Abstract
Abstract
Protein behavior, whether evolved in nature or intentionally designed, is governed by the energetics of amino acid interactions that bridge sequence to function. As an extension of the theory of protein folding funnels, local frustration quantifies the optimality of these residue-level interactions relative to alternatives. Although this framework has existed for decades, experimental validation has remained elusive. Past deep mutational scanning datasets enable experimental assessment across nearly 8,000 positions from 178 proteins. Results provide experimental evidence for theoretical predictions regarding the relationship between local frustration and protein sequence, structure, and evolution. An evaluation of three in silico methods shows modest agreement with benchmarks (r = 0.03 to 0.26), with the predominant local frustration predictor (Protein Frustratometer) capturing under 3% of the experimental variance. A recent deep learning model, Pythia, outperformed other tools across all evaluated metrics (r = 0.42). Finally, this scale of empirical data enables an evaluation of the strengths and limitations of energetic measures themselves, inspiring a complementary Sequence Probability INverse (SPIN) framework, which characterizes optimality through a Boltzmann-weighted selection probability within an ensemble of sequences. These findings help provide experimental grounding for the theoretical principles that govern protein sequence energetics.