Some promising initial alignments are obtained by applying projection methods or random starts on the entire dataset. Typically, random starts are used because they are cost efficient. The most promising sets of alignments are considered for further processing. These initial alignments are then converted into profile representation.