High-throughput experimental techniques have made possible the systematic sampling of the single mutation landscape for many proteins, defined as the change in protein fitness as the result of point mutation sequence changes. In a more limited number of cases, and for small proteins only, we also have nearly full coverage of all possible double mutants. By comparing the phenotypic effect of two simultaneous mutations with that of the individual amino acid changes, we can evaluate epistatic effects that reflect non-additive cooperative processes. The observation that epistatic residue pairs often are in contact in the 3D structure led to the hypothesis that a systematic epistatic screen contains sufficient information to identify the 3D fold of a protein.To test this hypothesis, we examined experimental double mutants for evidence of epistasis and identified residue contacts at 86% accuracy, including secondary structure elements and evidence for an alternative all--helical conformation. Positively epistatic contactscorresponding to compensatory mutations, restoring fitnesswere the most informative. Folded models generated from top-ranked epistatic pairs, when compared with the known structure, were accurate within 2.4 Å over 53 residues, indicating the possibility that 3D protein folds can be determined experimentally with good accuracy from functional assays of mutant libraries, at least for small proteins. These results suggest a new experimental approach for determining protein structure.

If disentangled properly, patterns distilled from evolutionarily related sequences of a given protein family can inform their traits - such as their structure and function. Recent years have seen an increase in the complexity of generative models towards capturing these patterns; from sitewise to pairwise to deep and variational. In this study we evaluate the degree of structure and fitness patterns learned by a suite of progressively complex models. We introduce pairwise saliency, a novel method for evaluating the degree of captured structural information. We also quantify the fitness information learned by these models by using them to predict the fitness of mutant sequences and then correlate these predictions against their measured fitness values. We observe that models that inform structure do not necessarily inform fitness and vice versa, contrasting recent claims in this field. Our work highlights a dearth of consistency across fitness assays as well as divergently provides a general approach for understanding the pairwise decomposable relations learned by a given generative sequence model.