cs.AISep 24, 2026
SavePredicting Transmembrane Protein Topology from 3D Structure
Organizations: Technical University of Denmark
Abstract
This paper presents a novel approach to infer protein topology using the state-of-the-art graph neural network (GNN), SchNet. The model is trained on the same dataset used to develop the recent DeepTMHMM model with 5-fold cross-validation. Unlike the conventional approaches based on using only the protein sequences or the -carbons as features, we have decoded our classifier in this way, so all atom-level embeddings are used. Without applying any pre-trained weight, the final results have shown great potential that GNNs can be used for topological predictions.
Figures & tables
Fig. 1 : Simple illustration explaining protein topologies. A protein can have different topologies depending on where it is. Please notice that the barrel protein shown in the middle has more complicated topologies as its bottom part frequently changes location. On the other hand, helices have relatively simpler topologies but they might have signal peptide bound to them, which needs to be distinguished. Illustration inspired by [ 1 ] .
Fig. 2 : Methodology throughout the project. All blocks are numbered for later reference. The 3D protein structure in block 2 is from AlphaDB 3D viewer.
| Protein type | alpha TM | alpha SP+TM | beta barrel | Globular | SP+Globular |
|---|---|---|---|---|---|
| Amount | 383 (387) | 102 (106) | 82 (82) | 1,982 (2,000) | 994 (1,000) |
Table 1: Number and name of each protein type used in the DeepTMHMM study. TM : Transmembrane, SP : Signal Peptide. The numbers without parentheses indicate sequences with both 1D sequences and 3D structures available while the numbers within parentheses indicate proteins with only 1D sequences.
Fig. 3 : Random graph illustrating the basic message passing of GNN. Notice that the AGGREGATE operation is part of the UPDATE function (indicated by a sum) and depends on the number of neighbors. The graph embedding is then a sum of all the node embeddings.
Fig. 4 : Major voting approach used in this study. The final topological prediction is based on the majority of the atom-level topological predictions.
Fig. 5 : Class distribution in the first fold of the 5-fold cross-validation. This distribution indicates that most protein parts intracellular proteins. The distributions in the other CV folds are very similar (not shown).
Fig. 6 : a): Loss curves for the 5 SchNet models trained and validated on each of the five CV folds. b): Accuracy curves for the 5 SchNet models trained and validated on each of the five CV folds. The abrupt jump around epoch 100 is caused by the exponential decaying learning scheduler. Please notice that the training for CV setup 5 was early-stopped around epoch 90.
Table 2: Topological and residue-level evaluation performed on the protein types alpha TM , SP+TM , beta barrel , Globular and SP+Globular presented in Table 2 . Please notice that the overlap criterion is set to be 5 by default for the topological evaluation of Globular and SP+Globular as these proteins contain both -helices and -sheets [ 1 ] .
Table 3: McNemars test and the corresponding 95 % confidence intervals (CIs) between the baseline and the trained SchNet model for each CV setup. The trained SchNet is always superior regardless of which setup.
Fig. 7 : SchNet trained and validated on CV setup 1 using -carbon approach with Gaussian smoothing. The hyper-parameters are fine-tuned and an exponential learning schedule is applied.
Fig. 8 : Training and validation curves of GCPNet and EGNN trained and validated on CV setup 1 with the proposed major voting method. Both models seem to have failed to learn the representation from the CV setup 1 dataset. Gaussian smoothing is applied to the models.