Hypothesis

Intermittent, nutrient‑contextualized mTOR inhibition rewires specific functional modules of the mTOR protein‑interaction network (PIN) in a temporally ordered fashion, producing a hormetic shift toward stress resistance without sacrificing the translational “yellow” module’s baseline activity. Continuous suppression locks the PIN in a survival‑only configuration, degrading specialized functions. A temporal graph attention network (TempGNN) trained on time‑resolved PIN data across fed‑fast‑refeed cycles will predict which intermittent dosing schedules maximize longevity‑associated outcomes (e.g., healthspan, stress‑resistance biomarkers) in model organisms.

Mechanistic Rationale

The mTOR PIN is not a monolithic switch but a set of context‑dependent modules—e.g., a translation‑driven yellow module (EIF4E/p70S6) and a stress‑responsive blue module (mTOR‑Raptor)【https://www.frontiersin.org/journals/genetics/articles/10.3389/fgene.2024.1452339/full】. Nutrient fluctuations trigger non‑linear, sequential disassembly and reassembly of these modules: during starvation, the yellow module attenuates while the blue module persists; upon refeeding, rapamycin‑sensitive blue module disassembly precedes yellow module recovery, creating a window where translation is partially restrained but stress signaling remains active【https://pubs.rsc.org/en/content/articlelanding/2020/ra/d0ra02297g】. This dynamic produces a transient “civilization‑lite” state where cells retain enough biosynthetic capacity for repair while activating survival pathways. Continuous mTOR inhibition, however, suppresses both modules indiscriminately, pushing the network into a static survival attractor that attenuates specialized functions such as synaptic plasticity or immune surveillance.

A TempGNN that incorporates edge‑level attention over time can capture these sequential rewiring patterns. By feeding it timestamped co‑IP or proximity‑labeling data (e.g., quantitative multiplex co‑IP across 0, 2, 6, 12 h of fasting/refeeding) the network learns to associate specific temporal motifs of module interaction strength with downstream phenotypes like autophagy flux, senescence markers, or IGF‑1 signaling【https://pmc.ncbi.nlm.nih.gov/articles/PMC10594569】【https://pmc.ncbi.nlm.nih.gov/articles/PMC10418199】. Importantly, the attention weights can be interpreted as the relative contribution of each module at each time point, providing a mechanistic readout of the “dial” position.

Testable Predictions

1. Prediction – TempGNN models trained on pulsed rapamycin (e.g., 2 h on/22 h off) will assign higher predictive power to intermittent regimens that preserve yellow‑module activity during early refeeding while enhancing blue‑module stress signaling during late fasting, correlating with improved median lifespan and delayed frailty in C. elegans or mice.
   • Falsification – If continuous rapamycin yields equal or better TempGNN‑predicted longevity scores than any pulsed schedule, the hypothesis is refuted.
2. Prediction – The temporal motif identified by the TempGNN (e.g., transient yellow‑module suppression followed by blue‑module rebound) will be necessary for longevity; genetically disrupting the blue module’s re‑assembly (via Raptor knock‑down) will abolish the benefit of intermittent mTOR inhibition despite identical drug exposure.
   • Falsification – If longevity persists despite blocked blue‑module dynamics, the hypothesized mechanism is incorrect.
3. Prediction – Feature importance analysis will show that edge‑weight changes between mTOR and EIF4E (yellow) and between mTOR and Raptor (blue) across specific time windows are the top predictors of healthspan, outperforming static network metrics.
   • Falsification – If static metrics (e.g., average degree) outperform temporal features, the temporal GNN adds no value.

Potential Confounds and Controls

• Control for off‑target drug effects by using structurally distinct mTORC1 inhibitors (e.g., Torin1) and genetic knock‑downs.
• Measure autophagy flux, lysosomal activity, and immune cytokine panels to ensure observed phenotypes are not secondary to generic stress.
• Use isogenic lines and litter‑matched controls to eliminate genetic background noise.
• Validate TempGNN predictions in an external cohort (different species or strain) to test generalizability.

By framing mTOR as a dial whose position is read out by the temporal topology of its interactome, this hypothesis bridges network pharmacology, geroscience, and dynamic deep learning, offering a concrete, falsifiable roadmap for designing interventions that extend lifespan without surrendering the organism’s civilizational functions.