Fig. 2: In silico evaluation on SKEMPI and the HER2 binders test set.

Comparative analysis of Per-PDB Spearman (a) and Pearson (b) correlations between predictions of various models and experimental data across SKEMPI subsets with varying difficulty levels. PDB codes in SKEMPI are categorized into “easy" (50+ similar data points in the training set), “medium" (1–50), and “hard" (0) targets based on the number of training data points having a high structural similarity (TM-score  >0.8) to it. The number of PDB codes for each difficulty is annotated in the figure legends. The box spans the inter-quartile range (25th to 75th percentile), with a solid line inside marking the median. Outliers are determined by 1× inter-quartile range. c Benchmark results on the HER2 binders test set (n = 419) show Pearson and Spearman correlations for various models. The deep-learning models are trained on SKEMPI. d Change of performance metrics in HER2 binders test set when excluding various models from the FoldX + Bind-ddG + GearBind(+P) ensemble. e Change of performance when changing GearBind architecture design, as quantified by cross-validation performance on SKEMPI (n = 5729). f Change of performance on SKEMPI (n = 5729) when excluding different models from the FoldX + Flex-ddG + Bind-ddG + GearBind(+P) ensemble.