Motivation: Learning-based model quality assessment programs have been quite successful at discriminating between high-and low-quality protein structures. Here, we show that it is possible to improve this performance significantly by restricting the learning space to a specific context, in this case membrane proteins. Since these are among the most important structures from a pharmaceutical point-of-view, it is particularly interesting to resolve local model quality for regions corresponding, e. g. to binding sites. Results: Our new ProQM method uses a support vector machine with a combination of general and membrane protein-specific features. For the transmembrane region, ProQM clearly outperforms all methods developed for generic proteins, and it does so while maintaining performance for extra-membrane domains; in this region it is only matched by ProQres. The predictor is shown to accurately predict quality both on the global and local level when applied to GPCR models, and clearly outperforms consensus-based scoring. Finally, the combination of ProQM and the Rosetta low-resolution energy function achieve a 7-fold enrichment in selection of near-native structural models, at very limited computational cost.