Construction of a Diagnostic Model and Drug Prediction for Postischemic Stroke Cognitive Impairment Based on Machine Learning Screening of Lactate Metabolism- and Pyroptosis-Related Genes.

Abstract

Reliable molecular biomarkers for poststroke cognitive impairment (PSCI) remain limited. Using publicly available bulk transcriptomic and single-cell RNA-seq datasets from GEO, we investigated lactate metabolism- and pyroptosis-related signatures and developed a diagnostic model. Differential expression analysis, KEGG pathway enrichment, and weighted gene coexpression network analysis (WGCNA) were performed, followed by multialgorithm feature selection (LASSO, SVM-RFE, and random forest). A logistic regression classifier was trained in the discovery cohort and externally validated in an independent cohort. Glycolysis/lactate metabolism, HIF-1 signaling, and NOD-like receptor-related pathways were enriched in PSCI-associated samples, and key coexpression modules were strongly correlated with ischemic injury traits. Cross-model consensus identified LDHA, GSDMD, and CASP1 as hub genes, yielding an AUC of 0.912 (95% bootstrap CI: 0.841-0.983) in the training cohort and 0.885 (95% bootstrap CI: 0.798-0.972) in the validation cohort. Immune deconvolution and scRNA-seq validation suggested increased proinflammatory microglia-associated signals, with relatively higher LDHA expression in microglia than in neurons; cell-cell communication analysis highlighted inflammatory interactions including IL1B-IL1R1. Connectivity map (CMap) analysis nominated candidate compounds, and molecular docking predicted favorable binding between oxamate and LDHA (binding energy = -9.5 kcal/mol). Collectively, these findings propose a compact LDHA/GSDMD/CASP1 biomarker panel for PSCI diagnosis and provide hypothesis-generating therapeutic leads that warrant further experimental validation.