Predicting transmembrane protein topology from 3D structure with SchNet (arXiv:2609.30446v1)
This arXiv preprint (arXiv:2609.30446v1) presents a method that applies the graph neural network SchNet to infer transmembrane protein topology from full-atom 3D structures. The model was trained with 5-fold cross-validation on the same dataset used for DeepTMHMM, uses all-atom embeddings instead of sequence-only or α‑carbon features, requires no pre-trained weights, and reports promising results that GNNs can be effective for topology prediction.