The Graph Is Lying: Why Our Models Fail at Edge Dynamics

I spend my days mapping aging using Graph Neural Networks (GNNs), effectively treating protein-protein interaction networks like static transit maps. We build these layered architectures and feed them omics data, hoping the hidden layers will somehow extract an "aging signal." But I owe this community some honesty: we're fundamentally misrepresenting the biology.

We treat the interactome as a fixed topology, but aging isn’t a state—it’s a topological drift. Our current models struggle because they assume nodes and edges stay stable over time, aside from minor weight fluctuations. Yet, the data suggests that during aging, the connectivity of our cellular machinery undergoes a phase transition. We’re trying to predict the destination of a moving train while pretending the tracks are permanent.

How do we actually capture stochastic rewiring? How do we differentiate between an edge that’s just functionally dampened and one that’s been structurally dismantled? My models are great at finding "hubs," but they’re notoriously blind to the transient, low-degree interactions that likely signal the onset of proteostatic collapse.

I’m stuck on a few core problems:

• Can we build a GNN that treats dynamic edge formation as the primary variable, instead of focusing on node expression levels?
• Is the loss of network modularity a driver of senescence, or is the network simply "de-differentiating" to minimize metabolic overhead?
• Are we looking at the wrong biological resolution, hunting for global rules in a system that might be fundamentally localized and fragmented?

If we keep forcing static graphs to explain a kinetic disaster, we’re just curve-fitting our own ignorance. I’m starting to suspect that the most important features aren’t even in our current training sets—they’re the interactions that simply stopped occurring.