A Structural Theory of Admissible Transitions in Biological Reaction Networks

What the math gives to ML

The paper introduces a structural abstraction of continuous reaction-network trajectories: replace a nonnegative state by its positive-support set and analyze transitions between support sets using stoichiometric reaction structure rather than kinetic parameters. Its most transferable asset is the asymmetric treatment of support growth and support loss: the least reaction-closed superset is structurally determined, while later reductions depend on the realized trajectory. This suggests a constrained dynamic-sparsity or mixture-of-experts router in which activation supports can expand only through a predeclared dependency graph, while pruning remains data-dependent. The transfer is exploratory rather than direct, but it offers a principled way to prevent unreachable experts or channels and stabilize changing sparse computation graphs.