Fig. 2. Graphinity model performance for ΔΔG prediction.
a, Correlation of Graphinity predictions with true experimental values for the Experimental_ΔΔG_645 + Reverse Mutations + Non-Binders dataset, with a random train–validation–test split. Reverse mutations were used for training and validation only and were not included in the test dataset. An ensemble of ten models was trained for 500 epochs with 10-fold cross-validation (CV) on the datasets. The trendline, shown in red, is a least-squares polynomial fit. b, The effect of train–validation–test CDR sequence identity cutoffs on Graphinity performance. This figure is included with error bars representing the standard deviation across the ten folds in Supplementary Fig. 2. c, Correlation of Graphinity predictions with true synthetic values for the Synthetic_FoldX_ΔΔG_942723 dataset with a 90% length-matched CDR sequence identity cutoff applied for the train–validation–test split. An ensemble of ten models was trained for ten epochs with 10-fold cross-validation. The trendline, shown in blue, is a least-squares polynomial fit. d, Histograms of the true and predicted FoldX ΔΔG values shown in c (x axis limited to −8 to +5 kcal mol−1 for clarity). The solid lines are kernel density estimates (KDEs). The Pearson’s correlation (r) values are shown in a and c.