Abstract Transmembrane (TM) proteins are major drug targets, indicated by the high percentage of prescription drugs acting on them. For a rational drug design and an understanding of mutational effects on protein function, structural data at atomic resolution are required. However, hydrophobic TM proteins often resist experimental structure determination and in spite of the increasing number of cryo-EM structures, the available TM folds are still limited in the Protein Data Bank. Recently, the DeepMind’s AlphaFold2 machine learning method greatly expanded the structural coverage of sequences, with high accuracy. Since the employed algorithm did not take specific properties of TM proteins into account, the validity of the generated TM structures should be assessed. Therefore, we investigated the quality of structures at genome scales, at the level of ABC protein superfamily folds, and also in specific individual cases. We tested template-free structure prediction also with a new TM fold, dimer modeling, and stability in molecular dynamics simulations. Our results strongly suggest that AlphaFold2 performs astoundingly well in the case of TM proteins and that its neural network is not overfitted. We conclude that a careful application of its structural models will advance TM protein associated studies at an unexpected level.