Bioinformatics-driven discovery and cross-cohort validation of a two-gene signature (RPH3A and HIGD1B) associated with Alzheimer's disease.

Background

Transcriptomic biomarker discovery for Alzheimer's disease (AD) has yielded numerous candidate signatures, yet many fail independent external validations due to overfitting or cohort-specific confounders. We aimed to identify a minimal reproducible gene signature and evaluate its generalizability under locked-model validation.

Methods

Three GEO datasets were analyzed: GSE118553 (training), GSE122063 (internal validation), and GSE5281 (external validation). Differentially expressed genes were identified in the training set, followed by LASSO and Random Forest feature selection, and intersecting genes were filtered for concordant fold-change direction and P < 0.05 across cohorts. An elastic net model was evaluated using locked training parameters by ROC, calibration, and decision curve analyses,with exploratory threshold optimization performed to assess model performance.

Results

Four genes overlapped between machine-learning methods, but only RPH3A and HIGD1B passed the consistency filter. The locked model achieved AUCs of 0.818 and 0.974 in the internal and external cohorts, respectively. Internal validation showed calibration drift (intercept -2.697, slope 0.705), resulting in 0.000 specificity at the locked threshold. Cohort-specific threshold optimization increased specificity to 1.000 with 0.607 sensitivity. External validation showed good calibration (intercept 1.065, slope 1.203) and favorable decision-curve performance.

Conclusion

RPH3A and HIGD1B constitute a reproducible AD-associated molecular signature with robust discrimination across independent postmortem brain cohorts, regions, and platforms. As all datasets were postmortem brain tissue, these findings reflect disease-associated molecular alterations. Calibration, confusion matrix, and decision curve analyses further support model performance and translational relevance.