MetaProm: a neural network based meta-predictor for alternative human promoter prediction

Area of Science:

• Genomics
• Bioinformatics
• Gene Regulation

Background:

• De novo eukaryotic promoter prediction is crucial for novel gene discovery and understanding gene regulation.
• Current promoter prediction programs (PPPs) exhibit poor overall performance and lack prediction overlap.
• Most PPPs are trained and tested on upstream promoters, with limited assessment of alternative promoter prediction accuracy.

Purpose of the Study:

• To evaluate the performance of major PPPs on a large-scale human genome dataset, focusing on alternative promoters.
• To develop and assess an artificial neural network (ANN) based meta-predictor integrating multiple PPPs and CpG island information.
• To investigate the association of alternative promoters with CpG islands.

Main Methods:

• Genome-wide performance evaluation of six major PPPs (PSPA, FirstEF, McPromoter, DragonGSF, DragonPF, FProm) using 42,536 human gene promoters.
• Development of an ANN-based meta-predictor combining predictions from individual PPPs and CpG island data.
• Analysis of the relationship between promoter location (3' vs. 5') and CpG island overlap for alternative promoters.

Main Results:

• Individual PPPs show varied strengths and weaknesses across 1.06 x 10^9 base pairs of human genome sequence.
• The developed ANN meta-predictor significantly outperforms any single PPP in both sensitivity and specificity.
• A notable finding is that 74% of 5' alternative promoters overlap with CpG islands, compared to 41% for 3' alternative promoters.

Conclusions:

• The study highlights the limitations of current PPPs, particularly for alternative promoter prediction.
• The novel meta-predictor offers a more accurate and reliable approach to eukaryotic promoter prediction.
• 5' alternative promoters are significantly more likely to be associated with CpG islands, providing insights into their regulatory mechanisms.