Fig. 1.

3D dynamic descriptor (3DDPD) generation overview. First, a selection of residues and atoms is made. XYZ coordinates are collected for the selected atoms over all frames of the trajectory. The full simulation ranging from 0 to 500 ns is divided into sub-trajectories and atomic coordinate statistics (average, SD, and median) are computed for each of them. Two routes are possible from this point to generate either one-hot encoded residue-specific rs3DDPDs or embedding-like protein-specific ps3DDPDs. Respectively, atomic data is grouped and standardized either per target or for all targets and PCA is computed. A number of PCs for each atom are then selected and, in the case of rs3DDPDs, grouped per residue by calculating the average and SD. A second dimensionality reduction step is applied to residue data and the selected n number of PCs are mapped to their corresponding positions in a MSA of the targets of interest. This results in a vector rs3DDPD of length n * L, where L is the length of the protein or the MSA. For ps3DDPDs, the m number of PCs selected per atom are grouped per target by calculating average, median, and SD, therefore resulting in the final vector descriptor of length m * 3