A major bottleneck in protein structure prediction is the selection of correct models from a pool of decoys. Relative activities of ∼1,200 individual single-site mutants in a saturation library of the bacterial toxin CcdB were estimated by determining their relative populations using deep sequencing. This phenotypic information was used to define an empirical score for each residue (RankScore), which correlated with the residue depth, and identify active-site residues. Using these correlations, ∼98% of correct models of CcdB (RMSD ≤ 4Å) were identified from a large set of decoys. The model-discrimination methodology was further validated on eleven different monomeric proteins using simulated RankScore values. The methodology is also a rapid, accurate way to obtain relative activities of each mutant in a large pool and derive sequence-structure-function relationships without protein isolation or characterization. It can be applied to any system in which mutational effects can be monitored by a phenotypic readout.