We hypothesize that aging‑specific drug target prediction fails with static GNNs because they ignore the time‑dependent rewiring of protein interactomes and the cell‑type‑specific context that drives degenerative phenotypes. By integrating longitudinal multi‑omics data (transcriptome, proteome, phosphoproteome) from young to old individuals into a dynamic hypergraph neural network (DHGNN) that models higher‑order interactions and weights edges by temporal change and cellular abundance, we will obtain representations that capture aging‑driven network topology shifts. Drugs will be encoded with 3D conformers and directional interaction features, allowing the DHGNN to learn predictive patterns that generalize to unseen compounds and targets.

Static GNNs suffer from structural leakage: edge removal splits preserve network motifs that let the model retrieve known interactions rather than discover new ones. Temporal hypergraphs break this leakage because each time slice presents a distinct wiring pattern; removing an edge in one age‑specific layer does not expose the same topology in another layer, forcing the model to infer genuine age‑dependent associations. Moreover, cold‑start failure stems from learning dataset‑specific degree biases that over‑represent well‑studied proteins. By weighting nodes with longitudinal expression variance and cell‑type specificity, the DHGNN reduces the influence of static hub bias and forces the model to rely on dynamic signatures that transfer across drugs and targets. Incorporating 3D drug conformers and direction‑aware interaction descriptors further corrects the omission of spatial and orientational information that plagues SMILES‑only or sequence‑only representations.

To evaluate the hypothesis we will first assemble tissue‑specific temporal PPI hypergraphs. Using publicly available aging atlases such as GTEx and the Human Aging Atlas, we will extract matched transcriptome, proteome and phosphoproteome profiles from donors stratified into five age bins (20‑30, 40‑50, 60‑70, 80‑90, 100+ years). For each bin we will compute differential abundance and co‑variation scores to construct weighted hyperedges that group proteins showing concerted age‑dependent changes. These hypergraphs will be stacked into a temporal series and fed into a DHGNN architecture that employs hypergraph convolution layers followed by a temporal attention mechanism. Drug nodes will be generated from RDKit‑derived 3D conformers, enriched with electrostatic potentials and hydrogen‑bond directionality, and concatenated with protein hypergraph embeddings via a bilinear decoder. The model will be trained on known drug‑target pairs from BindingDB and Davis, but validated exclusively through prospective leave‑one‑age‑out cross‑validation: all interactions from the oldest age bin are held out during training and used only for final testing.

Predicted interactions will be ranked and the top 50 compounds per age bin will be taken forward for experimental validation. In Caenorhabditis elegans we will employ RNAi knock‑down of the predicted target combined with drug exposure, measuring mean lifespan, thrashing rate and accumulation of age‑related lipofuscin. Parallel studies in male C57BL/6 mice will use CRISPRi to dampen target expression in liver and administer the compound, assessing frailty index, grip strength and plasma inflammatory cytokines. A outcome is considered supportive if at least three independent drug‑target pairs show a statistically significant (p<0.05, Bonferroni‑corrected) extension of lifespan or delay of two or more age‑related phenotypes relative to controls. Conversely, if none of the tested pairs produce a reproducible phenotypic improvement across both species, the hypothesis that temporal hypergraph modeling captures biologically relevant aging‑specific drug‑target interactions would be falsified.