FIG. 2.

(a) The Spearman correlation $\rho$ between LDDT and $\phi \left({\mathit{S}}_n,{\mathit{S}}_p\right)$ plotted versus the information content used $I$ for the ${\mathrm{C}}_{\alpha }$ contact maps (“Contacts”, blue squares), core identify (red circles), secondary structure (“SS”, yellow upward triangles), hydrogen-bonding satisfaction (“H-bond”, green downward triangles), and the 3Di embedding (purple diamonds). Inset: The amount of information $I^{\mathrm{*}}$ needed to obtain $\rho =0.9$ for each physical metric and machine-learning embedding in panel (a). The arrows for secondary structure and hydrogenbond satisfaction indicate they do not reach $\rho =0.9$. (b) $\rho$ plotted versus $I$ when using ESM2 embeddings to predict the contact maps (blue squares) and core identity (red circles) on a log₁₀-scale.