Integrating machine learning with multilayer transcriptomics pins JAK2 and ANXA5 as key genes linking obstructive sleep apnea to oxidative stress, validated in patient adipose tissue, intermittent-hypoxia mice, and post-CPAP samples

The study identifies a pair of oppositely directed hub genes in OSA adipose tissue, with JAK2 upregulated and ANXA5 downregulated, reproduces the same trends in the independent GSE135917 dataset, and reports ROC AUCs of 0.838 (training) and 0.790 (validation) for JAK2 and 0.775 and 0.919 for ANXA5. Prior work linking OSA to oxidative stress often focused on a single molecular pathway without systematic multilayer transcriptomic integration; this work intersects differential expression, co-expression modules, and an external oxidative-stress gene set, then takes the intersection of three machine learning algorithms to converge on two genes. Transcriptomic evidence from GSE38792 (10 OSA, 8 controls) for training and GSE135917 (34 OSA, 8 controls) for validation, reported with ROC curves and 95% confidence intervals, constituting computational reproducibility evidence.

Single-cell analysis shows JAK2 and ANXA5 are highly expressed in monocytes of OSA samples and identifies eight monocyte subtypes, with the Mono-inflammatory subset expanded in OSA patients, the Mono-non-classical population relatively reduced, and the Mono-ISG population also altered. The study localizes the two hub genes to specific immune cell subsets and describes monocyte subtype remodeling in OSA, going beyond tissue-level expression differences. Based on human single-cell dataset GSE214865, quality-controlled on mitochondrial percentage and gene counts, yielding 16 clusters annotated into 14 major cell populations, with subtype identity supported by dot and violin plots of markers such as CD14, FCGR3A, CX3CR1, ISG15, and IL1B.

Virtual knockout of JAK2 or ANXA5 with scTenifoldKnk strongly perturbs immune and inflammation genes including S100P, ALOX5AP, PROK2, and PADI4 (log2FC > 6), and ANXA5 knockout also dysregulates HLA genes such as HLA-DPB1, HLA-DPA1, HLA-DQA1, and HLA-DRA, with enrichment pointing to antigen processing and presentation, interferon response, and inflammation regulation. The study uses computational loss-of-function simulation to predict downstream regulatory networks and links ANXA5 to antigen presentation, proposing a previously underappreciated immune regulatory role. A machine learning inference from single-cell data, presented across bar, volcano, MA, and dot plots with reported enrichment significance (e.g., p < 1e-15, −log10(p) > 20), but computational rather than experimental knockout.

Clinical and animal validation shows significantly elevated JAK2 mRNA and protein and reduced ANXA5 in OSA patient adipose tissue, the same direction in intermittent-hypoxia mouse adipose tissue, decreased JAK2 and increased ANXA5 in 10 OSA patients after CPAP treatment, and a more pronounced dysregulation in OSA patients than in non-OSA obese or metabolically obese groups. The study advances bioinformatic screening into experimental validation across human adipose tissue, a mouse model, and paired pre/post-treatment samples, closing a loop from computation to experiment. RT-qPCR and Western blotting on samples from 40 subjects (20 OSA, 20 controls) at the Department of Otolaryngology and Head and Neck Surgery, Shanghai Xinhua Hospital, plus C57BL/6 intermittent-hypoxia mice at n = 5 per group, with the pre/post-CPAP comparison based on 10 patients.

The work offers a reusable analytical path for researchers interested in OSA molecular mechanisms, biomarkers, and immunometabolism: starting from public transcriptomic data, converging candidate genes via WGCNA and machine learning, then confirming layer by layer with single-cell analysis, virtual knockout, and experiments. Its conclusions apply to adipose tissue (visceral fat in the training data, soft palate fat in the validation samples) and to the monocyte cellular context; the authors propose that combined detection of JAK2 and ANXA5 could serve as an auxiliary diagnostic approach beyond PSG, and suggest JAK2 as a possible adjuvant therapy direction for CPAP-intolerant patients.

Several open questions remain: the core transcriptomic cohorts are small (18 training, 42 validation subjects), clinical validation is single-center with 40 subjects, and the pre/post-CPAP comparison covers only 10 patients, so how these sizes affect stability awaits larger multicenter prospective cohorts; virtual knockout is computational inference, and its predicted changes in antigen presentation and interferon pathways still need real knockout or intervention experiments; the study focuses on adipose tissue and monocytes, leaving other organs and cell types unknown; and because this is a full-text parse, the specific values and statistical details behind Figures 1 to 16 are not itemized in the text, so readers checking effect sizes should return to the original figures.