a long-standing practice due to several key reasons. Firstly, animal models offer significant physiological and anatomical similarities to humans, particularly mammals, enabling the study of complex immune responses and multi-tissue interactions that are crucial for understanding drug efficacy and safety. This similarity helps inform the design of clinical trials and mitigates the risk of adverse effects in human subjects. Rodents, in particular, share orthologous genes and physiological homologies with humans, making them especially suitable for studying human pathologies [31]. Secondly, the controlled experimental environment provided by animal models allows researchers to manipulate variables and isolate specific mechanisms, which can be challenging in studies involving humans due to genetic and environmental variability [32]. Finally, ethical and logistical constraints in human studies, especially those involving invasive procedures, underscore the practical importance of animal models for initial hypothesis testing and therapeutic evaluation prior to clinical trials [31]. Therefore, the models facilitate detailed mechanistic investigations into pathogenic processes such as NASH, including lipid metabolism and inflammatory signaling pathways, which are difficult to explore directly in human subjects.

Through extensive research on animal models, the authors have made significant strides in the search for a prophylactic drug for NASH. These investigations focused on various approaches, including targeting the liver-gut axis to modulate the gut microbiota using substances like kefir, the probiotic "Flora 20–14 Ultra Strength," the prebiotic inulin fiber "Greena," and synbiotics (the mixture of both probiotic and prebiotic). Additionally, anti-fibrotic strategies that directly addressed hepatic stellate cell inflammation, such as Mutaflor, were explored, moreover, rosavin was used to hinder inflammatory cell recruitment or block inflammatory signaling. To gain deeper insights, a panel of gene measurements involved in NASH pathogenesis and its epigenetic regulators were implemented. Alongside biochemical, histological, and immunohistochemistry analyses, these comprehensive evaluations provided valuable indications of the therapeutic candidates' biological actions and potential clinical benefits [33–36].

Machine learning (ML) allows autonomous learning from data that aided a variety of genomics studies [37–40]. Drug response prediction (DRP), a specialized application of ML, forecasts the phenotypic reactions of biological specimens using their molecular [41, 42]. These predictors often provide valuable insights [43]. Another significant research challenge is determining how a clinical drug will react, or how sensitive it will be, to a particular form of cancer [44]. Accurate prediction of clinical drug responses allows clinicians to comprehend variations in drug sensitivity outcomes among patients, ultimately reducing the time and costs involved in identifying effective drug candidates [45].

This study aimed to leverage machine learning (ML) algorithms to identify key genes that can reliably predict treatment responses in an animal model of NASH. By integrating biochemical, and molecular markers obtained through bioinformatics techniques, the research seeks to evaluate the efficacy of microbiome-targeted therapies and herbal drugs in managing NASH. The primary objective is to elucidate the genetic and epigenetic RNA networks involved in NASH progression and determine their potential as predictive markers for drug responses. Through rigorous validation of established biochemical-RNA signature networks in a larger animal cohort using bioinformatics and ML methodologies, the study aimed to enhance the accuracy of predicting drug efficacy in NASH treatment. Ultimately, this research underscores the significance of RNA-based predictors in advancing personalized treatment strategies for NASH.

Materials and methods

Experimental animal design

A hundred and thirty male Wistar rats 140–160 g were obtained from Ain Shams University's Scientific Research Centre “MASRI” for this study (Fig. 1A). The rats were maintained in a regulated environment with a 12-h light/dark cycle and a temperature of 20±2. Ethical approval (FMASU-R 111/2022, FWA 000017585) was obtained from Ain Shams Faculty of Medicine's Ethics Committee, following NIH guidelines for laboratory animal care and use. Rats were fed a high sucrose and high-fat (HSHF) diet (10% sucrose, 20% fat, supplemented with 68.75% standard chow, 1% cholesterol, and 0.25% bile salts) for 12 weeks to induce NASH [46]. After seven days of acclimating, rats were divided randomly into two groups: standard chow-fed (Normal group, n=10) and high-sucrose/high-fat-fed diet (n=120). The high-sucrose/high-fat-fed rats were subdivided into ten groups (n=12): NASH-12wk (fed the diet for 12 weeks), NASH-9wk (fed the diet for 9 weeks), FUS (administered 4×109 CFU bacteria/kg/day of FUS probiotic) [47], INU (administered 2 g/kg/day of INU prebiotic fibers) [48], FUS+INU (administered both FUS probiotic mixtures and INU prebiotic fibers), Mutaflor (administered 4×108 CFU of Escherichia coli Nissle 1917/kg bw/day) [49], Kefir (administered 4×107 CFU/ml ready-made kefir milk/kg bw/day) [50], and three Rosavin-treated groups (Ros-10, Ros-20, Ros-30) administered with three doses (10, 20, and 30 mg/kg/day) of herbal rosavin treatment [51].

Open in a separate window

Fig. 1

Schematic representation of the experimental protocol. A summarizes the study design and laboratory work. B illustrates the ML-based model findings

Microbiome-targeted therapies were administered via gastric gavage, while herbal rosavin treatment was administered intraperitoneally, both daily for 12 weeks. Prior to sacrifice, the rats were fasted for 12 h, anesthetized with urethane (1.2 g/kg) intraperitoneally, and blood samples were collected from the retro-orbital vein. The serum was separated by centrifugation (3000 rpm, 10 min, -20°C). Liver tissues were dissected, with one part used for RNA and protein analysis, and the rest fixed in 10% formalin for histopathology and immunohistochemistry. The study design workflow is depicted in Fig. 1.

Machine learning model building

In this study (Fig. 2), a comprehensive ML experiment was conducted utilizing a dataset comprising 130 samples with 2 different groups of signatures. The first group consisted of 20 molecular signature features (HSPD1, TNF, MMP14, ITGB1, YAP1, LATS1, NF2, FOXA2, TEAD2, LATS2, SOX11, SMAD4, AMOTL2, miRNA-6881-5P,miR-1205, miR-650,miR-6807-5p, lnc-SPARCL1-1:2, lncRNA SRD5A3-AS1, LncRNA RPARP AS-1), while the second group contained 14 biochemical features (ALT, AST, ALP, GGT, T. Bilirubin, D. Bilirubin, AFP, Albumin, TC, TG, HDL-C, LDL-c, IL-6, TGFB1).

Open in a separate window

Fig. 2

Sequential Steps of the ML Algorithms Illustrated in a Flowchart

The analysis began by conducting machine learning predictions individually for each group of features (Molecular and Biochemical) which allowed us to identify the best features within each group separately utilizing the robust Random Forest algorithm. Then, for the combined features six ML models were used as follows: Random Forest (RF), Logistic Regression (LR), K-nearest neighbor (KNN), Gradient Boosting, Decision tree (DT), and linear support vector machines (SVMs) to comprehensively identify the potential set of features that are highly predictive for the drug response with high accuracy compared to the invasive histopathological interventions, creating a total of three different modeling phases. was split into two sections: 80% was allocated for training purposes, while the remaining 20% served as a validation set for all trials. Additionally, various methods were utilized to mitigate overfitting, boost efficiency, and enhance the models' generalizability. First, Min–Max scaling was employed to normalize the dataset, adjusting features to a range between 0 and 1 to prevent any single feature from dominating the model. Next, outlier detection was conducted to ensure the dataset lacked significant outliers, preventing anomalous data points from biasing the model training. Lastly, Recursive Feature Elimination with Cross-Validation (RFECV) was implemented to determine the optimal subset of features for predicting NASH.

Results

Histological findings

The liver from the normal rats’ group displayed the normal structure of the brownish-red color that has a smooth and shiny morphology without any pathological features. However, livers from the NASH-induced groups or the group treated with herbal rosavin appeared enlarged, yellowish, and harder. Conversely, compared to those of the NASH groups, the livers of rats treated with Microbiome-targeted therapies looked redder and lighter. Histological examination of the livers in the NASH groups revealed severe parenchymal damage, ballooned hepatocytes, and micro and macrovesicular steatosis. Additionally, the NASH-12wk group exhibited substantial fibrosis. While the microbiome-targeted and herbal rosavin-treated groups indicated significant improvements in hepatocyte vacuolation, NASH score, inflammatory response, and fibrosis (Fig. 3) (Appendix Table A.3). Although the improvements in the Ros-10 group were less notable compared to other treated groups, these findings suggest that Microbiome-targeted therapies or herbal rosavin can potentially improve NASH progression. Overall, the group treated with a high dose of rosavin, and the FUS/INU combination group exhibited a significantly greater ameliorative effect compared to other treatments. The groups treated with probiotics (FUS, Mutaflor, and Kefir) and the Ros-20 group showed a similar pattern of improvement.

Open in a separate window

Fig. 3

Microbiome-targeted therapies and herbal rosavin impact on hepatic steatosis, inflammation, and fibrosis assessed using H&E staining. Histological examinations revealed (A) Normal group: normal architecture. B NASH-12 wk and (C) NASH-9wk groups: showed severe parenchymal damage, ballooned liver cells, and micro and macro-vesicular steatosis. Additionally, the NASH-12wk group showed substantial fibrosis. Notable histological improvements in steatosis, inflammation, fibrosis, and NASH score in (D) FUS, (E) INU, (F) FUS+INU, (G) Mutaflor, (H) Kefir, (I) Ros-10, (J) Ros-20 and (K) Ros-30 treated groups (Magnifications:×100)

Effect of microbiome-targeted therapies and herbal rosavin on the expression profile of RNA signatures

Both NASH groups experienced significant increases in the YAP1, HSPD1, TNF, TEAD2, ITGB1, FOXA2, MMP14, and SMAD4, as well as miR-1205, miR-6881-5p, miR-650, miR-6807-5p, and SPARCL1-1:2 expression. At the same time, LATS1, LATS2, AMOTL2, SOX11, NF2, LncRNAs-RPARP-AS1, and SRD5A3-AS1 were significantly suppressed in both NASH groups, in comparison to the Normal group. However, the daily treatment with either microbiome-targeted therapies or herbal rosavin slightly normalized the significant expression of hepatic RNA signatures (Fig. 4). The pronounced effect was observed in the ros-30 treated group and the FUS/INU combination group.

Open in a separate window

Fig. 4

Differential expression of RNA signatures in the hepatic tissues. Values mean±SD. *P<0.001 compared to the Normal group. ##P<0.001 and #P<0.05 compared to NASH-12wk group, aP<0.05 compared to NASH-9wk. One-way ANOVA followed by Tukey’s multiple comparison test

The ML results

Correlation analysis

The correlation matrix, as illustrated in Fig. 5, displayed the levels of correlation between pairs of features over the entire dataset for the molecular, biochemical, and both of them combined, providing insight into feature relationships. The correlation coefficient was contained in each cell of the matrix. Based on the connection with the target group feature, Tables ​Tables22 and ​and33 displayed the sorted features that had a high association.

Open in a separate window

Fig. 5

Heatmaps represent the correlation analysis. A Correlation matrix for the Molecular features. B Correlation matrix for the Biochemical features. C Correlation matrix for the combined features

Table 2

Sorted molecular features based on correlation with the target group feature

Molecular Features	Correlation Degree
ITGB1	0.82
YAP1	0.8
HSPD1	0.79
miR-6807-5p	0.78
miR-1205	0.77
TEAD2	0.77
SMAD4	0.77
FOXA2	0.76
Lnc-SPARCL1-1:2	0.74
MMP14	0.74
miR-650	0.74
Mir-9881-5p	0.72
TNF	0.71
NF2	0.7
LATS1	0.66
LATS2	0.66
RPARP-AS1	0.66
SRD5A3-AS1	0.68
AMOTL2	0.6
SOX11	0.54

Open in a separate window

Table 3

Sorted molecular features based on correlation with the target group feature

Biochemical Features	Correlation Degree
AFP	0.82
ALP	0.73
ALT	0.7
Albumin	0.69
IL-6	0.69
AST	0.67
GGT	0.65
TG	0.61
T. bilirubin	0.54
TGFB1	0.54
TC	0.39
D. bilirubin	0.34
HDL-C	0.32
LDL-C	0.31

Open in a separate window

Feature selection and PCA

For the Molecular signatures, the RFECV was adopted [70]. The method selected 2 out of 20 features which kept the same prediction accuracy level. Figure 6A depicts the performance of distinct features. The two selected features are HSPD1 and miR-6807-5p. Feature selection phases identified these 2 genes as being the two most important genes in the prediction of improvement. PCA was performed on sets of features before and after the feature selection phase. The PCA analysis shows how the prediction strength could be when using distinct sets of features. The set of selected features for the molecular signatures could keep the same level of prediction accuracy although 18 features were dropped. The selected set of features showed good discrimination between the two reconstructed categories of Improved case scores when examined by the PCA plot as shown in Fig. 6B&C.

Open in a separate window

Fig. 6

A Prediction performance for different numbers of features. PCA analysis for the features B before feature selection with 20 features, and C after feature selection (2 features)

However, the feature selection for the Biochemical Signatures did not show any improvement when features were reduced. Thus, the whole set of 14 features is used for predictions as they could together differentiate between predicted classes. The discriminative ability of the features is shown in the PCA (Fig. 7A). Also, the discriminative ability of the combined 34 features is shown in the following PCA (Fig. 7B).

Open in a separate window

Fig. 7

A PCA analysis for the biochemical features, B PCA analysis for the combined features

Confusion matrix of random forest classifier

The confusion matrix presented in Table 4 shows the correctness of the predictions for both classes, improved (HPS: 0→4) and Not-Improved (HPS: 5→8) based on the molecular data only. In the test set results, the Improved samples that were predicted correctly were 18 out of 20), while the Not-Improved samples were all correctly predicted (6 out of 6) based on the molecular data only. The correct portions of prediction are highlighted in bold in the above confusion matrix. While the confusion matrix presented in Table 5 shows the correctness of the predictions for both classes, improved (HPS: 0→4) and Not-Improved (HPS: 5→8) based on the biochemical data. In the test set results, the Improved samples that were predicted correctly were 20 out of 20), while the Not-Improved samples were all correctly predicted (6 out of 6). The correct portions of prediction are highlighted in bold in the above confusion matrix. Finally, the correctness of the predictions for both classes is presented in Table 6, Improved (HPS: 0→4) and Not-Improved (HPS: 5→8) based on both the molecular and biochemical features. In the test set results, the Improved samples that were predicted correctly are (20 out of 20), while the Not-Improved samples were all correctly predicted (6 out of 6). The correct portions of prediction are highlighted in bold in the above confusion matrix.

Table 4

Confusion matrix for the molecular data

Actual	Predicted
	Improved	Not-Improved
	18	2
	0	6

Open in a separate window

Table 5

Confusion matrix for the molecular data

Actual	Predicted
	Improved	Not-Improved
	20	0
	0	6

Open in a separate window

Table 6

Confusion matrix for both the molecular and biochemical data

Actual	Predicted
	Improved	Not-Improved
	20	0
	0	6

Open in a separate window

Area Under the Curve (AUC)

For the molecular features, the cross-validation results for both feature sets showed almost identical results. Also, the average area under the ROC curve was very high with the selected feature set predictions. Figure 8 shows the comparison between the full and selected feature set models when plotted on the AUC plot, displaying how accurate the prediction models were. Only a few numbers of instances were misclassified. The average accuracy achieved was 98%. The True positives refer to the improved samples that were predicted correctly. While the True negatives refer to the Not Improved samples that were correctly predicted. On the other hand, the cross-validation results of the biochemical features showed that the average accuracy achieved was 97%. The average AUC was 99% (Fig. 8). Additionally, the cross-validation results of both the molecular and the biochemical features showed that the average accuracy achieved was 98%. The average AUC was 100% (Fig. 8).

Open in a separate window

Fig. 8

A AUC for the Molecular features. B AUC for the Biochemical features. C AUC for the Combined features

Discussion

Despite the availability of several scoring systems for NASH severity classifications such as the NASH Clinical Research Network (NASH CRN) and the NAFLD Activity Score (NAS) that categorize NASH based on fibrosis stages or multifactorial scores including steatosis, cellular ballooning, and inflammation. These systems have limitations in capturing the full spectrum of cellular and stromal damage that drives the disease, and they are not sensitive enough to accurately indicate changes in the severity of the disease over time. Additionally, the currently available scoring systems illustrate only moderate-to-fair reproducibility [52]. Despite these drawbacks, the current framework for approving NASH therapies depends on pathological scoring methods performed manually [71, 72].

Intestinal microbes play a crucial role in metabolizing dietary compounds and producing bioactive molecules, impacting both physiological and pathological conditions that have the potential to promote HCCs and chronic liver diseases’ development and progression [73]. Accordingly, targeting the intestinal permeability as well as gut microbiome through interventions like fecal microbiota transplantation, diet, synbiotics, probiotics, and prebiotics shows promise as one of the innovative approaches for MAFLD/NASH treatment [74]. Animal studies have shown that probiotics have a profound effect on NASH, decreasing fatty acid production, metabolic endotoxemia, and inflammation. Wagnerberger et al. [75] proved that the administration of Lactobacillus casei strain Shirota probiotic are found to mitigate fructose-driven NAFLD by attenuating the hepatic Toll-like receptor 4 signaling cascade. A few of the research works have reported improvement in liver functions and histology with probiotic treatments, but additional randomized controlled trials are required to confirm their therapeutic effectiveness [76].

Recently, numerous research works have diagnosed MAFLD and/or NASH using ML focusing on particular avenues of the disease such as quantifying lipid content, staging by imaging, fibrotic status, or even using clinical practice records [77, 78]. Moreover, integrating ML algorithms into established bioinformatics could present tools that accurately predict or assist in the diagnosis of NASH, and identify therapeutic targets [79]. Sorino et al. [80] developed a web application on the basis of a Neural Network (NN) that utilized medical records to predict NAFLD, especially its absence. The application demonstrated a high level of accuracy of 77.0%, specificity of 100%, and sensitivity of 73%.

Therefore, the present research aims to use bioinformatics and ML techniques to develop a reliable model to predict drug efficacy in NASH. To achieve this, the previously established RNA-biochemical signatures were re-validated on a larger sample size of animals across different NASH therapy modules, involving microbiome-targeted therapy and herbal drugs. Our goal is to present a group of features that can assist in predicting the response to different therapeutic modalities for NASH. Herein, various public microarray databases were utilized for selecting functionally linked RNAs and biochemical signatures involved in pathogenic mechanisms driving the NASH pathogenesis, at genetic, epigenetic, and proteomic levels. Consequently, mRNAs (TEAD2, YAP1, LATS1, LATS2, NF2, FOXA2, AMOTL2, SOX11, SMAD4, MMP14, HSPD1, ITGB1, TNF), miRNAs (miR-1205, miR-6881-5p, miR-650, miR-6807-5p), lncRNAs (RPARP-AS1, SPARCL1-1:2, and SRD5A3-AS1) and target effectors (IL-6 and TGF-β1) were retrieved in addition to the serum liver enzymes and lipid profile biomarkers (ALT, AST, ALP, Alb, GGT, T. Bilirubin, D. Bilirubin, TC, TG, and LDL-C) and the AFP.

In the current study, the treatment of animals with either microbiome-targeted therapies or herbal rosavin for 12 weeks diminished the histopathological and metabolic abnormalities induced in the NASH groups. They decreased the degree of NASH score, the hepatic level of IL-6 and TGF-β1 as well as improvement in both liver biomarkers and lipid panel. The hepatoprotective effect of either microbiome-targeted therapies or herbal rosavin may be attributed to diminished pro-inflammatory cytokines’ production, inhibition of inflammatory signaling, antioxidant signaling pathways’ activation, enhanced cellular immunity, and decreased tissue apoptosis [34, 36].

Regarding mRNA signatures, there were significant increases in the hepatic expression of YAP1, HSPD1, TNF, TEAD2, ITGB1, FOXA2, MMP14, and SMAD4, with significant decreases in LATS1, LATS2, AMOTL2, SOX11, and NF2 in both NASH groups, as against the Normal group. The daily treatment with either microbiome-targeted therapies or herbal rosavin modulated the expression of hepatic mRNA signatures by ameliorating the signaling pathways contributing to NASH progression including Hippo signaling, TGF-β signaling, TNF signaling, and apoptosis. The Hippo signaling pathway is a highly conserved regulatory system that controls the hepatic size, liver regeneration, and hepatic cell differentiation and survival [81]. YAP serves as a pivotal downstream mediator of the Hippo signaling pathway. Upon dephosphorylation, it translocates to the nucleus where it forms a complex with the transcription factor TEAD, facilitating essential cell proliferation during liver regeneration. However, prolonged activation of YAP can lead to the development of malignant tumors. NF2 and angiomotin (AMOT), situated upstream in the Hippo pathway, contribute to YAP phosphorylation via LATS1/2 during cell fate determination and developmental processes. SMAD4, a member of the SMAD protein family, exhibits dual roles as both a tumor suppressor and an oncogene in cancers such as hepatocellular carcinoma. FOXA2 expressed prominently during early fetal liver development but absent in mature hepatocytes or bile duct cells, represents a potential marker for hepatic cell proliferation [81]. TGF-β signaling mechanism performs a prominent part in the maintenance of normal homeostasis of the liver and is involved in different stages of liver fibrosis, which further contributes to the progression of NAS [82]. TNF signaling is one of the primary regulators of inflammation. It has decisive functions in tissue homeostasis, regulating cell death, and inflammation, which suggests a potential interlinkage between the severity of NASH and its distinct levels [83]. Evidence shows that activated Kupffer cells can worsen liver injury by activating TGF-β and TNF signaling, leading to the production of new extracellular matrix (ECM) proteins like collagen, which contribute to liver fibrosis. This process activates matrix metalloproteinases (MMPs) such as MMP14, which induce the expression of hepatic vascular adhesion molecules. Activated integrins, such as ITGB1, on immune cells interact with adhesion molecules to promote cell detachment from the ECM. Dying hepatocytes release stress signals that affect nearby hepatic cells, triggering inflammation and fibrosis [36]. Therefore, the modulatory effect of microbiome-targeted therapies or herbal rosavin on mRNA signatures may result from inhibition of the Hippo, TGF, and TNF signaling pathways, and activation of antioxidant signaling pathways. In the current study, the combination of probiotics and prebiotics is more ameliorative than administering them individually because prebiotics serve as a food source for probiotics, enhancing their growth and activity. This synergistic effect can lead to a more robust and effective modulation of the gut microbiome, resulting in improved NASH progression. Furthermore, the combination group and the high-dose rosavin-treated group exhibited a similar modulatory effect. This may be owing to their action on similar pathways or mechanisms.

In the current study, a 12-week duration was chosen based on several considerations, including previous studies, the nature of the experimental model, and practical constraints. Many studies have successfully employed a 12-week duration to evaluate the efficacy of treatments in similar experimental models of metabolic disorders [84–87] finding this period sufficient to observe significant physiological and metabolic changes and providing meaningful insights into the immediate effectiveness of the therapies. Additionally, the treatments applied in our study have previously been tested in human subjects, showing prospective effects and demonstrating safety and efficacy over extended periods in clinical settings (Table S3). These prior human studies provide a strong foundation for our choice of treatments.

Additionally, the outcomes revealed a significant upsurge in the hepatic expression of miR-1205, miR-6881-5p, miR-650, miR-6807-5p, and SPARCL1-1:2 in the two NASH groups (NASH-12wk and NASH-9wk) compared to the Normal group. Moreover, there was a notable decrease in the levels. The expression of these miRNA and lncRNA signatures has been reversed with microbiome-targeted and herbal rosavin therapies, along with diet manipulation. The functional enrichment of these epigenetic signatures showed their involvement in NASH-related pathogenic mechanisms including TGF-β and Hippo signaling. This can explain their differential expression level among the NASH groups and the other treated groups. The important part of miRNAs and lncRNAs in developing NASH are widely recognized. miRNAs regulate gene expression by binding to specific mRNA sequences, resulting in either the degradation of mRNA or the suppression of translation [88]. Meanwhile, lncRNAs act as miRNA sponges, preventing the degradation of mRNAs targeted by miRNAs, thereby preserving their stability [89]. These molecules are involved in the progression of NASH [90], owing to their strong association with the inflammatory and fibrotic mechanisms and pathways that drive the disease pathogenesis [91]. Herein, different machine-learning models; (RF, LR, KNN, GBoost, DT, and SVM) were used to build models that can differentiate between our two classes (HPS Improved), or not (HPS Not-Improved). The data was sectioned into two partitions: 80% for training and 20% for validation. Also, adjustments were made to balance samples belonging to different HPS levels by using up sampling to increase the number of cases in specific HPS levels. The model's effectiveness was assessed using various performance evaluation metrics, including accuracy, precision, and recall, in addition to the analysis of the confusion matrix. For the combined molecular and biochemical features, GBoost was the best model outperforming other algorithms in terms of all evaluation metrics achieving an accuracy of 98%. However, after a gradual decrement of features, the results showed a better performance when using the Random Forest classifier, achieving an accuracy of 98.4%. The top selected set of molecular features was YAP1, LATS1, NF2, SRD5A3-AS1, FOXA2, TEAD2, miR-650, MMP14, ITGB1, miR-6881-5P, while the biochemicals were ALT, ALP, T. Bilirubin, AFP, TG, LDL-C. This variation of the result may suggest the dynamic nature of machine learning model selection and its sensitivity to feature composition [92]. The principal component analysis (PCA) before and after the feature selection phase showed that the 9 selected features could keep the same level of prediction accuracy and showed better discrimination between classes (NASH vs. Non-NASH). Ultimately, the RF classifier containing 9 features showed very high accuracy and the average area under the ROC curve with the selected 9 features was 0.99 which means that the 9 features have superior diagnostic characteristics for NASH disease (Fig. 1B).

In this study, different machine learning methods were tried to compare performance over our dataset. Iterative experiments with optimizations showed that Random Forest is performing well over other methods. The set of algorithms was selected based on either the simplicity or interpretability of the results. In addition, other methods were used for comparison and evaluation of performance.

The process began with Decision Tree methods, which highlight important features that achieve higher information gain during the splitting of tree nodes. Decision Trees are easily interpretable models able to capture non-linear relationships and feature importance. Random Forest provides the same ability while being more resistant to overfitting by using an ensemble of trees. Knowing the important features is crucial to help reduce experimental efforts. RF handles high-dimensional data and is robust to outliers and noisy data.

Other methods used for comparing performance included: Logistic Regression K-Nearest Neighbors (KNN), SVM, and lastly Gradient Boosting which is an ensemble method that sequentially builds models to correct the mistakes of previous models and achieves high predictive accuracy. Although it has a high predictive ability, it is more prone to overfitting and less effective in deciding feature importance when compared to Random Forest.

To tackle concerns about overfitting, a multi-step strategy was developed to strengthen our model's ability to generalize effectively. Initially, The data was rigorously preprocessed, cleansing it of any null values or duplicates to ensure integrity and remove potential biases. RFECV was then utilized to identify the most crucial features of the model. This method systematically prunes less relevant features, homing in on those that carry the most discriminatory power and reducing the risk of overfitting. In summary, ensuring the model's generalizability was a priority. The approach encompassed thorough data preprocessing, feature selection via RFECV, and the application of diverse algorithms. These combined efforts are believed to enhance the robustness and reliability of the findings.

AST, ALT, and ALP are typically utilized as indicators of liver injury, but they lack specificity in diagnosing NASH [11, 93]. Ma et al. [94] developed a machine learning predictive model for NAFLD, by selecting 5 features such as weight, TG levels, ALT, GGT, and serum uric acid levels that provided the most accurate predictions in their Bayesian network model achieving an accuracy rate of 83%, a specificity of 0.878, and a sensitivity of 0.675. By leveraging extensive electronic health records from the United States, Fialoke et al. [95] established a machine-learning model that was capable of accurately predicting NASH. This model incorporated longitudinal data of key biomarkers such as AST, ALT, AST/ALT ratio, and platelet count, along with basic demographic information and diabetes status (AUC=0.88).

Costello et al. [96] emphasized predicting drug response based using genomic, epigenomic, and proteomic datasets collected from human breast cancer cell lines. Ali and Aittokallio [97] proposed a Bayesian multi-task learning model on a dataset of 53 breast cancer cell lines that integrates data from multiple omics sources including RNA-sequencing, point mutations, somatic DNA copy number variation, protein abundance, transcript expression, and DNA methylation to predict drug response reporting that this multi-task learning model allows the model to learn a more inclusive view of the cancer cells and to make more accurate predictions of drug response suggesting the machine learning a promising tool for precision oncology.

Stetson et al. [98] reported three extensive pharmacogenomic investigations, involving the screening of anticancer compounds across more than 1,000 diverse human cancer cell lines. The datasets were combined for creating and authenticating multi-omics predictors for drug responses. The study makes a comparison of drug response models constructed via various techniques, including a support vector machine, penalized linear regression model, as well as random forest. The reliability and accuracy of each drug response model were rigorously evaluated through cross-validation across three independent datasets. Notably, in comparison to the commonly used elastic net regression as well as SVM, the RF yielded more robust and precise prediction signatures. The resulting drug response signatures can be used to stratify patients into treatment groups based on their tumor biology, offering two significant advantages: facilitating the repositioning and repurposing of existing anticancer therapies and expediting the process of bringing preclinical drugs to market.

Conclusion

The objective was to develop an algorithm to evaluate the efficacy of conventional liver blood tests combined with a panel of interacting mRNAs-miRNAs-LncRNAs in identifying a specific set of features that can reliably predict drug efficacy in Nonalcoholic Steatohepatitis (NASH). Utilizing noninvasive biomarkers for predicting treatment response reduces the need for invasive procedures like liver biopsies, thereby minimizing patient discomfort and associated risks. This approach can drive the development of new therapeutic targets and strategies, potentially leading to more effective treatments for NASH.

Acknowledgements

Abbreviations

Authors’ contributions

M. M., & S. H., R. K. & A. K.; Shared in Study design, methodology, statistical analysis, and drafting and revising the manuscript. N. E., L. M. F., M. B. A., M. M., M. S. E. & M.  A. Shared in shared in methodology and drafting the manuscript. All authors have read and agreed to the published version of the manuscript.

Funding

Open access funding provided by The Science, Technology & Innovation Funding Authority (STDF) in cooperation with The Egyptian Knowledge Bank (EKB). This work was supported by Mounir Aramnious Research Center( M.A.R.C),2022-1.

Availability of data and materials

All data generated during this study are included in this article. The source code for rebuilding the models used in the study can be found at: https://github.com/ibrahimzb/MLModels_NASHAnimal Supplementary Table A4.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Footnotes

References