Machine learning finds Non-APOE gene signals too weak to predict Alzheimer’s

A two-stage sequencing study in Korean adults across the Alzheimer’s disease continuum has tested whether candidate genetic variants outside the APOE region can add predictive power to machine learning models of the disease. The answer, reported in Alzheimer’s Research & Therapy, is largely no: when the researchers stripped away clinically proximal information and APOE itself, model performance fell sharply, and the Non-APOE candidate variants contributed little beyond chance.

The team applied a whole-exome sequencing stage to an amnestic mild cognitive impairment cohort of 346 individuals, where candidate loci were first nominated. These candidates were then evaluated in an independent whole-genome sequencing cohort of 995 participants, together with epidemiologic and neuropsychological variables. Gradient-boosted decision trees implemented in XGBoost were trained under a common stratified five-fold cross-validation scheme with pooled out-of-fold predictions, and all preprocessing, including one-hot encoding, was fitted within each training fold only to prevent information leaking from validation data.

The full clinical-genomic model achieved near-perfect internal discrimination, with an area under the receiver operating characteristic curve of 0.998 (95% CI 0.996–0.999). The authors caution, however, that this model included diagnostic-proximal clinical and neuropsychological variables, meaning it should be read as an upper-bound estimate of clinical separability rather than evidence of prospective predictive utility in real-world screening.

The more informative result came from a strict leakage-reduced sensitivity analysis. There, the AUROC dropped to 0.713 (95% CI 0.682–0.745), and fell further to 0.625 (95% CI 0.589–0.658) after removal of rs429358 and APOE proxy information. This stepwise decline illustrates how much of the apparent predictive signal in clinically enriched datasets reflects variables closely tied to the diagnostic process itself rather than underlying biology.

Focusing on genetics alone, a five-SNP model showed modest discrimination with an AUROC of 0.637 (95% CI 0.605–0.670), while a model restricted to Non-APOE SNPs was weak, at 0.533 (95% CI 0.497–0.568) — barely above the 0.5 expected from random guessing. In fold-wise permutation analyses of the strict model, Test_Age ranked first among all features, and rs429358, the well-established APOE-linked variant, was the highest-ranking genetic feature. The Non-APOE candidates showed small or near-zero mean AUROC decreases after permutation, indicating minimal contribution.

The study’s background notes that Alzheimer’s disease has a genetic architecture dominated by APOE, while predictive modeling in clinically enriched cohorts is often driven by neuropsychological and clinical variables. That combination, the authors argue, limits the interpretability of secondary genetic signals within multimodal frameworks — a problem their staged evaluation design was built to expose.

The conclusions are deliberately restrained. High multimodal discrimination in clinically enriched Alzheimer’s datasets can reflect diagnostic-proximal clinical separability, the authors write, and Non-APOE candidate variants should be interpreted as hypothesis-generating model features rather than confirmed susceptibility loci, independent predictors, or causal variants.

Important limitations frame these findings. No second independent Korean cohort with harmonized candidate genotypes and comparable clinical and neuropsychological variables was available, so direct external replication was not possible. In addition, the whole-exome and whole-genome stages used different clinical endpoints. The work therefore stands as an exploratory candidate-evaluation framework — one that demonstrates, with careful leakage control, how easily inflated performance metrics can arise in clinical-genomic modeling, and how rigorously such claims need to be tested before genetic variants beyond APOE are treated as meaningful predictors of Alzheimer’s disease.

Subject of Research: Exploratory clinical-genomic modeling of Non-APOE genetic variants in Alzheimer's disease

Article Title: Integrative clinical–genomic modeling evaluates exploratory Non-APOE candidate signals in Alzheimer’s disease

Article References: Paik, E. K., Lee, J., Lee, E.-S., Park, J. K., Song, J.-Y., Kim, S., Lee, J.-Y., Koh, I., Kim, W., & Park, Y. (2026). Integrative clinical–genomic modeling evaluates exploratory Non-APOE candidate signals in Alzheimer’s disease. Alzheimer's Research & Therapy. https://doi.org/10.1186/s13195-026-02231-0

Image Credits: AI Generated

DOI: 10.1186/s13195-026-02231-0

Keywords: Alzheimer's disease, APOE, whole-exome sequencing, whole-genome sequencing, XGBoost, machine learning, genetic variants, mild cognitive impairment, AUROC, predictive modeling, leakage control, Korean cohort