Integrative transcriptomics and machine learning identify temporal stress signatures and astrocytic neurocan associated with hypothalamic injury

Methods

We integrated hypothalamic long-RNA sequencing, small-RNA sequencing, single-nucleus RNA sequencing, and a multi-layer machine learning framework to characterize transcriptomic dynamics during stress progression and identify discriminative molecular signatures. Results: mRNAs exhibited strong temporal sensitivity and captured stage-specific transcriptional remodeling, whereas miRNAs primarily reflected the presence of stress regardless of duration.

By integrating differential expression analysis, Mfuzz, WGCNA, RF, LASSO, ElasticNet-RFE, and XGBoost, we established a six mRNA (Ncan, Mt2, Nab2, Ptch1, Catspere1/Catspere2) signature that accurately distinguished control, acute stress, and chronic stress, achieving an accuracy of 1.000 in independent testing. In contrast, an eleven miRNA signature demonstrated robust stress-state classification, with the best-performing RBF-SVM model reaching an accuracy of 0.944.

RT-qPCR further confirmed the reproducibility of both signatures, with the mRNA model achieving an accuracy of 0.898 in nested cross-validation and the miRNA model reaching an accuracy of 0.890. Among all candidates, Ncan was consistently identified as the top-ranked biomarker across machine learning models. It was predominantly expressed in astrocytes, progressively downregulated during stress, and its supplementation restored perineuronal nets (PNN) and alleviated depressive-like behaviors.

Virtual Ncan knockout preferentially disrupted the acute-stress regulatory network, with Ctcf predicted as its upstream regulator. Conclusion: Our study establishes a machine learning-based framework for stress molecular stratification, demonstrates that mRNAs encode stress-stage information whereas miRNAs primarily define stress states, and identifies astrocytic Ncan as a functionally validated biomarker linking PNN homeostasis to stress-induced neural injury.