Machine learning finds six shared aging genes linking Alzheimer's disease and osteoporosis
A 130-model machine learning screen plus cell experiments flags six aging genes common to Alzheimer's disease and osteoporosis, pointing to shared biology.
Summary
Alzheimer's disease and osteoporosis both become far more common with age and share risk factors such as female sex, genetics and lifestyle, but the shared biology is poorly understood. Researchers combined public gene-expression datasets from both diseases with a list of more than 2,000 known aging-related genes. They found 228 aging genes altered in both conditions. A screen of 130 machine learning models narrowed these to six: EGR1, NFKBIA, UQCRC2, RB1, CREBBP and NDUFV1. The genes relate to cellular senescence, inflammation, mitochondrial energy production and cell-cycle control. In mouse brain and bone-precursor cells aged artificially with D-galactose, all six changed in brain cells, and four changed in bone cells. This is an exploratory computational study and not a clinical test, but it points to shared aging pathways that might be targeted together.
Detailed Summary
Alzheimer's disease (AD) and osteoporosis (OP) are both diseases of late life. AD incidence doubles every five years after age 65, with nearly half of people 85 and older affected. OP prevalence is projected to double between 2010 and 2040. Earlier work links the two through amyloid-beta stimulation of osteoclasts, declining bone-derived osteocalcin, impaired Wnt/β-catenin signaling and shared inflammatory cytokines such as TNF-α and IL-6. The shared aging-related molecular drivers, however, have not been mapped at a systems level.
The authors combined four public GEO microarray datasets with 2,285 curated aging-related genes from the Human Ageing Genomic Resources and GeneCards. The AD cohorts were GSE132903 (97 cases, 98 controls) and GSE5281 (87 cases, 74 controls). The OP cohorts were GSE35958 (5 cases, 4 controls) and GSE7158 (12 cases, 14 controls). Each dataset was analyzed separately with limma (adjusted P < 0.05, absolute log2 fold change > 0.585) rather than pooled, and the results were then intersected. Weighted gene co-expression network analysis (WGCNA) identified disease-linked gene modules.
The analysis found 5,889 AD-related and 2,348 OP-related differentially expressed genes, with 725 shared. WGCNA identified a brown module positively correlated with AD (r = 0.57, P = 5 × 10⁻¹⁵; 1,576 genes) and a red module negatively correlated with OP (r = −0.50, P = 0.01; 711 genes). Integrating these with the differential expression results yielded 1,660 shared genes, and intersecting them with aging genes left 228. These were enriched for PI3K-Akt, longevity-regulating, cellular senescence, FoxO, p53, TNF and apoptosis pathways. The protein interaction network highlighted 30 hub genes, including TP53, EGFR, NDUFV1 and UQCRC2.
The team then built 130 predictive models from 11 algorithms. A Lasso plus gradient boosting (GBM) ensemble performed best, with AUCs of 0.915 (GBM) and 0.909 (Lasso), and selected six genes: EGR1, NFKBIA, UQCRC2, RB1, CREBBP and NDUFV1. Individual AUCs ranged from 0.646 (RB1) to 0.853 (NFKBIA). SHAP analysis ranked EGR1 (0.104) and NFKBIA (0.095) as the most influential predictors. A six-gene nomogram showed acceptable calibration and net benefit on decision curve analysis.
In the GSE33000 cohort (310 AD, 157 controls), CIBERSORT and ssGSEA showed that the hub genes correlated with immune cell infiltration patterns. For experimental validation, D-galactose-induced senescence models were used. All six genes were significantly altered in HT22 mouse hippocampal neurons. In MC3T3-E1 mouse pre-osteoblasts, EGR1, NFKBIA, UQCRC2 and NDUFV1 changed significantly, while RB1 and CREBBP did not, so the signature is partly cell-type specific.
The findings point to inflammation, mitochondrial oxidative phosphorylation, senescence and cell-cycle control as shared aging axes. Caveats are substantial. The OP datasets are very small, and the datasets come from different tissues and platforms. The nomogram for "co-occurrence" was built from separate single-disease cohorts, not patients with both conditions. Validation was mRNA-only in mouse cell lines using a chemical aging model, with no protein, in vivo or causal testing. The results are hypothesis-generating rather than clinically actionable.
Key Findings
- Identified 228 aging-related genes dysregulated in both Alzheimer's disease and osteoporosis, from 725 shared differentially expressed genes and 2,285 curated aging genes
- Screened 130 machine learning models from 11 algorithms; Lasso + GBM ensemble performed best (AUC 0.915 for GBM, 0.909 for Lasso)
- Six hub genes selected: EGR1, NFKBIA, UQCRC2, RB1, CREBBP, NDUFV1; individual AUCs ranged from 0.646 (RB1) to 0.853 (NFKBIA)
- SHAP ranked EGR1 (0.104) and NFKBIA (0.095) as the most influential predictors in the model
- WGCNA found an AD-linked brown module (r = 0.57, P = 5 × 10⁻¹⁵) and an OP-linked red module (r = −0.50, P = 0.01)
- Enrichment linked the shared genes to cellular senescence, p53, FoxO, PI3K-Akt, TNF signaling, and mitochondrial oxidative phosphorylation
- In D-galactose senescence models, all 6 genes changed in HT22 neurons, but only 4 changed in MC3T3-E1 pre-osteoblasts (RB1 and CREBBP did not)
Methodology
This was a computational study using four public GEO microarray datasets (AD: 97/98 and 87/74 cases/controls; OP: 5/4 and 12/14), with GSE33000 (310 AD, 157 controls) for immune deconvolution. Differential expression (limma) was run separately per dataset and intersected, followed by WGCNA, GO/KEGG enrichment and STRING protein interaction analysis. Models were trained with stratified 10-fold cross-validation and ranked by AUC; SHAP, nomogram, calibration and decision curve analysis were applied, and immune infiltration was assessed by CIBERSORT and ssGSEA. Six genes were validated by RT-qPCR (GAPDH reference, 2^−ΔΔCt, three independent experiments) in D-galactose-treated mouse HT22 and MC3T3-E1 cells, using Wilcoxon tests.
Study Limitations
The OP cohorts were very small (5 vs 4 and 12 vs 14), and the datasets spanned different tissues, cohorts and platforms. The diagnostic nomogram for AD-OP co-occurrence was built from separate disease datasets rather than patients with both conditions, and RB1 had a modest AUC (0.646). Validation was limited to mRNA changes in D-galactose-treated mouse cell lines, and RB1 and CREBBP did not change in bone-lineage cells. No conflict-of-interest information was available in the supplied text.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: