Since the first efforts to predict protein secondary structures in the early 1970s, protein secondary structure prediction has made great strides. This is a result not only of vast increses in the number of protein structures which can be used for predictions, but also of a number of sophisticated methodologies being brought to bear on the problem. I will attempt to introduce and evaluate several of these new methods: statistical methods based on information theory, nearest neighbor methods, Markov models, and neural nets. Similarities between the methods will be discussed, as well as similarities in the approaches which have yielded the most success.

Of course, a central question in secondary structure prediction, as in any scientific endeavor, is "What does the researcher hope to gain from work in this field?" The answer most commonly given by secondary structure predictors is that secondary structure prediction will hopefully be a useful bridge from known primary structure to the Holy Grail of a priori tertiary structure prediction. The usefulness of secondary structure assignments in performing sequence alignments is also pointed to, as is preported usefulness in drug design. A more cynical answer might be that secondary structure prediction is popular because it is easier than tertiary structure prediction, but that unfortunately ease is inversely proportional to relevance in this case. This question of relevance will be addressed in more detail at the end of this paper.

Since this paper is rather broad in scope, the reader may find that one or more of the methods dealt with below are too briefly treated. I refer you to the review by Barton (Barton, 1995) and the excellent review by Bohm (Bohm, 1995), as well as the specific papers used in this work.