Gene prediction: compare and CONTRAST

Abstract

CONTRAST, a new gene-prediction algorithm that uses sophisticated machine-learning techniques, has pushed de novo prediction accuracy to new heights, and has significantly closed the gap between de novo and evidence-based methods for human genome annotation.

Figure 1

Increase in the accuracy of de novo gene prediction over time. The gene sensitivity and specificity and the exon sensitivity and specificity on the EGASP test set [5] are shown for several programs by year of initial publication. Included are GENSCAN (1997), TWINSCAN (2001), N-SCAN (2005) and CONTRAST (2007). Note the significant decrease in false positive predictions (as measured by the rise in TWINSCAN's exon specificity) with the inital use of evolutionarily related genome sequences. By comparison, the accuracy of the Ensembl evidence-based gene predictions used in the EGASP experiment at the gene level were 71.6% sensitivity and 67.3% specificity and 77.5% sensitivity and 82.7% specificity at the exon level.