Base Pairing

Table 1 illustrates the wide range of operations that BEDTools support. Many of the tools have extensive parameters that allow user-defined overlap criteria and fine control over how results are reported. Importantly, we have also defined a concise format (BEDPE) to facilitate comparisons of discontinuous features (e.g. paired-end sequence reads) to each other (pairToPair), and to genomic features in traditional BED format (pairToBed). This functionality is crucial for interpreting genomic rearrangements detected by paired-end mapping, and for identifying fusion genes or alternative splicing patterns by RNA-seq. To facilitate comparisons with data produced by current DNA sequencing technologies, intersectBed and pairToBed compute overlaps between sequence alignments in BAM format (Li et al., 2009 (link)), and a general purpose tool is provided to convert BAM alignments to BED format, thus facilitating the use of BAM alignments with all other BEDTools (Table 1). The following examples illustrate the use of intersectBed to isolate single nucleotide polymorphisms (SNPs) that overlap with genes, pairToBed to create a BAM file containing only those alignments that overlap with exons and intersectBed coupled with samtools to create a SAM file of alignments that do not intersect (-v) with repeats.
Table 1.

Summary of supported operations available in the BEDTools suite

Utility	Description
intersectBed*	Returns overlaps between two BED files.
pairToBed	Returns overlaps between a BEDPE file and a BED file.
bamToBed	Converts BAM alignments to BED or BEDPE format.
pairToPair	Returns overlaps between two BEDPE files.
windowBed	Returns overlaps between two BED files within a user-defined window.
closestBed	Returns the closest feature to each entry in a BED file.
subtractBed*	Removes the portion of an interval that is overlapped by another feature.
mergeBed*	Merges overlapping features into a single feature.
coverageBed*	Summarizes the depth and breadth of coverage of features in one BED file relative to another.
genomeCoverageBed	Histogram or a ‘per base’ report of genome coverage.
fastaFromBed	Creates FASTA sequences from BED intervals.
maskFastaFromBed	Masks a FASTA file based upon BED coordinates.
shuffleBed	Permutes the locations of features within a genome.
slopBed	Adjusts features by a requested number of base pairs.
sortBed	Sorts BED files in useful ways.
linksBed	Creates HTML links from a BED file.
complementBed*	Returns intervals not spanned by features in a BED file.

Utilities in bold support sequence alignments in BAM. Utilities with an asterisk were compared with Galaxy and found to yield identical results.

Other notable tools include coverageBed, which calculates the depth and breadth of genomic coverage of one feature set (e.g. mapped sequence reads) relative to another; shuffleBed, which permutes the genomic positions of BED features to allow calculations of statistical enrichment; mergeBed, which combines overlapping features; and utilities that search for nearby yet non-overlapping features (closestBed and windowBed). BEDTools also includes utilities for extracting and masking FASTA sequences (Pearson and Lipman, 1988 (link)) based upon BED intervals. Tools with similar functionality to those provided by Galaxy were directly compared for correctness using the ‘knownGene’ and ‘RepeatMasker’ tracks from the hg19 build of the human genome. The results from all analogous tools were found to be identical (Table 1).

Some sequences, or even entire reads, can be overrepresented in FASTQ data. Analysis of these overrepresented sequences provides an overview of certain sequencing artifacts such as PCR over-duplication, polyG tails and adapter contamination. FASTQC offers an overrepresented sequence analysis module, however, according to the author’s introduction, FASTQC only tracks the first 1 M reads of the input file to conserve memory. We suggest that inferring the overall distribution from the first 1 M reads is not a reliable solution as the initial reads in Illumina FASTQ data usually originate from the edges of flowcell lanes, which may have lower quality and different patterns than the overall distribution.
Unlike FASTQC, fastp samples all reads evenly to evaluate overrepresented sequences and eliminate partial distribution bias. To achieve an efficient implementation of this feature, we designed a two-step method. In the first step, fastp completely analyzes the first 1.5 M base pairs of the input FASTQ to obtain a list of sequences with relatively high occurrence frequency in different sizes. In the second step, fastp samples the entire file and counts the occurrence of each sequence. Finally, the sequences with high occurrence frequency are reported.
Besides the occurrence frequency, fastp also records the positions of overrepresented sequences. This information is quite useful for diagnosing sequence quality issues. Some sequences tend to appear in the read head whereas others appear more often in the read tail. The distribution of overrepresented sequences is visualized in the HTML report. Figure 5 shows a demonstration of overrepresented sequence analysis results.

Genomic analyses often seek to compare features that are discovered in an experiment to known annotations for the same species. When genomic features from two distinct sets share at least one base pair in common, they are defined as ‘intersecting’ or ‘overlapping’. For example, a typical question might be ‘Which of my novel genetic variants overlap with exons?’ One straightforward approach to identify overlapping features is to iterate through each feature in set A and repeatedly ask if it overlaps with any of the features in set B. While effective, this approach is unreasonably slow when screening for overlaps between, for example, millions of DNA sequence alignments and the RepeatMasker (Smit et al., 1996–2004 ) track for the human genome. This inefficiency is compounded when asking more complicated questions involving many disparate sets of genomic features. BEDTools was developed to efficiently address such questions without requiring an installation of the UCSC or Galaxy browsers. The BEDTools suite is designed for use in a UNIX environment and works seamlessly with existing UNIX utilities (e.g. grep, awk, sort, etc.), thereby allowing complex experiments to be conducted with a single UNIX pipeline.

Output files from Illumina MiSeq were first run through FastQC (Andrews et al. 2018 (link)) to check read quality. The paired-end reads were merged using PEAR (Stamatakis et al. 2014 (link)) set to a minimum assembly length of 150 base pairs reads allowing for high quality scores at both ends of the sequence. Adapters were trimmed from the ends of the antibiotic resistance genes' coding sequence using Trimmomatic (Bolger et al. 2014 (link)). Enrich2 (Rubin et al. 2017 ) was used to count the frequency of each allele for use in calculating selection coefficients and associated statistical measures. We set Enrich2 to filter out any reads containing bases with a quality score below 20, bases marked as N, or mutations at more than one codon.
Fitness of an allele (w_i) was calculated from the enrichment of the synonyms of the wild-type gene ( ${\epsilon }_{wt}$ ), the enrichment of allele i ( ${\epsilon }_i$ ) and the fold increase in the number of cells during the growth competition experiment (r) as described by Equation 1. We utilize the frequency of wildtype synonymous alleles as the reference instead of the frequency of wildtype because wildtype synonyms occurred more frequently in the library and wildtype sequencing counts are more prone to being affected by the artifact of PCR template jumping during the preparation of barcoded amplicons for deep-sequencing. Detailed derivations of the following equations (Equations 1–6) can be found in our previous work (Mehlhoff et al. 2020 (link)).
We calculate the variance in the fitness as
where the frequency of allele (f_i) is calculated from counts of that allele (c_i) and the total sequencing counts (c_T).
From the variance in fitness, we calculated a 99% confidence interval. Additionally, we calculated a P-value using a 2-tailed test. Details of the Z-score and P-value equations are available in Mehlhoff et al. (2020) (link).
We estimated the number of false positives that would be included at P < 0.01 and P < 0.001 significance in order to correct for multiple testing (Storey and Tibshirani 2003 (link)) in our DMS datasets as described previously (Mehlhoff et al. 2020 (link)). For TEM-1, we estimated that our data would contain approximately 55.0 false positives on average at P < 0.01 significance and an estimated 5.6 false positives on average at P < 0.001 significance for a single replica (Mehlhoff et al. 2020 (link)). Those values are 44.1 and 4.3 (CAT-I), 52.8 and 5.3 (NDM-1), and 33.8 and 3.4 (aadB) at P < 0.01 and P < 0.001 significance, respectively. We chose to report the frequency of mutations having fitness effects that met the P-value criteria in both replica experiments to limit the occurrence of false positives.