We have developed a method called DMPfold that predicts protein structures without templates using deep learning on sequence covariation data. It was applied to proteins in Pfam and was able to model 25% of families without templates at high predicted accuracy.
DMPfold can be downloaded and run locally under an open source license, and can also be run via the PSIPRED web server. It takes around 3 hours on a single core to predict the structure for a 200 residue protein.
Read our paper in Nature Communications for more on how DMPfold makes recent advances in protein structure prediction available to the community.