Gene finding: putting the parts together

Any isolated signal of a gene is hard to predict. Current methods for promoter prediction, for instance, will have either a very low specificity or a very bad sensitivity, such that they will either predict a huge number of false positives (fake promoters) or a very small number of true promoters. The same is essentially true for splice site prediction: if looked at in isolation, splice sites are very hard to recognize with good accuracy. This may seem like a contradiction, because there are programs that perform well on this task such as those by Brunak, Engelbrecht, & Knudsen (1991) and Solovyev, Salamov, & Lawrence (1994). The reason for the success is that both of these methods also use the statistics of the coding exon region next to the splice site. Apart from doing a very careful job of describing the regions right around the splice site, they can therefore also rule out splice sites which do not sit next to something looking like a good coding region. In bird-watching the surroundings often gives the necessary clues in deciding which bird you are watching in the distance, whether it is seen in an open field or in a wood for instance. Signal detection in genes is much like bird-watching: it is necessary to take the surroundings into account. Therefore, to predict something like a splice site, you also need to predict coding exons and vice versa (disregarding the splice sites of introns in untranslated regions). In a long DNA sequence, you probably would not expect to see a coding exon with two associated splice sites unless there are other exons with which it can combine. In this way predictions of the various parts of a gene should influence each other, and prediction of the entire gene structure will also improve on the predictions of the individual signals. Therefore, in the last few years, gene prediction has moved more and more towards prediction of whole gene structures, and these methods typically use modules for recognition of coding regions, splice sites, translation initiation and termination sites, and some even use statistics of the 5’ and 3’ untranslated regions (UTRs), promoters, etc.. This combination of predictions has indeed improved the accuracy of gene prediction considerably, and as more knowledge is gained about transcription and translation, it is likely that the integration of other signals can improve it even further.