Hypothesis

Core claim: Incorporating sex‑specific epigenetic edge weights derived from methylation clocks into GNNs will rescue ranking performance (Hits@n) while reducing confounding bias in aging drug‑target prediction.

Rationale

• Existing GNNs achieve high accuracy but near‑zero Hits@n on DDI benchmarks [1]
• This paradox stems from models learning static topology that conflates true causal edges with confounding paths driven by latent variables such as sex, immune status, and diet [2]
• Mendelian randomization shows that unmeasured regulators bias edge inference [3]
• Prior work attempts to add confounder nodes [4] but remains untested in aging contexts.

Mechanistic Insight

We hypothesize that age‑related DNA‑methylation changes act as dynamic modifiers of protein‑protein interactions, effectively re‑weighting edges in the interactome in a sex‑dependent manner. When these epigenetic states are ignored, GNNs treat a static graph as if it were i.i.d., causing the model to assign high scores to many nodes that are merely correlated with confounding factors rather than causal drivers of aging. By encoding sex‑stratified methylation quantitative trait loci (meQTL) scores as edge attributes, the GNN can learn to down‑weight spurious confounder‑driven connections and up‑weight those that persist across epigenetic states, thereby improving the ability to rank true positives.

Testable Predictions

1. Ranking rescue: A GNN trained on a drug‑aging interactome augmented with sex‑specific meQTL edge weights will achieve Hits@10 ≥ 0.30 on a held‑out aging‑specific benchmark (e.g., DrugAge or Geroprotectors), whereas the same architecture without epigenetic weights will remain ≤ 0.02.
2. Confounder attenuation: The variance explained by sex and immune covariates in the model’s latent representations will drop by at least 40 % when epigenetic edge weights are included, as measured by linear probing.
3. Generalizability: Performance gains will persist under inductive splits where entire age‑cohorts are left out, demonstrating that the model captures causal, not merely correlative, patterns.

Experimental Design

• Build a multi‑layer interactome where each edge weight = baseline confidence × (1 + β_sex × ΔmeQTL_sex). ΔmeQTL derived from publicly available blood‑brain methylome QTL studies stratified by sex.
• Use a standard GNN backbone (e.g., GraphSAGE) to ensure any improvement is not due to architectural complexity.
• Train on known aging‑modulating compounds from DrugAge, validate on an independent set of geroprotectors from Geroprotectors.org.
• Report both classification accuracy (AUC) and ranking metrics (Hits@n, MRR). Perform ablation: remove epigenetic weights, shuffle sex labels, replace with random edge weights.

Falsifiability

If the augmented GNN fails to improve Hits@n beyond the baseline (≤ 0.02) or does not reduce covariate variance in latent space, the hypothesis is falsified. Conversely, a significant improvement would support the claim that sex‑stratified epigenetic confounding underlies the accuracy‑ranking paradox and that modeling it causally rescues predictive utility for aging drug discovery.