Deep chemical structure graph learning deciphers the lipotoxicity code of hypertriglyceridemic pancreatitis

Area of Science:

• Biochemistry
• Computational Biology
• Immunology

Background:

• Hyperlipidemic acute pancreatitis (HLAP) risk stratification is challenging due to discrepancies between triglyceride levels and clinical severity.
• Traditional lipidomics methods struggle to differentiate HLAP-specific lipid profiles from general hypertriglyceridemia (HTG).

Purpose of the Study:

• To develop a novel computational framework, DeepLipiDecipher, for identifying lipotoxic features in HLAP.
• To integrate lipid chemical structures and metabolic network topology for enhanced lipidomic analysis.

Main Methods:

• Development of DeepLipiDecipher, a knowledge-guided graph neural network.
• Retrospective analysis of a cohort of 433 subjects.
• Application of computational causal inference to elucidate pathogenic mechanisms.

Main Results:

• DeepLipiDecipher achieved robust classification performance (AUC=0.810), distinguishing HLAP phenotypes.
• HLAP susceptibility was associated with specific lipid structures (polyunsaturated and ether-linked phospholipids), not total lipid mass.
• Identified a potential pathogenic pathway involving SMPD3, ceramide accumulation, and PTGS2 hyperactivation.

Conclusions:

• Integrating network topology into lipidomics enhances the identification of disease-specific features.
• The SMPD3-Ceramide-PTGS2 immunometabolic axis is a potential therapeutic target for HLAP.
• DeepLipiDecipher offers a promising approach for understanding and stratifying HLAP.