A brief introduction to HNN
The HNN (Hierarchical Neural Network ) prediction method can be seen as an improvement on the famous classifier developed by Qian and
Sejnowski, and derived from the system NETtalk. As its predecessor, it is made up of two networks: a sequence-to-structure network and
a structure-to-structure network. The prediction is thus only based on local information. The improvements mainly deal with two points:
technical tricks (recurrent connections, shared weights...) have been used to increase the context on which the prediction is made
and concomitantly decrease by two orders of magnitude the number of parameters (weights),
physico-chemical data have been explicitly incorporated in the predictors used by the structure-to-structure network.
These modifications have significantly improved the error in generalization. This error could still be greatly lowered, would the
profile of multiple-alignment be used as input of the whole system, instead of the single primary structure of the sequence of
interest.
NPS@ is the orginal server for this method.