Fig. 1: Overall workflow for development and translation of the LDS in AD. | npj Digital Medicine

Fig. 1: Overall workflow for development and translation of the LDS in AD.

From: A machine learning–enabled blood transcriptomic signature for digital diagnosis and subtyping of Alzheimer’s disease

Fig. 1

Panel I: Two GEO brain transcriptomic datasets (GSE5281 and GSE84422; 102 NC and 121 AD samples) were integrated as the Training Cohort, processed by log2 transformation, quantile normalization, and ComBat batch correction, and 32 differentially expressed lactylation-related genes (DE-LRGs) were identified. Panel II: Using multiple feature-selection and machine-learning algorithms, 113 candidate models were screened and an RF + plsRglm LDS model based on seven key genes (GFAP, GTF2I, RB1, PFKM, BCLAF1, SPR, SMARCC1) was established. Panel III: LDS was validated and benchmarked against published signatures and single genes in seven independent GEO cohorts, demonstrating robust cross-platform diagnostic performance. Panel IV: In an independent clinical plasma cohort (NC = 180, aMCI = 90, AD = 270; total n = 540), the seven LDS genes (RT-qPCR) and p-tau181/217 (Simoa) were measured to evaluate diagnostic performance for AD and aMCI, identify AT⁺ individuals, and explore clinical utility and potential therapeutic targets. This figure was created with BioRender.com and is used under a permitted license.

Back to article page