This work introduces a novel approach for predicting transmembrane protein topology by leveraging the state‑of‑the‑art graph neural network SchNet. The model is trained on the same dataset used for the recent DeepTMHMM study, employing 5‑fold cross‑validation. Unlike conventional methods that rely solely on protein sequences or $\alpha$‑carbon atoms as features, our classifier incorporates embeddings for all atoms at the atomic level. Training is performed without any pre‑trained weights, and the results demonstrate that GNNs hold strong promise for topology prediction.
Review