Removing ~1% of the bonds in a random elastic network produces protein-like allosteric response

Rocks, Pashine, Bischofberger, Goodrich, Liu, and Nagel (PNAS 2017; arXiv:1607.08562) showed that a generic disordered spring network can be tuned into a long-range mechanical machine by deleting a tiny fraction of its bonds. "With nearly complete success, we are able to produce a strain between any pair of target nodes... by removing only ∼1% of the bonds," and the resulting behavior "is reminiscent of the long-range coupled conformational changes that often occur during allostery in proteins." A local perturbation applied at one site produces a designed response at a distant site — the defining signature of allostery — in a substrate with no biology in it.

The cheapness is the load-bearing point. Because long-range response turns out to be generic and easy to engineer in disordered networks, the authors offer the result as a clue to biology: it "may give insight into why allostery is a common means for the regulation of activity in biological molecules." The reading is that evolution reuses allostery not because it is special or hard to build, but because the soft-mode geometry that supports it is abundant and reachable by small changes — the trained artifact and the evolved protein occupy the same mechanical regime. (This is an interpretive extension the paper hedges as "may give insight," not a demonstrated evolutionary claim.)

This is the empirical anchor for the cross-domain arc of the capture that seeded this note. Node allostery is exactly the behavior a physical network can learn in situ by a local rule (claim-coupled-learning-elastic-networks-compute-without-a-processor), and the structure→function reading of such trained networks ties their task-critical edges to slow collective modes (claim-physical-networks-become-what-they-learn-soft-modes). Together they frame a protein as a physical learning machine on the same energy-based footing as the Boltzmann / Forward-Forward lineage (claim-hinton-forward-forward-boltzmann-lineage), and connect to backpropagation-gap's question of whether shared structure implies shared process.