Mitochondrial mt12361A>G increased risk of metabolic dysfunction-associated steatotic liver disease among non-diabetes

Abstract

BACKGROUND

Insulin resistance, lipotoxicity, and mitochondrial dysfunction contribute to the pathogenesis of metabolic dysfunction-associated steatotic liver disease (MASLD). Mitochondrial dysfunction impairs oxidative phosphorylation and increases reactive oxygen species production, leading to steatohepatitis and hepatic fibrosis. Artificial intelligence (AI) is a potent tool for disease diagnosis and risk stratification.

AIM

To investigate mitochondrial DNA polymorphisms in susceptibility to MASLD and establish an AI model for MASLD screening.

METHODS

Multiplex polymerase chain reaction was performed to comprehensively genotype 82 mitochondrial DNA variants in the screening dataset (n = 264). The significant mitochondrial single nucleotide polymorphism was validated in an independent cohort (n = 1046) using the Taqman^® allelic discrimination assay. Random forest, eXtreme gradient boosting, Naive Bayes, and logistic regression algorithms were employed to construct an AI model for MASLD.

RESULTS

In the screening dataset, only mt12361A>G was significantly associated with MASLD. mt12361A>G showed borderline significance in MASLD patients with 2-3 cardiometabolic traits compared with controls in the validation dataset (P = 0.055). Multivariate regression analysis confirmed that mt12361A>G was an independent risk factor of MASLD [odds ratio (OR) = 2.54, 95% confidence interval (CI): 1.19-5.43, P = 0.016]. The genetic effect of mt12361A>G was significant in the non-diabetic group but not in the diabetic group. mt12361G carriers had a 2.8-fold higher risk than A carriers in the non-diabetic group (OR = 2.80, 95%CI: 1.22-6.41, P = 0.015). By integrating clinical features and mt12361A>G, random forest outperformed other algorithms in detecting MASLD [training area under the receiver operating characteristic curve (AUROC) = 1.000, validation AUROC = 0.876].

CONCLUSION

The mt12361A>G variant increased the severity of MASLD in non-diabetic patients. AI supports the screening and management of MASLD in primary care settings.

Key Words: Metabolic dysfunction-associated steatotic liver disease; Mitochondrial DNA; Artificial intelligence; Machine learning; Algorithm

Core Tip: Metabolic dysfunction-associated steatotic liver disease (MASLD) is a systemic metabolic disorder affecting the liver-kidney-heart axis. Mitochondrial dysfunction drives the progression of liver steatosis into steatohepatitis and hepatic fibrosis. The mitochondrial mt12361A>G variant increased the severity of MASLD in the non-diabetic group but not in the diabetic group. By integrating genomics and machine learning, we established a random forest model to screen for MASLD with an area under the receiver operating characteristic curve of 1.000 in the training dataset and 0.876 in the validation dataset. The artificial intelligence model supports the prevention, screening, and management of MASLD in primary care.

INTRODUCTION

Metabolic dysfunction-associated steatotic liver disease (MASLD), previously termed non-alcoholic fatty liver disease (NAFLD), is defined as steatotic liver disease accompanied by one or more cardiometabolic risk factors[1]. The global prevalence of MASLD is estimated to be between 20% and 45% and steadily increases in obesity, type 2 diabetes, and metabolic syndrome[2,3]. Approximately 10%-20% of MASLD cases progress to metabolic dysfunction-associated steatohepatitis (MASH)[4]. MASH is characterized by steatosis with hepatocyte ballooning and fibrosis[5], which may progress to liver cirrhosis and hepatocellular carcinoma[6]. MASLD is a multisystem disease that not only elicits liver-related diseases but also accompanies extrahepatic comorbidities, such as type 2 diabetes mellitus, cardiovascular disease, chronic kidney disease, and extrahepatic malignancy[7].

The present study confirmed that the discrepancy between MASLD and NAFLD is minimal, suggesting that previous NAFLD studies remain valid under the new definition of MASLD[8]. “Multiple hits” including insulin resistance, lipotoxicity, mitochondrial dysfunction, cytokines, genes, and gut dysbiosis contribute to the pathogenesis of MASLD[7,9]. The role of mitochondrial dysfunction in MASLD has recently garnered considerable attention. Mitochondria serve as the primary sites of β-oxidation, a catabolic process of free fatty acids. Insulin resistance triggers hepatocytes susceptible to oxidative stress and compromises mitochondrial function[9,10]. Mitochondrial dysfunction causes disruption of respiratory chain activity, adenosine triphosphate (ATP) depletion, and overproduction of reactive oxygen species (ROS)[11]. Impaired mitochondrial β-oxidation results in the accumulation of fatty acids, leading to lipotoxicity and triggering pro-inflammatory signaling pathways[12,13]. Oxidative stress causes a decline in mitochondrial membrane potential and promotes programmed cell death, which aggravates hepatocyte damage and fibrosis[14,15]. On the other hand, hepatic mitochondria are flexible and can adapt to the ambient metabolic status to prevent excessive fat accumulation and lipotoxicity[16,17]. Mitochondrial capacity decline in steatohepatitis with type 2 diabetes[18]. Increased oxidative stress and loss of mitochondrial adaptation drive the progression of steatosis to steatohepatitis, fibrosis, and cirrhosis[19]. Mitochondria-targeted agents represent innovative approaches for treating MASLD[20]. MitoQ helps to preserve mitochondrial function and integrity by scavenging harmful ROS and reducing oxidative stress[21]. Inhibition of mitochondrial DNA (mtDNA) transcription acts to reverse diet-induced hepatosteatosis and obesity in mice[22].

The liver is an organ enriched with high-density mitochondria. Each hepatocyte contains approximately 500-4000 mitochondria that constitute 18% of the cell volume[23]. Mitochondria are dynamic organelles that undergo continuous fusion and fission. Mitochondrial biogenesis is crucial for maintaining mitochondrial functional integrity and mtDNA inheritance[24]. Compared with nuclear DNA, mtDNA is more vulnerable to damage due to exposure to excess ROS, lack of histones, and limited DNA repair capability[25]. Defects in mitophagy prevent the efficient removal of damaged mitochondria, exacerbating hepatocyte injuries[26]. Mutations and damage to mtDNA further disrupt mitochondrial function and contribute to metabolic dysfunction in MASLD[27]. Accumulated mtDNA damage increases the severity of fatty liver disease[28]. Patients with NAFLD have a higher mtDNA mutation rate and lower mtDNA copy number than healthy controls[27]. Previous studies have reported that mt12361A>G, mt16318C>A, and mt16129AA are associated with steatohepatitis and advanced liver fibrosis[29,30]. A genome-wide meta-analysis highlighted 17 genes related to mitochondria, cholesterol, and de novo lipogenesis that are implicated in MASLD predisposition[31].

Artificial intelligence (AI) is a powerful tool for big data analysis. Machine learning is a branch of AI that can learn hidden patterns from input data and gradually improve the accuracy[32,33]. Big data from clinical, laboratory, and multi-omics often contain complex, non-linear relationships that are difficult to analyze using traditional statistical methods[34]. Machine learning facilitates disease diagnosis and risk stratification[35]. Algorithms such as Naive Bayes, random forest, gradient boosting, and logistic regression have been employed for disease diagnosis with higher accuracy than conventional statistical methods[36,37].

The mechanisms underlying MASLD are not fully understood. Given that MASLD is a systemic metabolic disease beyond the liver, raising awareness about MASLD among all healthcare professionals is mandatory. Nonetheless, primary medical institutes may lack imaging facilities, and not all clinicians are familiar with ultrasound diagnosis. This study aimed to investigate the role of mtDNA polymorphisms in susceptibility to MASLD. By integrating genomics and machine learning, we aimed to establish an AI model to screen for MASLD, without relying on image diagnosis.

MATERIALS AND METHODS

Subjects

A total of 1310 subjects (screening dataset, n = 264; validation dataset, n = 1046) were enrolled from Kaohsiung Medical University Hospital between 2006 and 2008. The participants in the screening and validation datasets were hospital- and community-based populations, respectively. The study participants were outpatients or people participating in a community health screening program. The inclusion criterion was age > 20 years. The exclusion criteria included viral hepatitis, alcoholic liver disease, and any other specific etiology of steatosis. A detailed medical history, laboratory assessment, and abdominal ultrasonography were collected. The presence of hepatic steatosis was diagnosed by ultrasound sonography. Hepatologists trained at the same institution performed the abdominal ultrasonography (Toshiba SSA-250, Tokyo, Japan) to reduce interobserver variability. This study followed the ethical guidelines of the Declaration of Helsinki and was approved by the Ethics Committee of Kaohsiung Medical University Hospital [No. KMUHIRB-E(II)-20230129] and waiver of subject informed consent.

MASLD diagnosis

MASLD is defined as hepatic steatosis in the presence of at least one of the five cardiometabolic risk factors: (1) Body mass index ≥ 23 kg/m², or waist circumference ≥ 90 cm in men or ≥ 80 cm in women; (2) Glycosylated hemoglobin ≥ 5.7%, fasting plasma glucose ≥ 100 mg/dL, postprandial 2 hours plasma glucose ≥ 140 mg/dL, type 2 diabetes, or under anti-diabetic treatment; (3) Plasma triglycerides ≥ 150 mg/dL or under lipid-lowering treatment; (4) High-density lipoprotein ≤ 40 mg/dL for men or ≤ 50 mg/dL for women, or under lipid-lowering treatment; and (5) Blood pressure ≥ 130/85 mmHg or use of anti-hypertensive treatment[1,38].

Mitochondrial single nucleotide polymorphism selection and haplogroup classification

We selected 82 disease- or haplogroup-specific single nucleotide polymorphisms (SNPs) distributed along the mRNA, tRNA, rRNA, and D-loop regions of the mitochondrial genome using Mitomap (http://www.mitomap.org/) and the Human Mitochondrial Genome Database (http://www.genpat.uu.se/mtDB/) (Supplementary Table 1). Forty-eight of these 82 mitochondrial SNPs (mtSNPs) were previously reported in patients with diabetes, cancers, myopathy, Alzheimer’s disease, Parkinson’s disease, and various mitochondrial hereditary diseases. Thirty-four haplogroup-specific mtSNPs were used to identify 19 haplogroups in the Asian population (A, B, C, D4, D5, E, F, G, M7a, M7b, M8a, M9, M10, M27, N9a, R9, R11, R30, and Z) (Supplementary Tables 2 and 3).

Genotyping

DNA was extracted from peripheral blood using a DNeasy blood kit following the manufacturer’s instructions (Qiagen, Düsseldorf, Germany). During the screening stage, multiplex polymerase chain reaction (PCR) and allele-specific oligonucleotide dot blot hybridization were used for mtDNA variant genotyping. Five sets of multiplex PCR and primers were designed to cover the mtDNA regions containing these 82 mtSNPs[39]. Each 100 μL of PCR mixture contained 10× PCR buffer II (Applied Biosystems), 25 mmol/L magnesium chloride, 5 U/μL of Taq polymerase, 10 μmol/L of each primer, 8 mmol/L of deoxy-ribonucleoside triphosphate, and 50 ng/μL genomic DNA. The reaction mixture was denatured at 94 °C for 2 minutes followed by 30 cycles of 1 minute of denaturation at 94 °C, 1 minute of reannealing at 55 °C, and 2 minutes of extension at 72 °C. PCR was completed by a final extension at 72 °C for 5 minutes[39]. 1 μL of the PCR product was spotted onto the Zeta-Probe Blotting Membrane. Allele-specific oligonucleotide dot blot hybridization was performed according to the previously published protocols[39-41].

The Taqman^® allelic discrimination assay was used for genotyping the significant mtSNP in the validation dataset. The probe and primer sequences of mt12361A>G are as follows: TaqMan MGB probes: VIC-CACTACTATAACCACCCTAAC (wild type); FAM-ACTACTATAACCGCCCTAAC (variant); Forward primer: 5’-GGTGCAACTCCAAATAAAAGTAATAACCA-3’; Reverse primer: 5’-GTGGATGCGACAATGGATTTTACAT-3’.

Genotyping was performed according to the manufacturer’s instructions. The following amplification procedure was applied: 95 °C for 5 minutes, 40 cycles of 95 °C for 30 seconds, 60 °C for 30 seconds, and 72 °C for 1 minute. The reactions were performed in 96-well microplates using an ABI 9700 Thermal Cycler. Fluorescence was analyzed using the SDS software on an ABI 7500 real-time PCR System (version 1.2.3, Applied Biosystems, Foster City).

Machine learning analysis

A total of 1310 participants were randomly assigned to the training (70%) and validation (30%) datasets. Seventeen features, including sociodemographic variables, biochemical data, comorbidities, and mt12361A>G variants, were input into the machine learning models (Supplementary Table 4). The employed algorithms in the detection of MASLD included random forest, eXtreme gradient boosting (XGBoost), Naive Bayes, and logistic regression. The performance of the classifiers was assessed using confusion matrix, the area under the receiver operating characteristic curve (AUROC), precision-recall curve, and F1-score[42]. A precision-recall curve closer to the upper-right corner indicates higher accuracy. The F1-score, a harmonic mean of precision and recall, ranged from 0 to 1, with scores approaching 1 indicating better performance. Machine learning analysis was performed using Orange Data Mining software based on Python. (v.3.34.0, University of Ljubljana, Slovenia). The widget settings of the Orange Data Mining are presented in Supplementary Figure 1. Hyperparameter tuning for each algorithm is presented in Supplementary Table 5.

Statistical analysis

Student’s t-test was used to compare continuous variables. Categorical variables were assessed using the χ² or Fisher’s exact test. A trend test was conducted to determine the dose-dependent relationship between increasing cardiometabolic factors. Significant variables in the univariate analysis (P < 0.05) were chosen to construct the multivariate logistic regression model. Multivariate logistic regression analysis was used to explore the independent risk factors for MASLD. The Hosmer-Lemeshow test was used to assess the model’s goodness-of-fit. The statistical significance was defined as a two-tailed P value < 0.05. The statistical significance of multiple tests was adjusted using the Bonferroni correction. Statistical analyses were performed using the statistical package for social science program (version 26.0, statistical product and service solutions Inc., Chicago, IL, United States). The statistical methods used in this study were reviewed by Dr. Tsai PC of Kaohsiung Medical University.

RESULTS

Baseline demographics

Table 1 presents the demographic characteristics of participants. The prevalence of MASLD was 50.4% and 42.4% in the screening and validation datasets, respectively. The proportion of males was significantly higher in the screening dataset than in the validation dataset (59.5% vs 44.1%, P = 8.4 × 10^-6). Additionally, waist circumference was significantly greater in the screening dataset than in the validation dataset (83.2 ± 11.2 cm vs 81.2 ± 10.2 cm, P = 0.017). Cholesterol levels were significantly lower in the screening dataset than in the validation dataset. Aspartate aminotransferase (AST) and alanine aminotransferase (ALT) upper limit of normal ratios were considerably higher in the screening dataset than in the validation dataset.

Table 1 Baseline demographics, mean ± SD/n (%).

Variables	Total (n = 1310)	Screening (n = 264)	Validation (n = 1046)	P value
MASLD	577 (44.0)	133 (50.4)	444 (42.4)
Age (years)	48.6 ± 13.1	47.8 ± 12.3	48.8 ± 13.3	0.238
Sex (male)	617 (47.2)	157 (59.5)	460 (44.1)	8.4 × 10-6
Body mass index (kg/m2)	24.5 ± 3.8	24.3 ± 3.9	24.5 ± 3.8	0.545
Waist circumference (cm)	81.6 ± 10.5	83.2 ± 11.2	81.2 ± 10.2	0.017
Hip circumference (cm)	96.2 ± 7.4	96.5 ± 7.7	96.1 ± 7.4	0.483
Fasting glucose (mg/dL)	95.1 ± 29.4	98.1 ± 26.9	94.3 ± 30.0	0.063
Cholesterol (mg/dL)	196.7 ± 38.6	185.5 ± 31.3	199.5 ± 39.7	1.6 × 10-9
Triglyceride (mg/dL)	130.8 ± 97.1	127.4 ± 77.3	131.6 ± 101.5	0.530
High density lipoprotein (mg/dL)	56.7 ± 14.2	55.4 ± 13.1	57.0 ± 14.5	0.086
Low density lipoprotein (mg/dL)	126.1 ± 35.6	125.0 ± 35.1	126.4 ± 35.7	0.569
AST ULN ratio	0.58 ± 0.29	0.73 ± 0.27	0.54 ± 0.28	1.1 × 10-22
ALT ULN ratio	0.58 ± 0.41	0.73 ± 0.52	0.55 ± 0.37	9.0 × 10-8
Diabetes mellitus	74 (5.8)	18 (6.9)	56 (5.6)	0.410
Hypertension	213 (16.7)	42 (16.2)	171 (16.9)	0.798
Hyperlipidemia	142 (12.0)	34 (13.4)	108 (11.7)	0.455

MASLD: Metabolic dysfunction-associated steatotic liver disease; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; ULN: Upper limit of normal.

Screening for mtDNA variants

Among the 82 mtSNPs in the screening dataset (MASLD, n = 133; control, n = 131), none achieved statistical significance after Bonferroni adjustment. The severity of MASLD was further stratified based on the number of cardiometabolic traits. The frequency of the G allele of mt12361 was significantly higher in the patients with severe degree MASLD (≥ 3 cardiometabolic traits) than in the healthy controls (15.1% vs 1.5%, P = 3.0 × 10^-4) (Supplementary Table 1). Approximately 98.5% of participants were successfully classified into 19 common Asian mitochondrial haplogroups. No mitochondrial haplogroups were significantly correlated with MASLD in the screening dataset (Supplementary Table 2).

Validation of the mt12361A>G variants

Table 2 shows the correlation between mt12361A>G and steatotic liver disease with various cardiometabolic traits. In the screening dataset, the frequency of the G allele among patients with 0-1, 2-3, and 4-5 cardiometabolic traits was 1.3%, 10.9%, and 14.6%, respectively (trend test P = 1.5 × 10^-4). The mt12361G carriers with more than 4 cardiometabolic factors exhibited a 13.37-fold risk of developing MASLD than the controls (P = 0.001). Similarly, individuals with 2-3 cardiometabolic traits carrying the mt12361 G allele had a 9.58-fold higher risk of MASLD than the controls (P = 0.003).

Table 2 The association between mt12361 and steatotic liver disease.

CMRF	Screening					Validation					All
	mt12361		χ2		Trend test	mt12361		χ2		Trend test	mt12361		χ2		Trend test
	A (%)	G (%)	P value	OR (95%CI)	P value	A (%)	G (%)	P value	OR (95%CI)	P value	A (%)	G (%)	P value	OR (95%CI)	P value
4-5	35 (85.4)	6 (14.6)	0.001	13.37 (2.59-69.05)	1.5 × 10-4	102 (97.1)	3 (2.9)	0.445	1.60 (0.44-5.76)	0.165	137 (93.8)	9 (6.2)	0.001	3.79 (1.61-8.93)	1.7 × 10-4
2-3	57 (89.1)	7 (10.9)	0.003	9.58 (1.93-47.47)		212 (95.9)	9 (4.1)	0.055	2.31 (0.96-5.55)		269 (94.4)	16 (5.6)	4.6 × 10-4	3.43 (1.65-7.13)
0-1	156 (98.7)	2 (1.3)	Ref.	Ref.		652 (98.2)	12 (1.8)	Ref.	Ref.		808 (98.3)	14 (1.7)	Ref.	Ref.
3-5	62 (84.9)	11 (15.1)	3.0 × 10-4	11.36 (2.44-52.80)	1.3 × 10-4	198 (97.5)	5 (2.5)	0.637	1.29 (0.44-3.77)	0.441	260 (94.2)	16 (5.8)	0.001	3.27 (1.55-6.89)	0.001
1-2	58 (96.7)	2 (3.3)	0.592	2.21 (0.30-16.05)		205 (96.2)	8 (3.8)	0.135	2.00 (0.79-5.04)		263 (96.3)	10 (3.7)	0.093	2.02 (0.88-4.67)
0	128 (98.5)	2 (1.5)	Ref.	Ref.		563 (98.1)	11 (1.9)	Ref.	Ref.		691 (98.2)	13 (1.8)	Ref.	Ref.
3-5	62 (84.9)	11 (15.1)	2.1 × 10-4	8.25 (2.54-26.85)	5.1 × 10-5	198 (97.5)	5 (2.5)	0.968	1.02 (0.38-2.77)	0.968	260 (94.2)	16 (5.8)	3.6 × 10-3	2.55 (1.33-4.90)	0.004
0-2	186 (97.9)	4 (2.1)	Ref.	Ref.		768 (97.6)	19 (2.4)	Ref.	Ref.		954 (97.6)	23 (2.4)	Ref.	Ref.

CMRF: Cardiometabolic risk factors; OR: Odds ratio; CI: Confidence interval.

In the validation dataset, mt12361A>G showed borderline significance in MASLD patients with 2-3 cardiometabolic risk factors than the controls [odds ratio (OR) = 2.31, 95% confidence interval (CI): 0.96-5.55, P= 0.055]. However, this significance was not observed in patients with 4-5 cardiometabolic traits. In the overall dataset, MASLD patients with more cardiometabolic traits tended to have a higher frequency of the mt12361 G allele (trend test P = 1.7 × 10^-4). Individuals with 2-3 cardiometabolic traits who carried the mt12361G variant significantly increased the risk of MASLD compared with the controls (OR = 3.43, 95%CI: 1.65-7.13, P = 4.6 × 10^-4). Patients with 4-5 cardiometabolic traits carrying the G allele had a 3.79-fold higher risk of MASLD than the controls (OR = 3.79, 95%CI: 1.61-8.93, P = 0.001).

In the univariate analysis, MASLD patients were significantly older, predominantly male, and had elevated cholesterol, low-density lipoprotein cholesterol (LDL), AST, ALT levels, and a higher frequency of the G allele of mt12361. Multivariate logistic regression analysis revealed that age, sex, cholesterol, LDL, AST, ALT, and mt12361A>G were independent risk factors of MASLD after adjusting for factors with P < 0.05 identified in the univariate analysis. Patients with the mt12361 G allele had a 2.54-fold higher risk of developing MASLD than mt12361A carriers (adjusted OR = 2.54, 95%CI: 1.19-5.43, P = 0.016) (Table 3). The Hosmer-Lemeshow test confirmed that this regression model had acceptable goodness-of-fit in predicting MASLD (P = 0.088).

Table 3 Multivariate logistic regression analysis for the association between mt12361A>G and metabolic dysfunction-associated steatotic liver disease in the overall dataset, mean ± SD/n (%).

	Univariate			Multivariate
	MASLD	Control	P value	OR	95%CI	P value
n (%)	577 (44.0)	733 (56.0)
Age (years)	49.7 ± 12.1	47.7 ± 13.9	0.005	1.02	1.01-1.03	0.001
Sex (male)	340 (59.0)	277 (37.9)	3.5 × 10-14	1.54	1.19-1.99	9.3 × 10-4
Cholesterol (mg/dL)	200.4 ± 38.5	193.8 ± 38.4	0.002	0.99	0.99-1.00	4.9 × 10-4
LDL (mg/dL)	133.7 ± 36.1	120.2 ± 34.0	1.0 × 10-11	1.02	1.01-1.02	2.3 × 10-8
AST ULN ratio	0.61 ± 0.28	0.55 ± 0.29	2.7 × 10-4	0.15	0.06-0.40	1.4 × 10-4
ALT ULN ratio	0.70 ± 0.46	0.49 ± 0.35	1.1 × 10-17	11.43	5.88-22.22	6.7 × 10-13
mt12361A>G	26 (4.7)	13 (1.8)	0.003	2.54	1.19-5.43	0.016

OR: Odds ratio; CI: Confidence interval; MASLD: Metabolic dysfunction-associated steatotic liver disease; LDL: Low density lipoprotein cholesterol; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; ULN: Upper limit of normal.

Subgroup analysis (diabetes vs non-diabetes)

The study participants were further categorized into diabetic and non-diabetic subgroups. In the diabetic subgroup, no significant association was observed between the mt12361A>G polymorphism and hepatic steatosis. In the non-diabetic subgroup, the G allele frequency was 1.6%, 5.0%, and 7.2% among patients with 0-1, 2-3, and 4-5 cardiometabolic risk factors, respectively (trend test P = 5.3 × 10^-5). Non-diabetic individuals with 2-3 cardiometabolic factors carrying the G allele had a 3.37-fold higher risk than the controls (95%CI: 1.52-7.48, P = 1.6 × 10^-3). Non-diabetic patients with 4-5 cardiometabolic traits carrying the mt12361G variant significantly increased 4.93-fold risk of MASLD compared with the controls (95%CI: 1.97-12.35, P = 1.8 × 10^-4). The genetic effect of mt12361A>G was prominent in the non-diabetic subgroup but not in the diabetic subgroup (Table 4).

Table 4 The association between mt12361 and steatotic liver disease in the diabetes and non-diabetic subgroups.

CMRF	Diabetes					Non-diabetes
	mt12361		χ2		Trend test	mt12361		χ2		Trend test
	A (%)	G (%)	P value	OR (95%CI)	P value	A (%)	G (%)	P value	OR (95%CI)	P value
4-5	28 (96.6)	1 (3.4)	1.000	1.11 (0.07-18.55)	0.926	103 (92.8)	8 (7.2)	1.8 × 10-4	4.93 (1.97-12.35)	5.3 × 10-5
2-3	8 (80.0)	2 (20.0)	0.136	7.75 (0.62-96.63)		245 (95.0)	13 (5.0)	1.6 × 10-3	3.37 (1.52-7.48)
0-1	31 (96.9)	1 (3.1)	Ref.	Ref.		762 (98.4)	12 (1.6)	Ref.	Ref.
3-5	32 (94.1)	2 (5.9)	1.000	1.81 (0.16-21.06)	0.676	212 (93.8)	14 (6.2)	3.9 × 10-4	3.90 (1.74-8.71)	5.6 × 10-4
1-2	6 (85.7)	1 (14.3)	0.347	4.83 (0.26-88.53)		249 (96.9)	8 (3.1)	0.167	1.90 (0.75-4.77)
0	29 (96.7)	1 (3.3)	Ref.	Ref.		649 (98.3)	11 (1.7)	Ref.	Ref.
3-5	32 (94.1)	2 (5.9)	1.000	1.09 (0.15-8.23)	0.931	212 (93.8)	14 (6.2)	9.2 × 10-4	3.12 (1.54-6.33)	9.2 × 10-4
0-2	35 (94.6)	2 (5.4)	Ref.	Ref.		898 (97.9)	19 (2.1)	Ref.	Ref.

CMRF: Cardiometabolic risk factors; OR: Odds ratio; CI: Confidence interval.

In the non-diabetic subgroup, univariate analysis indicated that MASLD patients were older, predominantly male, and had significantly higher cholesterol, LDL, AST, and ALT levels, along with a higher frequency of mt12361A>G variants. Multivariate regression analysis confirmed that age, sex, cholesterol, LDL, AST, ALT, and mt12361A>G were significant risk factors for MASLD. Non-diabetic individuals carrying the mt12361G variant significantly increased 2.80-fold risk of MASLD compared with mt12361A carriers (adjusted OR = 2.80, 95%CI: 1.22-6.41, P = 0.015) (Table 5). This regression model had optimal goodness-of-fit in predicting MASLD among the non-diabetic subgroup (Hosmer-Lemeshow test, P = 0.197).

Table 5 Multivariate logistic regression analysis for the association between mt12361A>G and metabolic dysfunction-associated steatotic liver disease among the non-diabetic subgroup, mean ± SD/n (%).

	Univariate			Multivariate
	MASLD	Control	P value	OR	95%CI	P value
n (%)	508 (42.5)	687 (57.5)
Age (years)	49.1 ± 11.9	47.0 ± 13.7	0.005	1.02	1.01-1.03	0.002
Sex (male)	293 (57.8)	256 (37.4)	2.7 × 10-12	1.47	1.12-1.93	0.005
Cholesterol (mg/dL)	199.7 ± 37.6	193.5 ± 37.7	0.005	0.99	0.98-1.00	5.7 × 10-4
LDL (mg/dL)	133.3 ± 35.6	120.2 ± 33.9	2.0 × 10-10	1.02	1.01-1.02	2.3 × 10-7
AST ULN ratio	0.61 ± 0.28	0.55 ± 0.30	4.3 × 10-4	0.17	0.06-0.48	0.001
ALT ULN ratio	0.69 ± 0.46	0.49 ± 0.35	9.4 × 10-16	10.77	5.34-21.71	3.2 × 10-11
mt12361A>G	22 (4.6)	11 (1.7)	0.004	2.80	1.22-6.41	0.015

OR: Odds ratio; CI: Confidence interval; MASLD: Metabolic dysfunction-associated steatotic liver disease; LDL: Low density lipoprotein cholesterol; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; ULN: Upper limit of normal.

Performance of the AI models in MASLD prediction

In the training dataset, the AUROC was 1.000, 0.970, 0.824, and 0.817 for the random forest, XGBoost, Naive Bayes, and logistic regression algorithms, respectively. The random forest model outperformed the other models, achieving accuracy, F1 score, precision, and recall rate of up to 100% (Figure 1 and Table 6). In the validation dataset, the AUROC reached 0.876, 0.861, 0.849, and 0.853 for the random forest, XGBoost, Naive Bayes, and logistic regression algorithms, respectively. The accuracy, F1 score, precision, and recall rates of the random forest model were 0.802, 0.803, 0.807, and 0.802, respectively (Figure 1 and Table 6). The sensitivity, specificity, positive and negative predictive values of the random forest model were 81.9%, 78.9%, 73.9%, and 85.6%, respectively (Supplementary Table 6). The Delong test revealed that the performance of the random forest model was significantly better than that of the XGBoost, Naive Bayes, and logistic regression models on the training dataset (P < 0.001). In the validation dataset, random forest was superior to Naive Bayes (P = 0.011) and comparable to the XGBoost or logistic regression models (Supplementary Table 7).

Figure 1 Performance of artificial intelligence models for metabolic dysfunction-associated steatotic liver disease. Training dataset: A: Area under the receiver operating characteristic curve; B: Precision-Recall curve; Validation dataset: C: Area under the receiver operating characteristic curve; D: Precision-Recall curve. FP: False positive; TP: True positive; XGBoost: eXtreme gradient boosting.

Table 6 Performance of the artificial intelligence models.

Algorithm	AUC	Accuracy	F1	Precision	Recall
Training
Random forest	1.000	1.000	1.000	1.000	1.000
XGBoost	0.970	0.918	0.918	0.919	0.918
Naive Bayes	0.824	0.747	0.748	0.757	0.747
Logistic regression	0.817	0.745	0.744	0.744	0.745
Validation
Random forest	0.876	0.802	0.803	0.807	0.802
XGBoost	0.861	0.763	0.764	0.767	0.763
Naive Bayes	0.849	0.776	0.777	0.786	0.776
Logistic regression	0.853	0.756	0.754	0.754	0.756

AUC: Area under curve; XGBoost: eXtreme gradient boosting.

DISCUSSION

Individuals carrying the mt12361 G allele had a significantly increased risk of MASLD compared with mt12361A carriers in the screening dataset. Nevertheless, mt12361A>G showed only borderline significance among MASLD patients with 2-3 cardiometabolic traits in the validation dataset. Multivariate regression analysis confirmed that the mt12361A>G variant significantly increased 2.54-fold risk of developing MASLD. The genetic effect of mt12361A>G was prominent in the non-diabetic group but not in the diabetic group. The random forest model for MASLD prediction achieved an AUROC of 1.000 for the training dataset and 0.876 for the validation dataset. The AI model offers a potential solution for screening MASLD without relying on imaging facilities.

Inherited genes synergize with cardiometabolic risk factors in the progression of MASLD[43]. Insulin resistance is closely associated with the pathogenesis of MASLD, particularly in patients with type 2 diabetes and obesity[44]. Interestingly, the mt12361A>G variant increased the risk of MASLD in the non-diabetic group but not in the diabetic group. A previous study reported that the mt12361G variant increased the risk of moderate and severe NAFLD in a Chinese population[30]. The transition from NAFLD to MASLD reflects a shift in terminology that better captures the underlying pathophysiology. Despite the change in nomenclature, studies have indicated that the prevalence and risk factors of these conditions remain similar[45]. Consistent with our results, mtDNA damage has been reported to be correlated with hepatic steatosis[46]. The liver tissue in patients with NAFLD harbors complex mtDNA with a higher mutation rate and a greater degree of heteroplasmy than that in the controls[27]. mtDNA mutations that primarily affect the OXPHOS system influence the MASLD phenotypes. Genetic variability in the mitochondrial cytochrome b, a key component of the respiratory chain, drives the severity of fatty liver disease[28]. Patients carrying mt14766C>T missense mutations in the cytochrome b gene exhibit marked changes in mitochondrial morphology. Under electron microscopy, mitochondria appeared swollen, with a condensed mitochondrial matrix, and loss of mitochondrial membranes and cristae[27]. Progression of MASLD is linked to genetic variations in the D-loop hypervariable region of mtDNA. Specifically, the mt16318C>A mutation is correlated with steatohepatitis, whereas mt16129AA carriers have a higher degree of liver fibrosis[29]. Additionally, mitochondrial haplogroup G increases susceptibility to NAFLD, whereas haplogroup L protects against non-alcoholic steatohepatitis (NASH) and fibrosis[47,48]. However, we did not identify any haplogroups that were correlated with MASLD. A high degree of mtDNA homology (approximately 98%) has been observed between blood and liver tissues[27]. The mt12361A>G variant may partially explain the heritability of MASLD in non-diabetic individuals.

mtDNA is a maternally inherited 16.5-kb circular dsDNA that encodes 13 mRNA of OXPHOS complexes I, III, IV, and V, along with 22 tRNAs and 2 rRNAs for mitochondrial protein synthesis[49]. In recent decades, only sporadic reports have described the association between mt12361A>G and fatty liver[30]. The biological function of mt12361A>G remains unclear. The mt12361A>G variant causes the substitution of threonine with alanine at position 9 of the ND5 subunit of complex I, which is a major component of the mitochondrial respiratory train. The Thr → Ala substitution may diminish electron transport chain activity, potentially resulting in increased oxidative stress in the hepatocytes. Elevated oxidative stress and reduced ATP production can trigger apoptosis via activation of the mitochondrial permeability transition pore[50]. The extent of enzyme deficiency in patients with complex I defects has been linked to increased free radical production[51]. Mitochondrial dysfunction results in lipotoxicity, triggering pro-inflammatory pathways, and contributing to metabolic dysfunction in MASLD[12]. Moreover, threonine can form hydrogen bonds and is suitable for various post-translational modifications, whereas alanine is a non-reactive neutral amino acid that is typically not directly involved in protein function. Epigenetic modifications of hepatic mtDNA are correlated with obesity, diabetes, and histological severity of fatty liver disease[52-54]. The functional consequences of the mt12361A>G variant require further investigation.

Non-invasive tests can be conducted in primary care settings using simple scoring systems based on laboratory and clinical parameters. There are several preexisting surrogate markers of hepatic steatosis, such as the fatty liver index, hepatic steatosis index, NAFLD liver fat score, and SteatoTest-2, whose AUROC ranges from 0.77 to 0.85[55]. The performance of the random forest model in the detection of MASLD was better than that of these non-invasive tests. International guidelines recommend two-tier sequential testing in primary care for MASLD, starting with the fibrosis-4 index (FIB-4), followed by vibration-controlled elastography or an enhanced liver fibrosis test[56]. However, the FIB-4 index, and enhanced liver fibrosis test may underestimate the presence of significant liver disease in certain settings. Applying the FIB-4 index with a cut-off of 1.3, the type 2 diabetes population would miss up to 38% of those with advanced liver disease, predominantly younger patients with normal liver function tests[57]. Combining machine learning algorithms and the FIB-4 index may optimize patient referral to hepatologists.

AI models are a promising tool for screening MASLD in primary care. By integrating clinical data and mt12361A>G, the random forest algorithm could discriminate MASLD with an AUROC of 1.000 in the training dataset and 0.876 in the validation dataset. A previous study (n = 14439) developed a support vector machine model to screen for NAFLD with an AUROC of up to 0.850[58]. Another large study (n = 304145) constructed a machine learning-based framework to classify NAFLD, in which XGBoost achieved the best accuracy (AUROC = 0.951; accuracy = 0.880). However, this study lacked objective and unified standards for certain predictors such as dietary status, which may have diminished the accuracy of the model[59]. In a biopsy-proven study, a gradient-boosting algorithm improved the identification of NASH and at-risk NASH patients, achieving an AUROC of 0.71 and 0.83, respectively[35]. Although machine learning is powerful, most models remain as “black boxes” because the algorithms are too complex to interpret. The application of AI for the detection of MASLD is still in its infancy. Concerted efforts toward methodology, biomarker discovery, and high-quality data can enhance the performance of AI models.

The performance of different algorithms in this study varied due to their unique characteristics and assumptions. Random forest achieved the highest accuracy among all models. Random forest is an ensemble technique that aggregates multiple decision trees, reducing overfitting and improving generalization. It effectively handles non-linear relationships and interactions between features[60]. XGboost has a robust performance because it is an advanced gradient boosting technique that optimizes model performance through regularization and capturing complex feature interactions. It also includes built-in mechanisms to prevent overfitting[33]. By contrast, Naive Bayes assumes conditional independence between features, which may not hold in practice. It may perform poorly if there are correlations between features. Logistic regression assumes a linear relationship between the independent variables and the log odds of the dependent variable. It performs well on linearly separable problems but may struggle with non-linear relationships[61].

MASLD is a complex metabolic disorder that affects the systemic liver-kidney-heart axis. Increased awareness of MASLD and multidisciplinary collaboration are essential to reduce the burden of cardiovascular and liver-related mortality. Integrating primary and specialist care is important in the management of MASLD. Primary care physicians often play a crucial role in the prevention, screening, and care of MASLD[62]. However, not all primary medical institutes have imaging facilities, and not all clinicians are well-trained in ultrasound technology. The AI model provides a simple screening tool for discriminating between healthy and MASLD participants. This helps in selecting candidates referred to hepatologists and identifying high-risk MASLD patients who may benefit from intensive lifestyle modifications.

Our study had several limitations. The reported prevalence of MASLD in the present study was higher than that in the general population of Taiwan (36.4%)[63], likely due to selection bias. The association between mt12361A>G and MASLD achieved only borderline significance in the validation dataset. The prevalence of MASLD was higher in the hospital-based cohort than in the community-based cohort (50.4% vs 42.4%), which was consistent with the distribution of mt12361G allele frequency (5.7% vs 2.4%). Despite attempts to minimize differences using propensity score matching, there was some heterogeneity between the training and validation datasets. The performance of the AI model is suboptimal in the validation data. Potential overfitting and heterogeneity in the training dataset may have affected the generalizability of the model. Additionally, the classification of hepatic steatosis was solely dependent on ultrasonography, which may be subject to interobserver variation. If hepatic steatosis is less than 20%, ultrasound may miss it[64]. The consistency between AI models and histological changes in hepatic steatosis should be further confirmed. Liver fibrosis is the most important predictor of long-term outcomes in MASLD[65]. This study did not analyze the correlation between mtDNA variants and advanced liver fibrosis due to the lack of data regarding liver biopsy, FIB-4, or liver stiffness. Therefore, it is necessary to validate an AI model in external populations before its clinical application.

CONCLUSION

The mt12361A>G variant significantly increased the risk of developing MASLD. The genetic effect of mt12361A>G was significant in the non-diabetic group but not in the diabetic group. This study sheds light on the potential pathogenesis of mtDNA variants associated with the severity of MASLD. Machine learning is a promising solution for identifying high-risk patients with MASLD. It supports the prevention, screening, and management of MASLD through interdisciplinary collaboration.