A new graph neural network predicts protein topology straight from 3D atomic structure, skipping the sequence data most tools rely on.
Researchers trained SchNet, a graph neural network architecture, on the same dataset used to build DeepTMHMM, the current standard tool for mapping how proteins sit inside cell membranes. Instead of feeding the model a flat amino acid sequence or just the alpha-carbon backbone, as most existing approaches do, the team gave it embeddings for every atom in the 3D structure. They validated the approach with 5-fold cross-validation and used no pretrained weights, meaning the network learned everything from scratch on this one dataset. The paper describes the results as showing 'great potential' for graph neural networks in topology prediction.
That matters because topology, which parts of a protein poke outside the cell, sit in the membrane, or dangle inside, shapes how drugs can target it and how cells communicate. Sequence-based tools like DeepTMHMM have dominated this work for years, partly because reliable 3D structures used to be scarce. With protein structure prediction tools now making atomic-level models far more common, feeding that 3D detail directly into a network is a logical next experiment.
Logical, but unproven: the paper reports no head-to-head accuracy comparison against DeepTMHMM or any other tool, so 'great potential' is, for now, just one team's read on a first attempt.