Artificial intelligence as a ploy to delve into the intricate link between genetics and mitochondria in patients with...

Mitochondrial (mt-) D-loop and cell-free circulating (ccf-) mtDNA fragments, respectively reflecting mt-mass and tissue damage, are promising metabolic dysfunction-associated steatotic liver disease (MASLD) biomarkers. We previously found that PNPLA3/MBOAT7/TM6SF2 deficiency in HepG2 cells increased mt-mass, D-loop levels, and ccf-COXIII release. We explored mt-biogenesis and mt-biomarkers in patients with MASLD stratified by the number of risk variants (NRV = 3). We exploited GPT-4 to develop and validate new risk scores, predicting MASLD evolution, in two independent cohorts by integrating anthropometric and genetic data with mt-biomarkers.

A cohort of 28 patients with MASLD (Discovery cohort) was consecutively enrolled for hepatic mt-dynamics assessment by transmission electron microscopy and immunohistochemistry. Data were confirmed by quantitative real time-PCR in a retrospective cohort (Hepatic Validation, n = 184). D-loop and ccf-COXIII were retrospectively measured in peripheral blood mononuclear cells and serum samples of biopsied outpatients with MASLD (Serum Validation cohort, n = 824) and individuals with non-invasive MASLD diagnosis (n = 386, Non-invasive cohort). Risk scores were developed using random forest algorithms.

In the Discovery and Hepatic Validation cohorts, the PNPLA3/MBOAT7/TM6SF2 variants altered hepatic mt-dynamics, enhancing mt-content and D-loop levels (p <0.05) through the p38/PGC-1α pathway. Furthermore, NRV = 3 patients showed an increase in mt-fragmentation at transmission electron microscopy (TEM) and ccf-COXIII release (p <0.05). In the Serum Validation cohort, circulating D-loop and ccf-COXIII positively correlated with genetics [β _D-loop :0.17 (95% CI: 0.04-0.29), p = 0.01; β_ccf-COXIII:0.33 (95% CI: 0.19-0.46), p <0.0001] and MASLD severity [OR _D-loop :1.31 (95% CI: 1.01-1.71), p = 0.03; OR_ccf-COXIII:2.41 (95% CI: 1.69-3.44), p <0.0001] at multivariate analysis. Random forest allowed prediction models named Mitochondrial, Anthropometric, and Genetic Integration with Computational intelligence for assessing hepatocellular carcinoma risk (MAGIC-H), considering age, BMI genetics, D-loop, and ccf-COXIII. In both Serum and Non-invasive cohorts, the MAGIC-H score reached AUC >85% in identifying HCC cases regardless of cirrhosis, outperforming existing non-invasive tests.

Mt-biomarkers have a prognostic significance in genetically-predisposed patients with MASLD.

The study highlights that genetic variants in PNPLA3, MBOAT7, and TM6SF2 genes deeply contribute to metabolic dysfunction-associated steatotic liver disease (MASLD) progression by affecting hepatic mitochondrial adaptability. It also identified two novel biomarkers of mitochondrial origin which are strongly linked to disease severity and genetic background of patients with MASLD. The use of generative artificial intelligence tools, such as GPT-4, can enhance the use of biomarkers and polygenic risk scores for clinical risk stratification. We developed a customized version of GPT-4 (rsGPT-4), which identified a machine-learning approach (random forest) as the best method for creating prediction models for metabolic dysfunction-associated steatohepatitis, fibrosis, and hepatocellular carcinoma. The new scores combined the two mitochondrial biomarkers, genetic data, and anthropometric data and outperformed existing non-invasive tests for monitoring patients with MASLD.