Sleep deprivation induces systemic physiological stress accompanied by transcriptomic remodelling and immune dysregulation, yet objective molecular indicators for its assessment remain insufficient. This study integrated blood transcriptomic analysis with an interpretable machine learning framework to identify and validate candidate molecular signatures associated with sleep deprivation and their potential relevance to insomnia. Publicly available Gene Expression Omnibus datasets were used to construct an acute sleep deprivation training cohort, an independent sleep deprivation validation cohort, and a chronic insomnia validation cohort. Differentially expressed genes were first identified, followed by feature selection using six machine learning algorithms and Shapley additive explanations to improve model interpretability. Immune cell composition was inferred using CIBERSORT, and associations between candidate genes and immune cell subsets were further evaluated. Twenty-five differentially expressed genes were identified in the training cohort, from which eight high-priority candidate genes were selected by the interpretable machine learning framework. Among them, S100A3 showed consistent discriminatory performance across the training cohort, the independent sleep deprivation cohort, and the insomnia cohort, whereas VEGFB exhibited notable diagnostic potential, particularly in insomnia. Immune infiltration analysis indicated that sleep deprivation was associated with altered peripheral immune composition, including reduced resting natural killer cells and activated dendritic cells, together with changes in regulatory and naïve immune cell populations. Expression levels of S100A3 and VEGFB were significantly correlated with specific immune cell subsets, suggesting a link between these molecular signatures and stress-related immunomodulation. These findings identify S100A3 as a robust candidate biomarker shared by acute sleep deprivation and chronic insomnia, while VEGFB may reflect chronic metabolic or inflammatory adaptation. The proposed interpretable transcriptomic-machine learning framework provides a non-invasive strategy for discovering molecular indicators of sleep-related physiological stress and may support future risk stratification in sleep medicine.
Wang et al. (Fri,) studied this question.