SHREC 2021: Retrieval and classification of protein surfaces equipped with physical and chemical properties

Surface-based protein domains retrieval methods from a SHREC2021 challenge

“…In this deep learning method, our group exploits the availability of protein class labels from Ref. [35] to optimize the representation of protein surfaces without any additional properties. Particularly, we designed a message-passing graph convolutional neural network (MPGCNN) with the Edge Convolution (EdgeConv) paradigm [57] for the protein classification objective.…”

“…The main novelty of the SHREC ′ 21 track is arguably the availability of protein surfaces with electrostatic values, which has been shown to improve the retrieval performance of protein surfaces [11,35]. This additional feature might therefore allow to better distinguish structurally related proteins based on their surficial properties and improve the methods' performance.…”

“…As instance, in the SCOPe databases, the SH3 domain and the SH3_2 domain are classified in two different SCOPe domains, similarly to the Pfam classification. The DGCNN used the data from another SHREC ′ 21 track, the Retrieval and classification of protein surfaces equipped with physical & chemical properties track [35]. The organizers of this track, similarly to a previous track [33], proposed a set of shapes derived from NMR structures along their surficial physico-chemical properties to allow the participants to train their methods, and the resulting classification of the proteins was derived from the SCOPe database as well.…”

Real-Time Structure Search and Structure Classification for AlphaFold Protein Models

“…We used ADAM for parameter optimization with a binary cross entropy loss function. The learning rate was explored from 1e-3 to 7e-3 and 0.1-0.7 in our previous work and set to 0.005 15 . The accuracy of networks was evaluated on the negative and positive set generated from the 2,541 structures, which totals 167,872 pairs.…”

“…For this classification, we considered 1,101 folds in the class a (all  proteins), b (all  proteins), c ( proteins), d ( proteins), and g (small proteins) in the SCOPe database. The neural network takes 3DZDs of two protein structures and outputs the probability that the two structures belong to the same SCOPe fold 15 (Supplementary Figure 1; see Methods). We trained two networks, one that uses 3DZDs computed from full-atom protein surface and another one that takes 3DZDs computed from main-chain C, C, and N atoms 16 .…”