FIG. 3.

The Spearman correlation $\rho$ between LDDT and $\phi$ plotted against the fraction of flipped core identity labels compared to the native labels. Errors were introduced randomly starting from the true identities (black circles and solid line). Error bars indicate ± one standard deviation from the average over random samples. Core identity was also predicted from sequence using a series of models: the ESM2 model developed here (red x), Netsurfp3.0 (blue upward triangle), PaleAle6 (purple downward triangle), RaptorX (orange square), and E-pRSA (green diamond). The grey dotted line marks $\rho =0.9$, our cutoff for a high quality predictor. Inset: The rate $R$ of correct core identity labels plotted versus the entropy of the core identity labels $H$ colored by hydrophobicity $h_i$ (increasing from blue to yellow).