Dinucleotide controlled null models for comparative RNA gene prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Comparative RNA structure prediction aids in identifying functional noncoding RNAs.
• Existing RNA gene prediction programs may exhibit bias due to genomic dinucleotide content, particularly those using thermodynamic folding models.
• There is a need for dinucleotide-preserving control strategies to accurately assess RNA gene predictions, especially for multiple alignments where algorithms are lacking.

Purpose of the Study:

• To develop and present SISSIz, a novel program for simulating multiple alignments with controlled average dinucleotide content.
• To provide a robust null model for assessing RNA gene prediction significance.
• To create a dinucleotide-content-unbiased RNA gene finding program.

Main Methods:

• SISSIz simulates multiple alignments by preserving dinucleotide content, sequence diversity, local conservation, and gap patterns.
• A phylogenetic substitution model incorporating overlapping dependencies and site-specific rates is utilized.
• A distance-based approach with fast heuristics estimates a tree to guide the simulation process.

Main Results:

• The SISSIz program effectively simulates multiple alignments with controlled dinucleotide content.
• Testing on vertebrate genomic alignments demonstrated the algorithm's utility in studying RNA structure predictions.
• Integrating SISSIz with the RNAalifold algorithm resulted in a new RNA gene finding program unbiased by dinucleotide content.

Conclusions:

• SISSIz provides an efficient algorithm for randomizing multiple alignments while preserving dinucleotide content.
• The program enables more accurate estimation of false positive rates for existing prediction tools.
• SISSIz can be used for negative controls in machine learning training or as a standalone RNA gene finder.