Hypothesis

A relational graph convolutional network (R-GCN) that incorporates adversarial domain adaptation to remove tissue‑ and age‑batch effects, coupled with causal mediation analysis linking predicted drug‑target interactions to senescence‑associated secretory phenotype (SASP) output, will prioritize proteins whose perturbation causally reduces senescent cell burden in human fibroblasts. Conversely, targets selected by standard GNNs lacking these controls will show no consistent effect on SASP markers.

Mechanistic rationale

The 61‑protein, 120‑interaction druggable aging subnetwork[2] captures multi‑layer biology: protein‑protein interactions (PPI), transcriptional regulation, and metabolic pathways. Standard GNNs treat all edges uniformly, conflating distinct mechanistic routes and ignoring known confounders such as tissue‑specific expression batches and donor age variability[1]. By assigning edge‑type‑specific weight matrices, an R‑GCN can learn distinct transformations for PPI versus regulatory edges, preserving the directional influence of transcription factors on metabolic enzymes that drive senescence[3].

Adversarial domain adaptation layers will be trained to minimize the ability of a domain classifier to predict tissue source or chronological age from node embeddings, thereby forcing the network to encode aging‑relevant signals rather than batch artefacts[1]. Causal mediation analysis will then test whether the effect of a predicted drug‑target interaction on a functional aging readout (e.g., SA‑β‑gal positivity) is mediated through changes in SASP cytokine secretion (IL‑6, IL‑8). This moves beyond correlation‑based affinity metrics[1] to a causal chain: drug target → network perturbation → SASP modulation → phenotypic outcome.

Testable predictions

1. Prediction ranking – In age‑cold split validation (training on donors <45 years, testing on donors ≥65 years), the R‑GCN with adversarial debiasing will achieve a higher precision@10 for targets whose siRNA knock‑down reduces SASP secretion by ≥30 % compared to baseline, whereas a vanilla GNN (e.g., GPS‑DTI) will not exceed random precision.
2. Mediation significance – For the top‑5 R‑GCN predicted targets, the indirect effect of target knock‑down on SA‑β‑gal positivity via SASP reduction will be statistically significant (p < 0.01, bootstrap CI), while the direct effect will be non‑significant, supporting a mediation model.
3. Specificity of edge types – Ablation experiments removing either the transcriptional‑regulation edge set or the metabolic edge set from the R‑GCN will cause a measurable drop in mediation effect size (>20 %), demonstrating that both layers contribute to causal inference.
4. Falsifiability – If the adversarial debiasing layers are removed and the model retrained, the precision@10 for SASP‑reducing targets will fall to the level of the baseline GNN, confirming that batch correction is necessary for the observed improvement.

Experimental validation plan

• Data – Use the multilayer aging subnetwork[2] as the static graph; overlay tissue‑specific PPI, regulatory, and metabolic edges from public databases (STRING, TRANSFAC, Reactome).
• Model – Implement an R‑GCN with three edge‑type transforms, an adversarial loss gradient reversal layer, and a mediation head predicting SASP cytokine levels.
• Validation splits – Create tissue‑cold (leave‑one‑tissue‑out), age‑cold (young vs. old donor), and drug‑cold (known senescence modulators held out) splits.
• Orthogonal assays – Transfect senescent human fibroblasts (induced by irradiation) with siRNA against each top‑ranked target; measure SA‑β‑gal, p16^INK4a^, and secreted IL‑6/IL‑8 via ELISA after 72 h.
• Analysis – Compute precision, recall, and mediation effect sizes; compare against vanilla GNN and random siRNA controls using paired statistical tests.

If the hypothesis holds, the workflow will provide a rigorously vetted shortlist of targets whose network‑level perturbation demonstrably ameliorates senescence phenotypes, establishing a template for causal, confounder‑aware GNN applications in aging drug discovery.