Construction of Autophagy-Related Gene Classifier for Early Diagnosis, Prognosis and Predicting Immune Microenvironment Features in Sepsis by Machine Learning Algorithms.

Background: The immune system plays a fundamental role in the pathophysiology of sepsis, and autophagy and autophagy-related molecules are crucial in innate and adaptive immune responses; however, the potential roles of autophagy-related genes (ARGs) in sepsis are not comprehensively understood. Methods: A systematic search was conducted in ArrayExpress and Gene Expression Omnibus (GEO) cohorts from July 2005 to May 2022. Machine learning approaches, including modified Lasso penalized regression, support vector machine, and artificial neural network, were applied to identify hub ARGs, thereby developing a prediction model termed ARG classifier. Diagnostic and prognostic performance of the model was comprehensively analyzed using multi-transcriptome data. Subsequently, we systematically correlated the ARG classifier/hub ARGs with immunological characteristics of multiple aspects, including immune cell infiltration, immune and molecular pathways, cytokine levels, and immune-related genes. Further, we collected clinical specimens to preliminarily investigate ARG expression levels and to assess the diagnostic performance of ARG classifier. Results: A total of ten GEO and three ArrayExpress datasets were included in this study. Based on machine learning algorithms, eight key ARGs (ATG4C, BAX, BIRC5, ERBB2, FKBP1B, HIF1A, NCKAP1, and NFKB1) were integrated to establish ARG classifier. The model exhibited excellent diagnostic values (AUC > 0.85) in multiple datasets and multiple points in time and superiorly distinguished sepsis from other critical illnesses. ARG classifier showed significant correlations with clinical characteristics or endotypes and performed better in predicting mortality (AUC = 0.70) than other clinical characteristics. Additionally, the identified hub ARGs were significantly associated with immune cell infiltration (B, T, NK, dendritic, T regulatory, and myeloid-derived suppressor cells), immune and molecular pathways (inflammation-promoting pathways, HLA, cytolytic activity, apoptosis, type-II IFN response, complement and coagulation cascades), levels of several cytokines (PDGFRB, IL-10, IFNG, and TNF), which indicated that ARG classifier/hub ARGs adequately reflected the immune microenvironment during sepsis. Finally, using clinical specimens, the expression levels of key ARGs in patients with sepsis were found to differ significantly from those of control patients, and ARG classifier exhibited superior diagnostic performance, compared to procalcitonin and C-reactive protein. Conclusion: Collectively, a diagnostic and prognostic model (ARG classifier) based on eight ARGs was developed which may assist clinicians in diagnosis of sepsis and recognizing patient at high risk to guide personalized treatment. Additionally, the ARG classifier effectively reflected the immune microenvironment diversity of sepsis and may facilitate personalized counseling for specific therapy.