We describe Seder - our deep learning algorithm for protein structure prediction. Protein residue sequences are scored based on their similarity to the native protein structure of the sequence as in the PDB. We applied a recurrent two-hidden-layer feed-forward neural network with momentum and associative memory in our methodology. For optimization we use a hybrid of the Levenberg-Marquardt and back-propagation algorithms. The input features are the partial energy sums between pairs of residues, the distance profile of individual atoms to the solvent, and residue-residue contact maps with a cutoff distance of 5 and 7 Å. The output˚ layer is trained to predict the TM-score to native structure for the given model. We show that in some cases our Seder algorithm outperforms AlphaFold2 (AF2) in protein structure prediction, and we discuss the reason of such underperformance of AF2.