Can genetics predict Alzheimer's risk through imaging intermediates?

Genetics alone leaves much of the risk unexplained

Polygenic risk scores summarize many genetic variants into one number, and they are the standard way to predict common-disease risk from DNA. The paper's premise is that this approach has a ceiling: complex biological processes that sit between genes and symptoms are not fully captured by genetic variation alone, and traditional polygenic scores typically assume linear genetic effects. Imaging-based models, by contrast, can reach substantially higher accuracy for many diseases, but imaging is costly and unavailable for most people.

G2DBridge is designed around that asymmetry. It uses a cohort where genetics, imaging, and diagnosis are all measured to learn how genetic variation maps onto imaging-derived phenotypes, abbreviated IDPs. Those IDPs are quantitative brain measurements, such as volumes, cortical thickness, or diffusion metrics, that reflect structural and microstructural change. Once that mapping is learned, a new cohort needs only genotype data: the model infers the imaging features and passes them to a disease classifier.

Three imaging modalities carry largely non-redundant signal

The study used three IDP sets from the Alzheimer's Disease Neuroimaging Initiative (ADNI): diffusion tensor imaging (DTI) for white matter microstructure, FreeSurfer morphometric measures for brain volumes and cortical thickness, and Automatic Segmentation of Hippocampal Subfields (ASHS) for hippocampal regions. Analytic samples were 385 participants for DTI (265 features), 534 for FreeSurfer (323 features), and 534 for ASHS (35 features), with 355 participants common to all three sets.

On their own, these IDPs classified Alzheimer's disease status well. Best AUCs were 0.853 ± 0.063 for ASHS with Lasso, 0.952 ± 0.036 for DTI with XGBoost, and 0.922 ± 0.024 for FreeSurfer with a neural network. Combining all three in one integrated neural-network model raised AUC to 0.967 ± 0.018. The paper supports that gain with a correlation analysis: cross-modality agreement was low, with mean absolute correlations of 0.112 for DTI–FreeSurfer, 0.085 for DTI–ASHS, and 0.116 for FreeSurfer–ASHS, and block-level RV coefficients between 0.07 and 0.18. The authors read this as evidence that the modalities capture complementary rather than duplicate aspects of brain structure.

Predicted imaging features explain only a small share of variance

The second stage asks whether genotype can predict IDPs at all. Across the three IDP sources, a neural network gave the best balance of explained variance and error. For DTI it reached R² = 0.072 ± 0.019; for FreeSurfer and ASHS, XGBoost had the highest R² at 0.033 ± 0.007 and 0.051 ± 0.004, with the neural network close behind at 0.032 ± 0.008 and 0.051 ± 0.003 while minimizing error. Lasso performed worse, with lower R² and higher mean squared error.

Those R² values of roughly 0.03 to 0.07 are low in absolute terms, and the paper does not present them as accurate reconstruction of a person's scan. The authors argue this is expected in statistical genetics, where genotype-based models usually explain only a small fraction of variance in complex traits, and they frame the goal as isolating the heritable, genetically regulated component of each IDP rather than reproducing non-genetic noise. As supporting evidence, they report positive correlations between real and predicted IDPs in GWAS effect sizes across all modalities.

External validation favors G2DBridge over polygenic scores

The decisive comparison is the one run on genetic-only data. Models were trained and tuned on ADNI with five-fold cross-validation, and the Alzheimer's Disease Sequencing Project (ADSP) version 5 was reserved exclusively for external validation, with individuals appearing in both datasets removed from ADSP. The genotype-only ADSP cohort contained 25,113 individuals. Polygenic scores came from the PGS Catalog, computed with published SNP effect sizes from large-scale GWAS, including multiple scores derived from large European cohorts and one trained on multi-ancestry samples.

In ADNI internal validation, genetically inferred IDPs achieved higher AUC than all genetic-only baselines, though individual IDP-mediated models gave mixed results. In the ADSP external validation, G2DBridge outperformed multiple state-of-the-art polygenic risk score models while using the same target genetic data. The paper's explanation is that the two-stage design acts as an information bottleneck: predicting disease directly from a large SNP set is hard because there are far more variables than samples, so a direct model tends to memorize genetic noise and overfit, visible in higher validation loss. Routing prediction through imaging features restricts the model to genetic signals that map onto brain-structure change.

Separation of cases and controls matters at clinical thresholds

Beyond AUC, the paper reports two clinically oriented analyses. Distributional comparisons of normalized prediction scores showed significantly clearer separation between Alzheimer's cases and controls for G2DBridge than for polygenic risk score models. Decision Curve Analysis, which weighs net benefit against a risk threshold, showed that at clinically meaningful thresholds of 20% and 30%, most baseline models dropped to zero or negative net benefit, while G2DBridge maintained a positive benefit in most folds.

The authors interpret this as potential usefulness for prioritizing individuals for further screening or clinical follow-up. That is a statement about ranking and triage within the studied cohorts, not a claim that the model diagnoses Alzheimer's disease from DNA alone. The relevant comparison is between G2DBridge and polygenic scores on identical genetic inputs, and the reported advantage is in that comparison.

The model leans on regions consistent with Alzheimer's pathology

Feature-attribution analysis examined which inputs drove predictions. For DTI, diffusivity metrics from the cingulum–hippocampal tract and the uncinate fasciculus were among the strongest predictors, aligning with prior reports of white matter disruption in Alzheimer's disease. For FreeSurfer, cortical and subcortical features from temporal regions repeatedly ranked highly. For ASHS, medial temporal lobe structures including parahippocampal gyrus, entorhinal cortex, and hippocampal subfields were highlighted, regions long described as early markers of Alzheimer's-related neurodegeneration.

The paper also checked whether model-derived risk scores recover known associations. Across structural volumetric measures, predicted risk was negatively associated with regional brain volume, significant for all ASHS regions and for FreeSurfer hippocampal volume (p < 0.001), with the same trend but not significance for FreeSurfer entorhinal and parahippocampal volumes. For DTI, higher mean diffusivity in the superior longitudinal fasciculus and cingulum cingulate bundle was associated with higher predicted risk (p < 0.001), and lower fractional anisotropy in the corpus callosum and its splenium was associated with higher predicted risk (p < 0.001). These are consistency checks against established directions of effect, not independent causal findings.

The gain is bounded by paired sample size

The framework's advantage depends on having a training cohort with matched genetics, imaging, and diagnosis, and the paper identifies paired sample size as the main constraint. Because the three IDP sets were only weakly correlated, the authors argue that performance is limited mainly by the number of paired samples rather than by the feature space, making larger genetic and imaging cohorts the key to further gains. The ADNI imaging subsets analyzed here were in the hundreds of participants, while external validation used a genetic-only cohort of 25,113.

The paper also positions Alzheimer's disease as a case study chosen for its strong genetic and imaging biomarkers, and states that the approach is generalizable to other neurodegenerative, cardiovascular, and psychiatric diseases where paired genetic and imaging data exist. That is a stated extension, not a demonstrated result. For Alzheimer's risk prediction specifically, the practical priority the evidence supports is this: when only genotype is available, a model that routes prediction through genetically inferred imaging features can add predictive power and net benefit at 20% and 30% risk thresholds relative to polygenic scores built on the same genetic data, provided a suitably sized paired cohort was available for training.

Sources used in this answer