Requires: htslib (http://www.htslib.org/). Set HTSDIR in the make file to the htslib folder and make. Note that we merely call make on htslib, as such, its dependencies and system requirements need to be fulfilled.
Compile:
git clone https://github.com/samtools/htslib.git
make
Usage:
eagle -v variants.vcf -a alignment.bam -r reference.fasta > output.tab
scripts/example.sh describe some EAGLE workflows as a bash script
For EAGLE read classification options, see EAGLE-RC below
A tab-delimited text file with one row per variant and columns representing:
The read counts represent reads that are unambiguously for the reference or alternative sequence (2x the probability favoring one over the other), as opposed to aligned reads. Our model attempts to account for various uncertainties in the hypotheses.
-v --vcf [FILE] VCF file describing the variants, only the columns describing position and sequence are used [columns: 1,2,4,5].
-a --bam [FILE] BAM alignment data file, reference coordinated sorted with index [filename.bam.bai].
-r --ref [FILE] Reference genome [multi] fasta file.
-o --out [FILE] Output file name, defaults to stdout if not specified.
-t --nthread [INT] The number of processes to use. Default is 1.
-s --sharedr [INT] Group/chain nearby variants based on shared reads. Default is 0/off, which uses the distance based method. Option 1 will group variants if any read that crosses the first variant also cross the variant under consideration. Option 2 will group variants if any read that crosses any variant in the set also cross the variant under consideration.
-n --distlim [INT] Group/chain nearby variants within n bp of each other to be considered in the set of hypotheses for marginal probability calculations. Default is 10 bp (0 is off).
-w --maxdist [INT] Maximum number of bases between any two variants in the set of hypotheses. This sets a window size that allows for larger values of -n without chaining an excessive number of variants. Default is 0 bp (off).
-m --maxh [INT] The maximum number of hypotheses to be tested. Instead of looking at all 2^n combinations for a set of variants, if after the current k for n choose k combinations is considered and the number of tested hypthotheses exceeds the maximum, then do not consider more combinations. The solo variants and the "all variants" hypothesis are always tested first and do not count towards the maximum, thus any value greater than 0 ensures all doubles are tested. Default is 1024 (2^10).
--mvh Instead of the marginal probabilities over the hypotheses set, output only the maximum variant hypothesis (highest probability of phased variants) in the set of hypotheses. Note that maxh will limit the possible combinations tested.
--pao Use primary alignments only, based on SAM flag.
--isc Ignore soft-clipped bases in reads when calculating the probabilities, based on cigar string.
--nodup Ignore marked duplicate reads, based on SAM flag.
--splice Reads are from RNA-seq and potentially spliced, based on cigar string.
--bs [INT] Reads are bisulfite treated. In the probability model, this considers C to T (top strand) and G to A (both strand) as matches. 0: off, 1: top/forward strand, 2: bottom/reverse strand, 3: both. Default is 0 (off).
--dp Instead of the short read model, which assumes no indel errors, use dynamic programming (short in long) to calculate the likelihood. This allows handling of long reads which have higher rates of sequencing errors and indel errors.
--gap_op [INT] Gap open penalty for use with --dp. Default is 6. For long reads that contain indel errors, 2 may be a better.
--gap_ex [INT] Gap extend penalty for use with --dp. Default is 1.
--verbose Verbose mode. Output the likelihoods for every read seen for every hypothesis to stderr. Used in read classification with eagle-rc.
--lowmem Low memory usage mode. For SNPs, we use a method to quickly derive the alternative hypothesis probability from the reference hypothesis probability without constructing the alternative sequence in memory. For indels, which can be treated as a series of SNPs, this method may not be faster depending on read depth due to the number of frameshifted bases to account for. Though it will save memory which may allow for more threads without hitting some memory cap.
--phred64 Reads quality scores are in phred64. Default is phred33.
--hetbias [FLOAT] Prior probability bias towards heterozygous or homozygous mutations. Value between [0,1] where 1 is towards heterozygosity. Default is 0.5 (unbiased).
--omega [FLOAT] Prior probability of originating from outside paralogous source (i.e. not from reference genome and also not from the candidate variant genome). Value between [0,1]. Default is 1e-5.
--rc Wrapper for read classification settings: --omega=1.0e-40 --isc --mvh --verbose --lowmem.
compare2TruthData.py: Separate false positives and true positives based on truth data given as a VCF.
compile_likelihoods.py: Post-process the probabilities calculated by EAGLE and compile the results into tab-delimited table format. It can be used to find somatic mutations given positive (i.e. tumor) and negative (i.e. normal) results on the same set of variants. Likelihood ratio and allele frequency thresholds are then used to filter mutations.
combine_vcf_eagle.py: Annotate the vcf file with the results from compile_likelihoods.py. Raw EAGLE output can also be used if specified in the options.
Heterozygous non-reference variants (VCF: comma separated multiple alternative sequences) are output as separate entries. Furthermore, if they are near other variants, it will separately consider each alternative sequence in their own sets so that phasing is not an issue. This may result in entries with the first 4 columns being identical. The user will need to examine the variant set of these cases to determine which it belongs to. The script compileLikelihood.py will naively retain the one with the maximum probability, for entries with identical coordinates.
EAGLE-RC is a method for classifying whether a read belongs to one genomic hypothesis or another, given a set of genotype differences between them. This can be applicable for determining if reads originate from a specific allele or from a specific homeolog in allopolyploids.
EAGLE-RC can also classify reads directly via calculating the likelihood of the read given two alignments to different (sub)genome references, without knowing the genotype differences explicitly, implemented as no genotype info (--ngi) mode.
First with explicit genotype difference information, use EAGLE to calculate likelihoods for each read and for each hypothesis (--verbose) as well as phased variants likelihoods (--mvh) as output. Then the program eagle-rc can take these two inputs to classify reads and optionally split reads into bam files for each class. We also use a lower omega to be more tolerant to sequence differences outside the tested hypotheses. The wrapper option (--rc) takes care of these parameters. Other options (such as --dp for long reads, --splice for RNA-seq, --isc to ignore soft clipped bases, etc.) may or may not be applicable depending on the use case.
Usage, with more details in example.sh:
eagle -t 2 -v variants.vcf -a alignment.bam -r reference.fasta --rc 1> output.tab 2>readinfo.txt
eagle-rc -a alignment.bam -o out_prefix -v output.tab readinfo.txt > classified_reads.list
-v --var [FILE] EAGLE output file, containing variant likelihoods
-a --bam [FILE] BAM alignment data file to be processed and whose reads are to be classified, reference coordinated sorted with index
-o --out [String] Output file name prefix
-u --unique [FILE1,FILE2,...] Optionally, also output reads that are unique against other BAM files (comma separated list)
--listonly Print classified read list only (stdout) without processing BAM files
--readlist Read from classified read list instead of EAGLE outputs and process BAM files. This is useful if you previously outputed with listonly and then performed post-processing on list files to obtain, for example, a consensus list. This will allow the post-processed list to be used instead.
--refonly Write REF classified reads only when processing BAM file.
--paired Consider paired-end reads together.
--pao Use primary alignments only, based on SAM flag.
For no genotype information classification, the options in the default mode listed above are also applicable. Usage, where the classification is from the point of view of ref1 as the reference hypothesis and ref2 as the alternative hypothesis:
eagle-rc --ngi -o out --ref1=ref1.fa --ref2=ref2.fa --bam1=align1.bam --bam2=align2.bam > classified_reads.1vs2.list
Note: This does not have the advantage of the default usage where explicit genotype information is used in the form of variants. For example a read did not map to ref2 due to divergence between sample and reference. However, we could still evaluate this in the default usage by considering the explicit genotype difference between ref1 and ref2, with ref1 as the reference where the read was mappable. The evaluation showed that the read was more likely to be ref2 genotype despite being mappable only in ref1, in which case we designated this read as unknown UNK. This case has been observed to truly happen in our benchmarks where the truth being the read is derived from ref2 is known. Using --ngi in this case would give a false ref1 classification.
--ref1 [FILE] Reference genome 1 fasta file
--bam1 [FILE] Alignments to reference genome 1, --ref1, bam file
--ref2 [FILE] Reference genome 2 fasta file, we recommend that the chromosome names are different between ref1 and ref2.
--bam2 [FILE] Alignments to reference genome 2, --ref2, bam file
--isc Ignore soft-clipped bases in reads when calculating the probabilities, based on cigar string.
--nodup Ignore marked duplicate reads, based on SAM flag.
--splice Reads are from RNA-seq and potentially spliced, based on cigar string.
--bs [INT] Reads are bisulfite treated. In the probability model, this considers C to T (top strand) and G to A (both strand) as matches. 0: off, 1: top/forward strand, 2: bottom/reverse strand, 3: both. Default is 0 (off).
--phred64 Reads quality scores are in phred64. Default is phred33.
A tab-delimited text file with one row per read (pair if --paired) and columns representing:
Output BAM files which classify reads into 4 classes (reads from -a align.bam):
If --ngi then there are 2 sets of output bam files:
In the classified read list, keep in mind that the reference focus REF is in ref1 coordinates. If ref2 is listed then that is because it is more likely as ALT. One way to get the reversed list is to switch REF and ALT along with their likelihoods:
perl -ne 'chomp; @t=split(/\t/); if($t[1] eq "REF"){$t[1]="ALT";} elsif($t[1] eq "ALT"){$t[1]="REF";} $tmp=$t[4]; $t[4]=$t[5]; $t[5]=$tmp; print join("\t", @t)."\n";' classified_reads.1vs2.list > classified_reads.2vs1.list
Or rerun with:
eagle-rc --ngi --listonly --ref1=ref2.fa --ref2=ref1.fa --bam1=align2.bam --bam2=align1.bam > classified_reads.2vs1.list
The lists can then be passed on for further processing such as in hexaploid analysis.
Tony Kuo and Martin C Frith and Jun Sese and Paul Horton. EAGLE: Explicit Alternative Genome Likelihood Evaluator. BMC Medical Genomics. 11(Suppl 2):28. https://doi.org/10.1186/s12920-018-0342-1