Hypothesis: Sex‑Stratified Hormonal Confounding Inflates GNN Predictions for Aging Drug Targets

Core claim When graph neural networks (GNNs) are trained on static protein‑protein interaction (PPI) graphs to prioritize aging‑relevant drug targets (e.g., mTOR, SIRT1, AMPK), they inadvertently learn sex‑specific co‑expression patterns that arise from hormonal regulation rather than true causal binding propensity. This leads to inflated performance metrics for targets embedded in estrogen‑ or androgen‑responsive modules and poor generalizability to the opposite sex or to populations with divergent hormone levels.

Mechanistic rationale Aging is driven by multiple independent damage processes that require distinct repair strategies. Hormonal signaling, especially estrogen and androgen pathways, rewires interactome topology in a sex‑dependent manner: estrogen‑responsive genes (e.g., ESR1, GREB1) gain transient high‑weight edges to metabolic regulators, while androgen‑responsive nodes (e.g., AR, FKBP5) show similar flux. Standard GNNs that treat edge weights as static or derived solely from baseline expression ignore these dynamics, so the model attributes predictive power to structural motifs that are actually confounded by hormone‑driven co‑variation.

Testable predictions

1. Sex‑stratified performance gap – A GNN trained on a mixed‑sex PPI network will achieve significantly higher AUROC for aging targets when evaluated on a ligand set derived from male‑biased experimental data versus female‑biased data (or vice‑versa).
2. Confounder ablation – Removing the top principal components of sex‑stratified gene expression (as done in MRPC/GMAC 2) from node features will reduce the AUROC gap by at least 15 % without compromising overall accuracy.
3. Dynamic edge weighting – Incorporating sex‑specific edge weights computed from hormone‑responsive co‑expression networks (e.g., using context‑specific PPI scores from STRING filtered by estrogen/androgen response elements) will produce a GNN whose calibration curves converge across sexes and whose feature importance shifts toward known causal aging pathways (mTORC1, NAD+ salvage, autophagy) rather than hormone‑receptor clusters.
4. External validation – Prospective testing of the top‑5 predicted targets in primary hepatocytes from male and female donors will show comparable hit rates (e.g., ≥30 % confirming geroprotective activity) only when the model incorporates sex‑stratified confounding correction.

Experimental design

• Build three PPI graphs: (A) static baseline (average expression), (B) static + sex‑stratified PC covariates as extra nodes (following 2), (C) static + sex‑specific edge weights derived from hormone‑responsive expression (PEARSON > 0.7 in estrogen‑ or androgen‑treated RNA‑seq).
• Train identical GNN‑DTA architectures (e.g., equivariant GNN with Coulomb/Lennard‑Jones terms 1) on ligand‑affinity data from ChEMBL, enforcing ligand‑level splits 3 and site‑stratified CV 4.
• Evaluate on held‑out test sets partitioned by sex of the assay source (male‑derived vs female‑derived kinetic data).
• Statistical comparison using DeLong test for AUROC differences and calibration‑slope regression.

Falsifiability If the AUROC gap does not exceed 5 % between sexes, or if adding sex‑stratified PCs or edge weights fails to reduce the gap, the hypothesis is refuted. Likewise, if hormone‑responsive edge weighting does not improve calibration or shift feature importance away from hormone receptors, the mechanistic claim loses support.

Impact Demonstrating that hormonal confounding inflates GNN‑based aging target predictions would justify routine sex‑aware graph construction in geroprotective drug discovery, aligning machine‑learning pipelines with the biological reality that aging networks are hormonally plastic.