Fig. 3

Fitting biophysical models to DMS data assaying mutant effects on multiple phenotypes. a Library design and doubledeepPCA (ddPCA) functional assays (BindingPCA, bPCA and AbundancePCA, aPCA) used to interrogate the effects of all single and a subset of double AA substitutions on the cellular abundance and binding of PSD95-PDZ3 to its cognate ligand (CRIPT) [15, 16]. Green tick mark, yeast growth; red cross, yeast growth defect; DHF, dihydrofolate; THF, tetrahydrofolate. b Identical to panel a except ddPCA was applied to the oncoprotein KRAS to interrogate the effect of mutations on interactions with six different binding partners [15]. c, d Three-state equilibria, thermodynamic models, neural network architectures, and corresponding MoCHI model design tables used to infer the binding and folding free energy changes (∆∆G_f, ∆∆G_b) of the mutant libraries depicted in panels a and b, respectively. Target variable predictions for the three library blocks assaying KRAS-RAF1 bPCA are depicted; the additional 15 (5 × 3) observed phenotypes corresponding to the other 5 binding partners are not shown for simplicity. ∆G_b, Gibbs free energy of binding; ∆G_f, Gibbs free energy of folding; K_b, binding equilibrium constant; K_f, folding equilibrium constant; c, standard reference concentration; p_b, fraction bound; p_f, fraction folded; g_f, nonlinear function of ∆G_f; g_fb nonlinear function of ∆G_f and ∆G_b; R, gas constant; T, temperature in Kelvin. e Nonlinear relationship between observed bPCA fitness and inferred changes in free energies of binding and folding. Thermodynamic model fit shown in red. f Performance of three-state biophysical model predictions of bPCA fitness. R² is the proportion of variance explained. g, h Similar to panels e and f but corresponding to KRAS-RAF1 bPCA fitness for block 1. i Nonlinear relationship between observed aPCA fitness and inferred changes in free energy of binding. j Performance of two-state biophysical model predictions of aPCA fitness. k, l Similar to panels e and f but corresponding to KRAS aPCA fitness for block 1. m, n Comparisons of confident model-inferred free energy changes to previously reported in vitro measurements [49,50,51]. Error bars indicate 95% confidence intervals from a Monte Carlo simulation approach (n = 10 experiments). Pearson’s r is shown