Recalibration of mapping quality scores in Illumina short-read alignments improves SNP detection results in low-coverage sequencing data

Background Low-coverage sequencing is a cost-effective way to obtain reads spanning an entire genome. However, read depth at each locus is low, making sequencing error difficult to separate from actual variation. Prior to variant calling, sequencer reads are aligned to a reference genome, with alignments stored in Sequence Alignment/Map (SAM) files. Each alignment has a mapping quality (MAPQ) score indicating the probability a read is incorrectly aligned. This study investigated the recalibration of probability estimates used to compute MAPQ scores for improving variant calling performance in single-sample, low-coverage settings. Materials and methods Simulated tomato, hot pepper and rice genomes were implanted with known variants. From these, simulated paired-end reads were generated at low coverage and aligned to the original reference genomes. Features extracted from the SAM formatted alignment files for tomato were used to train machine learning models to detect incorrectly aligned reads and output estimates of the probability of misalignment for each read in all three data sets. MAPQ scores were then re-computed from these estimates. Next, the SAM files were updated with new MAPQ scores. Finally, Variant calling was performed on the original and recalibrated alignments and the results compared. Results Incorrectly aligned reads comprised only 0.16% of the reads in the training set. This severe class imbalance required special consideration for model training. The F1 score for detecting misaligned reads ranged from 0.76 to 0.82. The best performing model was used to compute new MAPQ scores. Single Nucleotide Polymorphism (SNP) detection was improved after mapping score recalibration. In rice, recall for called SNPs increased by 5.2%, while for tomato and pepper it increased by 3.1% and 1.5%, respectively. For all three data sets the precision of SNP calls ranged from 0.91 to 0.95, and was largely unchanged both before and after mapping score recalibration. Conclusion Recalibrating MAPQ scores delivers modest improvements in single-sample variant calling results. Some variant callers operate on multiple samples simultaneously. They exploit every sample's reads to compensate for the low read-depth of individual samples. This improves polymorphism detection and genotype inference. It may be that small improvements in single-sample settings translate to larger gains in a multi-sample experiment. A study to investigate this is ongoing.