Hypothesis

The observed age‑related decline in NAD+ is not a passive consequence of accumulating damage but an active, network‑level downregulation that reallocates cellular resources away from energetically costly maintenance programs toward survival‑priority pathways. In other words, the cell revises its "ambition budget" by coordinately suppressing NAD+‑dependent processes (DNA repair, sirtuin signaling, chromatin remodeling) while activating compensatory metabolic routes that sustain ATP production with lower NAD+ flux.

Mechanistic Insight

NAD+ sits at the hub of a highly interconnected metabolic‑signaling module that includes NAMPT, NMNAT, CD38, PARPs, and SIRT1‑7. If the downregulation is adaptive, we expect the aging interactome to preserve the overall topology of this module while shifting edge weights: consumption enzymes (CD38, PARPs) gain increased centrality, biosynthesis enzymes (NAMPT, NMNAT) show reduced but coordinated expression, and downstream sirtuin targets exhibit rewired connections to alternative deacetylases or AMPK‑driven pathways. Passive damage, by contrast, would erode edges randomly, decreasing clustering and modularity without compensatory rewiring.

Experimental Design

1. Construct age‑stratified PPI/metabolic networks for mouse muscle and human skin/brain using public interaction databases (STRING, BioGRID) layered with metabolite‑enzyme connections from Recon3D. Assign each node a age‑specific expression weight from GTEx, Tabula Muris, or single‑cell RNA‑seq datasets.
2. Encode tissue composition as a node feature: proportion of adipocytes, fibroblasts, myocytes, etc., derived from deconvolution algorithms (CIBERSORTx) applied to bulk tissue profiles.
3. Train a Graph Attention Network (GAT) to predict NAD+ concentration from the network topology and node features. Use Shapley value attention to attribute importance to each node and edge.
4. Compare two model families: (a) an adaptive‑rewiring model that allows edge‑weight modulation but penalizes loss of clustering coefficient; (b) a damage model that penalizes edge loss uniformly. Evaluate performance with cross‑validation across age groups.
5. Validate predictions by perturbing key hub nodes (e.g., CD38 knockdown, NAMPT overexpression) in cultured myocytes and measuring NAD+ flux and downstream sirtuin activity.

Predictions

• If NAD+ decline is adaptive: the GAT will achieve higher prediction accuracy when edge‑weight changes are constrained to preserve module modularity (ΔQ < 0.05) and when Shapley highlights coordinated up‑regulation of CD38/PARPs and down‑regulation of NAMPT/NMNAT. Compensatory pathways (e.g., increased glycolysis, enhanced NAD+ salvage via NRK2) will show gained centrality.
• If decline is due to passive damage: the best‑performing model will tolerate random edge loss, show decreased average clustering coefficient and increased path length, and Shapley values will highlight no consistent direction of change across NAD+ pathway genes.

Falsifiability

A clear falsification occurs if (i) the adaptive model does not outperform the damage model in cross‑validated NAD+ prediction despite controlling for tissue composition, or (ii) experimental perturbation of predicted hub shifts fails to alter NAD+ levels in the direction forecast by the model.

Controls for Tissue Composition

By explicitly feeding cell‑type proportion features into the GAT and performing ablation studies (removing these features), we can test whether apparent NAD+ changes persist after accounting for shifts in adipocyte/fibroblast content. A significant drop in model performance when composition features are removed would indicate confounding; persistence would support a cell‑intrinsic network effect.

Expected Outcomes

We anticipate that incorporating tissue composition will reduce the apparent NAD+ decline attributable to cellular heterogeneity by ~30%, revealing a residual, coordinated rewiring signature. This would provide the first mechanistic, network‑level evidence that NAD+ loss reflects a strategic budgeting decision rather than indiscriminate wear‑and‑tear.

**References

• Graph-in-Graph for DTI: https://arxiv.org/abs/2507.11757
• Two-stage pre‑training for GNN‑DTI: https://pmc.ncbi.nlm.nih.gov/articles/PMC11439641/
• NAD+ consumption and biosynthesis in aging: https://pmc.ncbi.nlm.nih.gov/articles/PMC7442590/
• Graph attention and interpretability tools: https://pmc.ncbi.nlm.nih.gov/articles/PMC12659342/