Hypothesis

NAD+ decline acts as a conditional resource‑allocation signal: in proliferating cells it limits repair capacity and drives aging, whereas in senescent cells it enforces a low‑energy state that suppresses the metabolically costly SASP. Restoring NAD+ uniformly therefore benefits healthy tissues but may inadvertently reactivate SASP in senescent cells, confounding intervention outcomes. We hypothesize that a graph neural network trained on multi‑omics networks stratified by cell‑state can identify NAD+‑modulating compounds that selectively enhance NAD+ in non‑senescent cells while preserving or further lowering NAD+ in senescent cells, thereby improving healthspan without aggravating inflammation.

Mechanistic Basis

NAD+ is consumed by PARPs during DNA repair and by CD38 in immune signaling, linking its levels to upstream damage [https://digitalcommons.wustl.edu/cgi/viewcontent.cgi?article=7625&context=open_access_pubs]. Low NAD+ reduces mitochondrial ATP production, which attenuates the SASP—a phenotype that depends on high glycolytic and secretory flux [https://onlinelibrary.wiley.com/doi/10.1111/acel.13920]. Thus, NAD+ depletion can be interpreted as a cell‑intrinsic budget cut that prioritizes survival over elaborate paracrine signaling. PARP inhibition rescues NAD+ and delays aging phenotypes, confirming that DNA damage drives the decline [https://pmc.ncbi.nlm.nih.gov/articles/PMC12177089]. However, NAD+ boosters such as NMN/NR ignore the divergent functional outcomes of NAD+ restoration across cellular contexts.

Modeling Approach

We will construct a heterogeneous graph where nodes represent drugs, proteins, metabolites, and transcripts, and edges capture known interactions (binding, regulation, correlation). Node features will include omics measurements separated by cell‑state labels (e.g., young proliferative, aged non‑senescent, senescent) derived from single‑cell RNA‑seq and proteomics atlases. A hierarchical GNN (similar to H²GnnDTI) will learn state‑specific embeddings, allowing the model to predict how a compound shifts NAD+ levels in each subpopulation. The loss function will penalize predictions that increase NAD+ in senescent nodes while rewarding those that raise NAD+ in non‑senescent nodes.

Testable Predictions

1. In silico: The GNN will rank known NAD+ modulators (e.g., PARP inhibitors, CD38 blockers, NAMPT activators) higher for non‑senescent cells and lower for senescent cells compared with random compounds.
2. In vitro: Treatment of human fibroblasts with top‑ranked predicted modulators will raise NAD+ and reduce γ‑H2AX foci in proliferating cells, while senescent cells treated with the same compounds will show unchanged or decreased NAD+ and reduced IL‑6/IL‑8 secretion (SASP suppression).
3. In vivo: Mice receiving the top‑ranked modulator will exhibit improved metabolic health (glucose tolerance, activity) without increased circulating SASP cytokines, whereas mice given a non‑selective NAD+ booster (NMN) will show mixed outcomes—improved metabolism alongside elevated serum IL‑6.

Falsifiability

If the GNN fails to discriminate compounds by cell‑state (i.e., predicted activity correlates equally with NAD+ changes in both senescent and non‑senescent cells) or if experimental validation shows no differential effect on SASP, the hypothesis is refuted. Conversely, a consistent cell‑state‑specific NAD+ shift coupled with divergent functional outcomes will support the idea that NAD+ decline is a context‑dependent budget cut rather than a uniform driver of aging.