Prediction performance was much higher in constitutive exons than the other categories of alternatively spliced exons. Lack of sequence conservation partly explained the decrease in specificity. The highest specificity levels in the training set were found to occur when the two informant species D. yakuba and D. erecta were available. In the hold out set, 86% of the constitutive exons matched to D. yakuba and D. erecta sequences compared with 75% of the alternatively spliced exons. (In the remaining cases some other combination of one or two informant species were found.) Thus, in some cases ExAlt could not optimally use evidence of sequence conservation to limit false positive predictions. The ab initio versions of ExAlt (ExAlt- Frame-ab initio and ExAlt-Default-ab initio) also got a smaller percentage of cassette exons correct compared to constitutive exons. Many of the cassette exons were less than 100 bases long and the single species ExAlt (ExAlt-Frame-ab initio in Table 3) correctly identified a majority of these exons (62%). However, short exons (less than 100 bases) made up 69% of the cases where single species ExAlt did not get both splice sites exactly correctly. In contrast, all but 1 of the short constitutive exons were correctly identified.