Hypothesis

A graph neural network (GNN) that integrates temporal edge weights derived from single‑cell SASP trajectories can prioritize small‑molecule modulators which reprogram senescent cells to secrete a pro‑repair secretome (high PDGF‑AA, CCN1/2; low IL‑1β, TNF‑α) without eliminating the cells. In diabetic wound models, treatment with GNN‑predicted modulators will accelerate closure, reduce fibrosis, and shift macrophage polarization toward M2, whereas senolytic clearance will impair early granulation tissue formation.

Mechanistic Rationale

Senescent cells function as dynamic signaling hubs whose secretome is rewired by feedback loops between NF‑κB, C/EBPβ, and chromatin remodelers [6]. These loops generate co‑expression patterns that evolve over hours to days, creating transiently beneficial SASP phases (e.g., PDGF‑AA‑driven myofibroblast differentiation) and later detrimental phases (IL‑1β‑driven inflammation). Static molecular graphs ignore this temporal topology, causing GNNs to learn spurious correlations between node presence and drug affinity.

We propose to encode each time point of a single‑cell secretome trajectory as a node and the correlation of SASP factor co‑expression between successive time points as a weighted edge. The resulting temporal graph captures causal influence: a high weight from IL‑1β at t₁ to TNF‑α at t₂ suggests a feed‑forward inflammatory cascade. By training a Graph‑in‑Graph (GiG) architecture on these temporal graphs paired with known SASP‑modulating compounds (CXCR2 antagonists, IL‑1R blockers, JAK inhibitors) we expect the GNN to learn edge‑weight perturbations that flip the trajectory from a maladaptive to a reparative attractor state.

Experimental Design

1. Data generation – Isolate senescent fibroblasts from murine diabetic wounds at 0, 24, 48, 72 h post‑injury; perform scRNA‑seq + secreted protein profiling (Olink) to build SASP co‑expression matrices [7].
2. Temporal graph construction – Nodes = SASP factors; edges = Spearman correlation between factor i at time t and factor j at time t+Δt; weight = absolute correlation.
3. Model training – GiG model takes molecular graph of a compound and the temporal SASP graph as input; output = predicted change in edge‑weight profile (ΔW). Loss function minimizes MSE between predicted ΔW and observed ΔW from compound‑treated senescent cultures.
4. Validation – Top‑5 predicted modulators tested in vivo: diabetic db/db mice with full‑thickness dorsal wounds receive daily topical application (n=10 per group). Controls: vehicle, dasatinib+quercetin (senolytic), and CXCR2 antagonist (SB‑225002) as a benchmark.
5. Readouts – wound area over time (planimetry), histology (Masson’s trichrome for fibrosis, α‑SMA for myofibroblasts), flow cytometry for macrophage M1/M2 (CD86/CD206), SASP multiplex (IL‑1β, IL‑6, TNF‑α, PDGF‑AA, CCN2).

Falsifiable Predictions

• If the hypothesis is correct, GNN‑predicted modulators will increase the PDGF‑AA/IL‑1β ratio in wound tissue by ≥2‑fold relative to vehicle at day 4, while senolytics will decrease total SASP factor intensity but also reduce PDGF‑AA, delaying closure.
• If the hypothesis is false, no compound will significantly shift the SASP trajectory without altering senescent cell abundance, and wound healing metrics will not differ from senolytic or vehicle groups.

Potential Impact

Successful validation would establish a computational framework for targeting the qualitative state of senescence rather than its quantity, aligning with the hostage‑negotiator view: senescent cells are kept, but their negotiating stance is altered to favor tissue repair. This approach could be extended to other chronic pathologies where SASP duality drives disease progression (e.g., atherosclerosis, neurodegeneration).