Fig. 4: IFUM performance on real-world protein engineering and in silico screening applications.
From: Protein folding stability estimation with explicit consideration of unfolded states
a Scatter plots of experimental melting temperature (Tm) versus IFUM-predicted ΔG (ΔGpred) for IFN-λ, IL-10, and UGT76G1 sequence and backbone redesigns. The model achieved a Pearson Correlation Coefficient (PCC) of 0.75, 0.62, and 0.87, respectively. The marker size corresponds to protein length, and the marker color corresponds to the number of mutations relative to each wild-type protein. Markers with the labels indicate the wild-type proteins and IL-10M1. Rosetta FastRelaxed AF3 model structures were used in these scatter plots. The dashed lines are linear regression lines (measure of center), and the shaded regions indicate the 95% confidence interval (error bands) for each correlation, calculated via bootstrapping. b, c Performance on the in silico screening of designed proteins. b Composition of design datasets for five different protein folds, showing the number of computationally designed proteins that were experimentally found to express in E. coli in a soluble monomeric form (blue) or not (red). c Predictive performance, measured by the Area Under the Receiver Operating Characteristic curve (AUROC), for classifying non-expressing designs from expressing ones. The performance of IFUM’s predicted ΔGpred (blue) is compared with the predicted Local Distance Difference Test (plDDT) from ESMFold (orange) and AF3 (green). Source data are provided as a Source data file.
