Prediction of non-canonical polyadenylation signals in human genomic sequences based on a novel algorithm using a fuzzy membership function.

Computational prediction of polyadenylation signals (PASes) is essential for analysis of alternative polyadenylation that plays crucial roles in gene regulations by generating heterogeneity of 3'-UTR of mRNAs. To date, several algorithms that are mostly based on machine learning methods have been developed to predict PASes. Accuracies of predictions by those algorithms have improved significantly for the last decade. However, they are designed primarily for prediction of the most canonical AAUAAA and its common variant AUUAAA whereas other variants have been ignored in their predictions despite recent studies indicating that non-canonical variants of AAUAAA are more important in the polyadenylation process than commonly recognized. Here we present a new algorithm "PolyF" employing fuzzy logic to confer an advance in computational PAS prediction--enable prediction of the non-canonical variants, and improve the accuracies for the canonical A(A/U)UAAA prediction. PolyF is a simple computational algorithm that is composed of membership functions defining sequence features of downstream sequence element (DSE) and upstream sequence element (USE), together with an inference engine. As a result, PolyF successfully identified the 10 single-nucleotide variants with approximately the same or higher accuracies compared to those for A(A/U)UAAA. PolyF also achieved higher accuracies for A(A/U)UAAA prediction than those by commonly known PAS finder programs, Polyadq and Erpin. Incorporating the USE into the PolyF algorithm was found to enhance prediction accuracies for all the 12 PAS hexamers compared to those using only the DSE, suggesting an important contribution of the USE in the polyadenylation process.