Skip to main content
. Author manuscript; available in PMC: 2022 Dec 13.
Published in final edited form as: J Mol Graph Model. 2021 Dec 21;111:108103. doi: 10.1016/j.jmgm.2021.108103

Fig. 4.

Fig. 4.

APPFD-FK-GMM processing pipeline involving Phase 1 (fitting a GMM to all the keypoints or LSPs descriptor, i.e. local APPFDs from each 3D protein surface and for all database protein surfaces) and Phase 2 (computing a single compact descriptor called fisher-vector (FV) for each 3D protein by aggregating all its keypoints or local APPFDs using the fisher kernel (FK) framework and the trained GMM in Phase 1. Within each LSP around a keypoint, six different geometric features are first extracted, and each feature-dimension is binned into a 1D histogram with 35 bins, where all histograms are combined to form a 210-dimensional descriptor, i.e local APPFD for each LSP. All such LSP descriptors from each 3D protein are compacted into a 4210-dimensional FV for that protein model, as in Phase 2.