Fig. 3.

Optimization and analysis of graph convolutional neural networks employed in this work across 788 protein targets. a Network contraction observed as a function of hyperparameter optimization. b Mean AUC scores for default and optimized gCNN methods and a random forest (RF) benchmark for chronological test sets comprised of the latest 20% of data labels for each target (“full test”) and further divided into “early” (80–90%) and “late” (90–100%) test subsets. Average deviations (analogous to 95% confidence intervals) for all nine table values are ± 0.005 when assessed across the 788 protein targets