The energy landscape of a protein – the high-dimensional surface governing how it folds, misfolds, and fluctuates – has long been conjectured to possess ultrametric structure, where the distance between any two configurations satisfies a stronger-than-triangular inequality. Bikulov and Zubarev construct a minimal model of a 128-residue heteropolymer using point masses with repulsive, Lennard-Jones, screened Coulomb, and elastic bond potentials, then test for ultrametricity directly using replica overlap defined through Pearson correlation coefficients.

Their GPU simulations across multiple random sequences reveal that 90% of sequences contain predominantly ultrametric distance matrices, and of those, 97.8% display nontrivial ultrametricity – meaning the hierarchical structure is not a trivial consequence of widely separated energy scales but reflects genuine multi-level organization. The method borrows from spin glass theory but dispenses with disorder averaging, working instead with individual sequence realizations. This is important because real proteins are single sequences, not ensembles, and the relevant question is whether a specific sequence’s landscape is ultrametric rather than whether the average over random sequences is.

The result provides computational evidence for Frauenfelder’s hypothesis that protein energy landscapes are organized as taxonomic trees – hierarchies of basins within basins – and connects protein physics to the mathematics of p-adic spaces. If the landscape is genuinely ultrametric, then the dynamics of conformational transitions can be described by diffusion on a tree, and the slow relaxation observed in proteins at low temperatures is a natural consequence of the hierarchical barrier structure rather than of any particular chemical detail.

(arXiv:2603.13012)