Hypothesis: Incorporating explicit confounder nodes for age and sex into a GNN architecture that iteratively unrolls and de-confounds multi‑omics networks will eliminate spurious edges caused by shared upstream regulators, thereby yielding drug‑target affinity predictions that are causally grounded and generalize across chronological cohorts. Specifically, we predict that a GNN built on aging‑specific protein‑protein interaction (PPI) networks, augmented with GMAC‑style principal‑component confounder nodes and METALICA‑style unrolling layers, will achieve (i) a ≥15% reduction in false‑positive rate (FPR) for known non‑aging drug targets, (ii) a ≥10% increase in area under the precision‑recall curve (AUPRC) for aging‑specific targets when evaluated on temporally held‑out test sets, and (iii) no significant performance drop when the same model is applied to an independent dataset collected in a different sequencing batch.

Mechanistic rationale: Age is a pervasive latent variable that simultaneously drives global changes in gene expression, metabolite levels, and network topology [5][8]. Standard correlation‑based GNNs interpret any co‑variation between two proteins as a direct interaction, mistaking age‑driven co‑regulation for causal edges [5]. By modeling age (and sex) as explicit graph nodes whose attributes are derived from principal components of batch‑corrected omics data, we allow the GNN to condition on these confounders during message passing. The unrolling step then iteratively removes indirect paths that pass through these confounder nodes, isolating direct causal influences akin to METALICA’s de‑confounding on multi‑omics IBD networks [6]. This approach directly addresses the open question of scaling GNNs to dynamic, cyclic networks with latent confounders [8].

Experimental design: 1) Construct sex‑stratified PPI networks from HAGR and GTEx data, separating samples into young (20‑35 y) and old (60‑75 y) cohorts. 2) Derive confounder nodes for each sample using the top 5 PCs of genotype‑expression and metabolomics matrices, following GMAC [7]. 3) Implement a GNN with two novel layers: (a) a confounder‑conditioned message‑passing step that aggregates neighbor features weighted by confounder similarity, and (b) an unrolling‑deconfounding block that iteratively recomputes edge weights after subtracting confounder‑mediated paths (as in METALICA) [6]. 4) Train the model to predict drug‑target binding affinity from DGIdb and ChEMBL, using CASTER‑DAPT as a baseline [3][4]. 5) Evaluate using (i) chronological hold‑out (train on young, test on old) and (ii) batch hold‑out (train on batch A, test on batch C) splits, reporting FPR, AUPRC, and calibration error. 6) Perform permutation testing where confounder node attributes are shuffled to confirm that performance gains depend on correct confounder modeling.

Falsifiability: If the proposed architecture fails to reduce FPR or improve AUPRC relative to baseline GNNs under both chronological and batch splits, or if performance gains disappear when confounder nodes are correctly included, the hypothesis is refuted. Conversely, consistent improvements across splits would support the claim that explicit confounder modeling and causal unrolling are necessary for reliable GNN‑based aging drug‑target discovery.