Figure 3

Machine learning framework and predictions. (a) The accuracy of both individual classifiers as well as the ensemble was calculated based on its correct classification of alanine mutations on the A2 and the C2 domains, using a 10 cross-fold validation (“Methods”). The variation in the accuracy values is due to the relatively small input (344 instances). The bars depict mean values and error bars, the standard deviation. (b) The predicted chromogenic activity outputted by the classifiers were only moderately correlated (Pearson’s correlation coefficient, p-value < 0.01). (c) In general, the FVIII mutant constructs with chromogenic activity similar to the wild-type form received high scores from the classifier ensemble. Likewise, low-activity mutants correctly received low scores. The boxplots depict the median (center line), the first and third quartiles (lower- and upper-bounds), and 1.5 times the inter-quartile range (lower- and upper whiskers). Each dot in the plot is an amino acid mutation (i.e., an in vitro alanine mutant construct). Unpaired, two-sided Wilcoxon test (***Indicate p-values < 0.001). (d) Predicted chromogenic activity mapped into the FVIII structure (Supplementary Table S2 lists all predicted values). (e) The relation between the predicted chromogenic activity and the relative surface exposure of the residues of the A1, A3, and C1 domains, indicating that perturbations to the ~ 10%–20% most buried residues (within the 0.1–0.2 range) will likely result in a considerable reduction of the chromogenic activity of the mutant construct. Each dot represents one amino acid from the A1, A3 and C1 domains. Image created using the structure 2R7E (Ref.³) and Chimera 1.14 (Ref.⁶²).