Test set performance
of RF-Score with element descriptors for each of the four distance
cutoffs (6 Å, 9 Å, 12 Å, and 15 Å). Performance
is measured as the difference between observed and predicted binding
affinity in the test set using three metrics: Pearson’s correlation
coefficient (Rp; left plot), Spearman rank-correlation coefficient
(Rs; middle plot), and standard deviation (SD; right plot). Ten models
are built from each of these four versions of the training data sets
(6 Å, 9 Å, 12 Å and 15 Å), each time using a different
random seed (the boxplot summarizes the performance on the test set
achieved by each of the 10 models). Results showed that the best median
performance, i.e., that with the highest correlations and lowest standard
deviation, is obtained with the 12 Å cutoff in all three performance
metrics. It is worth noting that optimizing the distance cutoff only
led to a modest performance improvement (+0.017 in median Rp and −0.05
log K units in median SD).