Hypothesis

We hypothesize that a degree‑corrected graph neural network trained on the human longevity interactome will prioritize small molecules that modulate the ULK1‑PIK3C3 complex to shift autophagic flux from bulk recycling to selective resource reallocation, thereby mimicking a cellular siege response rather than generic cleanup.

Rationale

Longevity‑associated proteins form high‑degree hubs (median degree 7.0 vs 5.0 in the core PPI) [3], which can bias standard GNN message‑passing toward topology‑driven predictions. By penalizing high‑degree nodes during aggregation we force the model to learn functional patterns that survive degree normalization. Autophagy activation under nutrient stress is not a housekeeping act; ULK1/PIK3C3 dephosphorylation redirects organelles and metabolites to sustain ATP production [4][5]. This triage‑like flux redistribution determines whether a cell adapts to chronic siege or slips into senescence‑associated secretory phenotype. Most screens still rely on LC3‑II puncta, ignoring the metabolic rewiring that decides outcome.

Experimental Design

1. Network preparation – Assemble the 2 338‑node, 3 271‑edge longevity PPI [3]. Compute node degree and generate a degree‑corrected adjacency matrix where message weights are inversely proportional to log(degree+1).
2. Model architecture – Use a GraphSAGE‑style GNN with two message‑passing layers. Replace the standard mean aggregator with the degree‑corrected version. Train on known drug‑target pairs from BindingDB (filtered for human proteins) to predict binding affinity to ULK1 and PIK3C3.
3. Virtual screening – Dock the top‑500 predicted binders against the ULK1‑PIK3C3 interface using AutoDock Vina; retain compounds with ΔG < ‑7 kcal/mol.
4. Flux validation – Treat human fibroblasts with 10 µM of each candidate under EBSS starvation. Measure ¹³C‑glucose → TCA intermediates and ¹⁵N‑glutamine → aspartate flux via LC‑MS. Calculate the ratio of oxidative TCA flux to glycolytic flux as a readout of siege‑mode metabolism.
5. Phenotypic readout – Quantify senescence (β‑galactosidase) and SASP (IL‑6, IL‑8 ELISA) after 72 h treatment.

Predictions

• Degree‑corrected GNN will outperform a baseline GraphDTA (AUC increase > 0.07) specifically for ULK1/PIK3C3 targets, indicating that predictive power persists after topology bias removal.
• Top hits will increase oxidative TCA flux by ≥ 30 % relative to untreated starved cells while decreasing LC3‑II puncta accumulation, confirming a shift from bulk autophagy to selective resource reallocation.
• Cells treated with effective compounds will show lower SA‑β‑gal positivity (< 15 % vs > 40 % in controls) and reduced SASP cytokines, linking flux rewiring to delayed senescence.
• Inactive compounds (randomly selected from the same chemical space) will not alter flux ratios nor senescence markers.

Potential Pitfalls & Mitigations

• Over‑correction risk: Excessive penalization could erase true biological signals. We will test a range of penalty exponents (0.5–2.0) and select the one that maximizes validation AUC on a held‑out set of known aging drug targets.
• Compound solubility: Hydrophobic hits may aggregate in EBSS. Include a DMSO‑control titration step and discard any condition where turbidity exceeds 0.1 AU at 600 nm.
• Cell‑type specificity: Fibroblasts may not capture neuronal autophagy dynamics. Repeat key flux assays in iPSC‑derived neurons to assess generality.

If the degree‑corrected GNN fails to enrich for flux‑shifting molecules, the hypothesis that autophagy inhibition represents a siege‑rationing mechanism would need revision, suggesting that observed longevity effects stem from unrelated pathways. Conversely, success would validate targeting the ULK1‑PIK3C3 node as a rationing lever, opening a new class of gerotherapeutics that modulate the cell’s internal siege response rather than merely boosting autophagic flux.