Hypothesis

Integrating age-stratified, post‑translational modification (PTM)-aware edge weights and causal adjustment for batch effects into graph neural networks (GNNs) will yield significantly better and more calibrated predictions of drug‑target interactions in aging contexts than standard GNNs trained on static protein‑protein interaction (PPI) networks.

Mechanistic Rationale

Aging reshapes the interactome not only by altering protein abundance but also by rewiring PTM patterns—phosphorylation, acetylation, and ubiquitination—that create or destroy binding interfaces in a age‑dependent manner [1][2]. Static PPI networks treat every edge as invariant, causing GNNs to learn spurious correlations driven by confounding variables such as site‑specific batch effects or cohort‑level expression shifts [3][4]. When these confounders align with both network topology and drug‑response labels, the model inflates importance of hub nodes that are merely batch artifacts, a problem exacerbated by node‑degree bias and temporal leakage in knowledge graphs [5][6].

By constructing edge weights that reflect quantitative PTM changes measured across age strata (e.g., from phosphoproteomics or acetylomics datasets) and feeding these into a GNN, the model can distinguish true age‑specific interaction rewiring from uniform noise. Coupling this with a causal inference layer—such as inverse‑probability weighting derived from site‑stratified covariates—removes the influence of hidden batch variables while preserving signal, as demonstrated by RUVCorr’s ability to retain true biological relationships [2].

Testable Predictions

1. Performance Gain: On a held‑out set of drug‑target pairs curated from aging‑relevant studies (e.g., geroprotector screens), the PTM‑aware, causally adjusted GNN will achieve a higher AUROC and AUPRC than a baseline GNN using a static PPI network, with p < 0.01 in a paired DeLong test.
2. Calibration Improvement: Expected calibration error (ECE) will decrease by at least 20 % for the PTM‑aware model, indicating better probability estimates.
3. Confounding Robustness: When site‑batch labels are permuted, the performance advantage of the PTM‑aware model will disappear, confirming that the gain stems from confounding adjustment rather than mere added features.
4. Mechanistic Specificity: Shuffling PTM edge weights while preserving network topology will abolish the performance improvement, demonstrating that the age‑specific PTM information, not just extra edge density, drives the gain.

Experimental Design

• Data: Compile a multi‑cohort dataset of human PPIs (static), age‑stratified PTM quantifications (e.g., from CPTAC aging cohorts), and drug‑target binding affinities (Kd or IC50) from ChEMBL and aging‑focused screens.
• Models: (a) Baseline GNN (e.g., PhysDual‑GCN) on static PPI; (b) PTM‑aware GNN where edge features = static adjacency × age‑specific PTM score; (c) PTM‑aware GNN + causal adjustment layer (inverse‑probability weighting based on site, batch, and demographic covariates).
• Validation: Use ligand‑level splits to prevent leakage, and construct age‑stratified hold‑out sets ensuring no overlap of drug compounds or target proteins across folds. Evaluate with AUROC, AUPRC, and ECE.
• Falsifiability: If the PTM‑aware, causally adjusted model does not outperform the baseline by a statistically significant margin, or if performance persists after PTM shuffling, the hypothesis is falsified.

This hypothesis directly addresses the field’s reliance on static topology and insufficient confounder control, proposing a concrete, experimentally tractable route to improve GNN‑based drug‑target discovery for aging.