An artificial neural network system is used for pattern recognition in
protein side-chain-side-chain contact maps, A back-propagation networ
k was trained on a set of patterns which are popular in side-chain con
tact maps of protein structures, Several neural network architectures
and different training parameters were tested to decide on the best co
mbination for the neural network, The resulting network can distinguis
h between original (from protein structures) and randomized patterns w
ith an accuracy of 84.5% and a Matthews' coefficient of 0.72 for the t
esting set. Applications of this system for protein structure evaluati
on and refinement are also proposed. Examples include structures obtai
ned after the application of molecular dynamics to crystal structures,
structures obtained from X-ray crystallography at various stages of r
efinement, structures obtained from a de novo folding algorithm and de
liberately misfolded structures.