Neutral Networks

A neutral network is a connected set of genotypes (or phenotypes) that share the same fitness value, allowing populations to drift through sequence space without selection pressure. Neutral networks were first described by Peter Schuster and Walter Fontana in the context of RNA secondary structure: vastly different RNA sequences can fold into the same secondary structure, and these sequences are connected by single-point mutations that preserve structure. The population can wander along this network, exploring distant regions of sequence space that would be inaccessible if every mutation were subject to selection.

The existence of neutral networks solves a fundamental problem in evolvability. Selection alone cannot explain the origin of novel structures because it operates on available variants, and the space of possible variants is astronomically large. Neutral networks provide a mechanism for exploration: populations drift through neutral space, accumulating mutations that are individually invisible to selection but collectively enable access to new phenotypes when the environment changes or when a threshold number of mutations opens a new fitness peak.

The concept extends beyond molecular biology. In neural networks (the artificial kind), the loss landscape contains vast neutral regions where different parameter configurations produce identical performance. This is why stochastic gradient descent can find good solutions: it drifts through neutral space, and the geometry of that space determines which solutions are accessible. The parallel between biological and artificial neutral networks is not metaphorical. Both are instances of a general principle: high-dimensional spaces contain connected manifolds of equivalent function, and traversal of these manifolds is the mechanism of exploration.