Expectation-Maximization (EM) Algorithms for Mapping Short Reads Illustrated with FAIRE data and the TP53-WRAP53 Gene Region

TL;DR

This paper introduces an EM algorithm that probabilistically maps short sequencing reads to genomes, improving accuracy by considering error rates and multiple mapping locations, demonstrated with FAIRE data on TP53-WRAP53.

Contribution

It presents a novel EM-based method for mapping short reads that accounts for sequencing errors and multiple alignments, enhancing mapping accuracy over existing approaches.

Findings

Improved read allocation accuracy using EM algorithm.

Effective handling of multi-mapped reads with probabilistic assignment.

Application to FAIRE data reveals detailed genomic insights.

Abstract

Huge numbers of short reads are being generated for mapping back to the genome to discover the frequency of transcripts, miRNAs, DNAase hypersensitive sites, FAIRE regions, nucleosome occupancy, etc. Since these reads are typically short (e.g., 36 base pairs) and since many eukaryotic genomes, including humans, have highly repetitive sequences then many of these reads map to two or more locations in the genome. Current mapping of these reads, grading them according to 0, 1 or 2 mismatches wastes a great deal of information. These short sequences are typically mapped with no account of the accuracy of the sequence, even in company software when per base error rates are being reported by another part of the machine. Further, multiply mapping locations are frequently discarded altogether or allocated with no regard to where other reads are accumulating. Here we show how to combine probabilistic mapping of reads with an EM algorithm to iteratively improve the empirical likelihood of the allocation of short reads. Mapping using LAST takes into account the per base accuracy of the read, plus insertions and deletions, plus anticipated occasional errors or SNPs with respect to the parent genome. The probabilistic EM algorithm iteratively allocates reads based on the proportion of reads mapping within windows on the previous cycle, along with any prior information on where the read best maps. The methods are illustrated with FAIRE ENCODE data looking at the very important head-to-head gene combination of TP53 and WRAP 53.