Fig. 2

Optimization of the 3DDPD generation strategy. Ten PCM regression tasks with temporal split were trained with each variation of the 3DDPDs to select the optimal parameters. Pairwise differences were analyzed by their statistical significance in a Student’s T test, represented by asterisks in (a,b):. * = p-value < 0.05; ** = p-value < 0.01; *** = p-value < 0.001. a rs3DDPDs were optimized by testing different options for trajectory data (i.e. choices of statistical metrics for sub-trajectory grouped coordinate atomic data: “coordinate” includes all, “rigidity” only SD), number of frames in the sub-trajectory frame splits, number of PCs from the residue PCA, atom selection (i.e. all heavy atoms or “minus C”: non-carbon), and residue selection (i.e. full sequence or class A GPCRdb-annotated binding pocket). b ps3DDPDs were optimized based on trajectory data, variance covered by the selected number of atom PCA components, atom selection, and residue selection. c Residue selection options exemplified on the structure of adenosine A1 receptor PDB 5UEN. In orange, the residues that would be selected by each of the five possible definitions of a structural-driven binding pocket selection approach: full sequence, class A, family, subfamily, and target