Fig. 4

Benchmark of 3DDPD performance in PCM bioactivity modelling tasks against non-dynamic descriptors. Ten RF models with random seeds were trained and validated for each combination of protein descriptors with ECFP6 molecular fingerprints. A shade of green (the darker the better) represents better performance using a descriptor A instead of a descriptor B, as read in panel a. A shade of red (the darker the worse) represents worse performance using a descriptor A instead of a descriptor B. The statistical significance of the differences is derived from pairwise Student T-test and represented by asterisks: * = p-value < 0.05; ** = p-value < 0.01; *** = p-value < 0.001. Four PCM tasks were benchmarked: a Classification with validation based on an 80:20 random split. In classification tasks, MCC was used as an evaluation metric on the test set. b Regression with validation based on 80:20 random split. In regression tasks, Pearson r was used as an evaluation metric on the test set. c Classification with validation based on a temporal split, with 2013 as the cutoff year. d Regression with validation based on a temporal split, with 2013 as the cutoff year