Deep learning-driven proteomics analysis for gene annotation in the renin-angiotensin system

Abstract

The renin-angiotensin system (RAS) is central to cardiovascular diseases such as hypertension and cardiomyopathy, yet the functions of many RAS genes remain unclear. This study developed a multi-label deep learning model to systematically annotate RAS gene functions and elucidate their roles in biological pathways. A total of 39,463 RAS-related publications from PubMed and PMC were processed into text format. Feature matrices were generated using TF-IDF and token processing, followed by dimensionality reduction via Principal Component Analysis (PCA). A Multi-Layer Perceptron (MLP) was applied for multi-label classification, with performance evaluated using Precision, F1-Score, Ranking Loss, and ROC-AUC metrics. The model outperformed traditional methods (SVM, Random Forest), achieving a Precision of 0.7474 and ROC-AUC of 0.8697. Grouping into three major biological branches improved interpretability and performance (Precision: 0.8312; ROC-AUC: 0.9182). In silico predictions were validated using extracellular vesicle (EV) proteomics and capillary Western assays in DOCA-salt hypertensive mice. Key genes-AGTR2, IRAP (LNPEP), Ywhas (SFN), EDNRA, and ESR2-were identified as critical RAS components. Notably, IRAP was markedly upregulated in hypertension and showed regulatory interactions with 14-3-3 proteins, modulating Nedd4-2, ACE2, and AGTR1 signaling. To our knowledge, this is the first integration of multi-label AI modeling with EV proteomics for RAS pathway annotation. This framework captures complex gene-pathway relationships, advancing systems-level understanding of RAS biology and revealing a novel IRAP/Ywha(s)/Nedd4-2-ACE2 interaction axis as a potential therapeutic target.

Keywords: Hypertension; Machine learning; Multi-layer perceptron; Renin-angiotensin system.