SchNet Uses 3D Protein Structure to Predict Transmembrane Topology

Summary

This paper presents a method for inferring transmembrane protein topology from three-dimensional structure with SchNet, a graph neural network. The model is trained on the same dataset used to develop DeepTMHMM and evaluated with five-fold cross-validation. Unlike approaches based only on protein sequences or alpha-carbon features, the classifier uses embeddings for all atoms in the structure. The authors report that the approach shows potential for topology prediction without using pretrained weights, although the abstract does not provide detailed performance figures.