Extended Data Fig. 2. Residuals versus fitted values for linear and thermodynamic models fit to AbundancePCA data from combinatorial library 1.

a, Residual fitness (observed − predicted) versus predicted fitness for first-order linear models (left) and first-order thermodynamic models (right) evaluated on GRB2-SH3 combinatorial AbundancePCA data (combinatorial library 1; see Fig. 1). The smoothed conditional mean (generalized additive model) is shown in red. b, Similar to panel a except models include all first-order and second-order genetic interaction (energetic coupling) terms/coefficients. c, Similar to panel a except results are shown for models that were trained on GRB2-SH3 ddPCA data consisting of single and double aa substitutions only²³.