Hypothesis

Static protein‑protein interaction (PPI) graphs used by current GNN‑based drug‑target affinity models miss age‑specific rewiring driven by phosphorylation‑dependent edge modulation, causing transductive evaluations to overestimate predictive power for aging drugs.

Mechanistic Rationale

Aging alters kinase activity and phosphatase expression, leading to widespread changes in phospho‑sites that switch PPIs on or off without changing the underlying amino‑acid sequence [[https://pubs.rsc.org/en/content/articlelanding/2020/ra/d0ra02297g]]. When a GNN is trained on a static adjacency matrix derived from generic contact maps or sequence‑based predictors, it learns correlations that reflect the dominant (often young‑adult) interactome topology. In transductive splits where test nodes are present during training, the model can implicitly memorize age‑biased edge patterns, inflating AUROC/AUPRC on benchmarks that do not stratify by age.

Testable Predictions

1. A GNN that incorporates longitudinal phosphoproteomic edge weights (e.g., from GTEx or MAP kinase activity assays across decades) will show significantly lower performance drop when evaluated on an held‑out older cohort (≥65 y) compared to a static‑edge GNN.
2. The performance gap between static and dynamic models will correlate with the magnitude of age‑dependent phospho‑edge changes measured in the same tissue.
3. Drugs predicted as high‑affinity aging targets by the dynamic GNN will be enriched for compounds with documented geroprotective effects in model organisms or early‑phase human trials, whereas static GNN predictions will not show this enrichment.

Experimental Design

• Data construction: Build tissue‑specific PPI networks for skeletal muscle and liver from GTEx RNA‑seq (ages 20‑30, 40‑50, 60‑70, 80‑90). Overlay phosphosite quantitative data from the same donors (or from public phosphoproteomics atlases) to compute age‑specific edge confidence scores (e.g., weighted by change in phosphorylation of interacting partners).
• Modeling: Implement two GNN variants using the same architecture (e.g., BridgeDPI). Variant A uses a static adjacency matrix (binary edges). Variant B uses edge features derived from the age‑specific phospho‑weights, allowing message passing to modulate edge strength dynamically.
• Training: Train both variants on drug‑target affinity data from Davis and KIBA, using only compounds with known targets irrespective of donor age.
• Validation: Create three test sets: (a) young‑adult (20‑30 y) held‑out targets, (b) middle‑aged (40‑50 y), (c) older (≥65 y). Ensure no overlap of drug or target identifiers between train and test.
• Metrics: Report AUROC, AUPRC with 95 % confidence intervals over five random seeds. Calculate Δ performance (static – dynamic) per age stratum.
• Follow‑up: Take top‑5 predicted aging‑specific targets from the dynamic model, test in vitro binding assays, and assess lifespan extension in C. elegans or murine healthspan readouts.

Potential Confounds and Controls

• Control for overall expression level changes by regressing out transcript abundance when computing edge weights.
• Ensure that the phospho‑weight matrix is not circularly derived from the same samples used for affinity labels; use independent donors.
• Verify that observed improvements are not driven solely by increased graph density; match average degree across static and dynamic graphs via edge‑thresholding.

If the dynamic GNN consistently outperforms the static counterpart in the older strata and its predictions align with empirical geroprotective activity, the hypothesis that static PPI graphs confound GNN‑based aging drug discovery will be supported. Conversely, a lack of age‑stratified performance difference would falsify the claim that phosphorylation‑driven interactome rewiring is a major source of bias in current models.