There have been many attempts over the years to predict protein secondary structure based purely on sequence. The best accuracy reported in the literature to date to my knowledge is ~90%; it has been argued on theoretical grounds that accuracy much greater than 90% is probably not possible due to some uncertainty in precisely how secondary structure is defined. Most such efforts rely on either convolutional neural networks or multilayer LSTMs. This network is a work in progress (hence, the code is not yet extensively commented so may be hard to read). It's aimed at predicting secondary structure using a somewhat more unusual approach. Updates (and comments) coming soon!