Hypothesis

Sex-stratified topological leakage inflates GNN predictions for aging drug targets when transductive splits hide inductive failure.

Mechanistic reasoning

Aging-related protein-protein interaction (PPI) networks are modulated by sex hormones, creating sex-biased topological patterns (e.g., degree centrality of estrogen‑responsive hubs). Most PPI databases aggregate data from mixed‑sex sources without stratification, producing a static interactome that conflates true aging mechanisms with sex‑specific confounding. When GNNs are trained on these networks using random or edge‑based splits, the model can learn to predict drug‑target affinity by exploiting sex‑associated degree patterns rather than learning binding‑site features. Because the splits are transductive (nodes appear in both train and test), the model merely retrieves the sex‑biased topology that was present during training, leading to inflated performance on benchmarks like DUD‑E and PDBbind while failing on truly independent, sex‑balanced or temporally distinct sets such as ChEMBL and MUV.

Testable predictions

1. Models trained on mixed‑sex PPI data will show a significant drop in AUROC (≥0.20) when evaluated on a sex‑stratified hold‑out set where all test nodes belong to the opposite sex of the training nodes.
2. The performance gap will be largest for targets known to be hormone‑regulated (e.g., ESR1, AR) and minimal for sex‑neutral targets (e.g., housekeeping enzymes).
3. Removing node degree features or applying degree‑preserving edge rewiring will reduce the sex‑stratified performance gap, indicating that the model relies on topological proxies.
4. Incorporating explicit sex‑specific edge weights (derived from sex‑stratified PPI datasets) will improve generalization across sexes without sacrificing overall accuracy.

Experimental design

• Data construction: Extract two PPI networks from BioGRID, one comprising interactions reported primarily in male‑derived studies and the other in female‑derived studies (filter by PMID-associated sex metadata). Merge them to create a mixed‑sex baseline network.
• Model: Use a representative GNN (e.g., GraphSAGE) trained to predict drug‑target binding affinity from the mixed‑sex network, using identical hyperparameters across experiments.
• Splits: a. Standard random node split (transductive). b. Sex‑stratified split: train on male‑derived nodes, test on female‑derived nodes, and vice‑versa. c. Temporal split using publication dates to further control for leakage.
• Evaluation: Compute AUROC, AUPRC, and calibration error on each test set. Perform ablation studies where node degree is shuffled or replaced with random values.
• Validation: Prospectively test top‑ranked predictions from each model configuration in vitro using a panel of sex‑specific cell lines (e.g., MCF‑7 vs. LNCaP) to confirm whether predicted affinities hold across sexes.

If the hypothesis holds, we will observe a marked decline in performance under sex‑stratified splits, implicating sex‑biased topological leakage as a key source of overoptimistic GNN estimates for aging drug targets. Conversely, if performance remains stable, the hypothesis would be falsified, suggesting that other confounders (e.g., temporal bias or missing data) dominate the generalization gap.