Plasma metabolomics profiling and machining learning-driven prediction of nonalcoholic steatohepatitis

Abstract

Rationale: We performed targeted metabolomics with machine learning (ML)-based interpretation to identify metabolites that distinguish the progression of nonalcoholic fatty liver disease (NAFLD) in a cohort. Methods: We conducted plasma metabolomics analysis in healthy control subjects (n=25) and patients with NAFL (n=42) and nonalcoholic steatohepatitis (NASH, n=19) by gas chromatography-tandem mass spectrometry (MS/MS) and liquid chromatography-MS/MS as well as RNA sequencing (RNA-seq) analyses on liver tissues from patients with varying stages of NAFLD (n=12). The resulting metabolomic data were subjected to routine statistical and ML-based analyses and multiomics interpretation with RNA-seq data. Results: We found six metabolites that were significantly altered in NAFLD among 79 detected metabolites. Random-forest and multinomial logistic regression analyses showed that eight metabolites (glutamic acid, cis-aconitic acid, aspartic acid, isocitric acid, α-ketoglutaric acid, oxaloacetic acid, myristoleic acid, and tyrosine) could distinguish the three groups. Then, the recursive partitioning and regression tree algorithm selected three metabolites (glutamic acid, isocitric acid, and aspartic acid) from these eight metabolites. With these three metabolites, we formulated an equation, the MetaNASH score that distinguished NASH with excellent performance. Finally, metabolic map construction and correlation assays integrating metabolomics data into the transcriptome datasets of the liver showed correlations between the concentration of plasma metabolites and the expression of enzymes governing metabolism and specific alterations of these correlations in NASH. Conclusions: We found several metabolites that distinguish NASH from non-NASH via metabolomics analysis and ML approaches, developed the MetaNASH score, and suggested the pathophysiologic implications of metabolite profiles in relation to NAFLD progression.

Machine learning-featured Secretogranin V is a circulating diagnostic biomarker for pancreatic adenocarcinomas associated with adipopenia

“…Construction of models with Random Forest (RF) algorithm was done as described previously ( 28 ). To build models capable of distinguishing between NOP and PAC samples as well as higher and lower survival rates, we implemented RF algorithm-based machine learning with optimized parameters through a 10-fold repeated cross-validation using the R package “ caret ”.…”

“…One of the finest methods to deduce high-potential candidates is machine-learning-based categorization, which acts as a researcher's bias-free technique to discover traits that may function as diagnostic and prognostic biomarkers, giving an opportunity to uncover underlying pathophysiological principles and mechanisms (25)(26)(27)(28). Herein, we performed machine learning-based classification to identify diagnostic and prognostic markers for PAC.…”