Hypothesis

Incorporating tissue‑specific metabolic state variables (AMPK activity, mTORC1 signaling, cytosolic acetyl‑CoA concentration) as dynamic node features in graph neural networks will significantly improve the prediction of autophagy‑modulating drug efficacy across aged tissues, whereas models that ignore these variables will systematically overestimate benefit in metabolically stressed organs such as liver and underestimate it in reserve tissues like skeletal muscle.

Mechanistic Basis

Autophagy functions as a rationing system that is toggled by the cell’s perception of siege conditions. Nutrient‑sensing kinases (AMPK, mTOR) and metabolite sensors (acetyl‑CoA) directly gate the transcriptional and post‑translational control of core autophagy genes (e.g., BECN1, LC3) [4, 5]. In aged cells, chronic elevation of acetyl‑CoA suppresses Atg gene expression, while AMPK activation can override this block [5]. Consequently, the same drug that stimulates autophagy initiation may be futile in a tissue with high acetyl‑CoA/low AMPK (e.g., aged liver) but protective where AMPK dominates (e.g., aged muscle). Current GNN architectures for drug‑target interaction predict binding affinity in a static topological context and lack any representation of these metabolic switches [1, 2].

Prediction

A GNN that augments each protein or drug node with three continuous features—phospho‑AMPK/total AMPK ratio, phospho‑S6K/total S6K ratio (mTORC1 read‑out), and measured cytosolic acetyl‑CoA—will achieve a statistically significant increase in AUROC (≥0.07 improvement) when classifying compounds as autophagy‑effective versus ineffective in aged mouse tissues, compared to a baseline GNN that uses only sequence‑ and network‑derived features.

Experimental Design

1. Data curation – Compile a dataset of FDA‑approved and experimental compounds with reported effects on autophagy flux (LC3‑II turnover, p62 degradation) in primary cells from aged (≥24 mo) mouse liver, skeletal muscle, heart, and brain. Include quantitative tissue‑specific measurements of phospho‑AMPK, phospho‑S6K, and acetyl‑CoA from the same samples [3].
2. Model construction – Build two graph architectures:
   • Baseline: heterogeneous DTI-HETA‑style graph with GCN embeddings and GAT attention, using only protein sequence similarity, chemical fingerprints, and known interaction edges.
   • Metabolic‑augmented: identical topology but each protein node appended with the three metabolic features; drug nodes retain chemical descriptors.
3. Training & validation – Use five‑fold cross‑validation stratified by tissue to predict binary efficacy (effective if flux increase >30 % vs control). Optimize hyper‑parameters independently for each architecture.
4. Statistical test – Compare AUROCs via DeLong’s test; a p‑value <0.01 indicates significant improvement.

Potential Outcomes

• If the metabolic‑augmented GNN outperforms the baseline, the hypothesis is supported: encoding siege‑state metabolism enables models to discern when autophagy modulation will be beneficial versus futile, bridging the current methodological chasm between DTI prediction and aging‑context autophagy biology.
• If performance does not improve, the hypothesis is falsified, suggesting either that the selected metabolic readouts do not capture the relevant regulatory node or that additional layers (e.g., lysosomal capacity, redox state) are required for accurate prediction.

This framework directly links the siege‑state conception of autophagy to a concrete, testable modification of graph‑based drug‑target modeling, providing a path toward context‑aware therapeutic discovery in aging.