Sequencing data stored in BCL format are demultiplexed through a process that uses the index sequences unique to each sample to assign clusters to the library from which they originated. Each cluster contains two indexes (i7 and i5 sequences, one at each end of the library fragment). The combination of those index sequences are used to demultiplex the pooled libraries.
After demultiplexing, this process generates FASTQ files, which contain the sequencing reads for each individual sample library and the associated quality scores for each base call, excluding reads from any clusters that did not pass filter.
The following sections describe performance testing methods.
Illumina tests the analytical performance of variant calling using an approach that covers the entire workflow including library preparation, sequencing, and secondary analysis. This approach is used to test a diverse selection of variants. When the variant calling pipeline is expanded to call a new variant class, this approach is always used.
This version of the software includes results generated by features tested in silico and by beta features. Beta features have not been fully evaluated for performance, see Beta Features.
Illumina uses in silico testing to the test the ability of the software to call an expanded scope of clinically relevant variants, including rare variants. In silico testing is used as a complementary method to analytical performance testing with wet lab step to expand the scope of testing. For example, while Illumina has analytically verified the performance of the software for calling complex variants in EGFR, the in silico testing approach characterizes the ability of the software to call complex variants in other genes.
For in silico testing, variants of interest are extracted from public databases like Cosmic and ClinVar. Each variant is simulated at different VAF levels by, depending on the variant class, spiking in mutant reads into a normal FFPE background (for sequence variants) or by increasing or decreasing the coverage of exons in the normal FFPE sample (for CNVs, for example, exon-level CNVs. The simulated reads match the expected quality of typical FFPE samples, such as fragment length, error rate, and family size. After the simulation, the software processes samples with spiked-in variants and determines the results. This approach does not include library prep and sequencing of tumor FFPE samples that include the rare variants of interest. The software reports these variants, but analytical verification was not performed.
DRAGEN TruSight Oncology 500 Analysis Software includes the following features that were tested i_n silico_ for both TruSight Oncology 500 and TruSight Oncology 500 HT:
Complex variants in genes beyond EGFR
Insertions and deletions > 25 bp
CNV amplifications
CNV deletions
Variants in intron-exon junctions (2 bp – 10 bp into introns)
In addition, the following features were tested in silico for TruSight Oncology 500 HT:
Exon-level CNVs in BRCA1 and BRCA2
This version of DRAGEN TruSight Oncology 500 Analysis Software includes beta features which have not been verified by Illumina due to limited access to samples or lack of an appropriate orthogonal method to perform testing, and, the use of in silico testing alone is not sufficient for verification purposes.
Customers are responsible for evaluating and demonstrating performance of any beta features they choose to implement. Beta features are indicated as such in the CombinedVariantOutput.tsv file. Illumina will continue to evaluate beta features with intent to fully release upon completion of verification for each feature.
This version includes the following beta features that may be used with the TruSight Oncology 500 HRD Assay:
Tumor fraction (beta)
Ploidy (beta)
Absolute copy numbers (ACN) (beta)
Gene-level loss of heterozygosity (LOH) events (beta)
Beta feature results are included in the Combined Variant Output file and other files. However, disclaimers that the results are generated by beta features are only provided in the Combined Variant Output file.
The software processes sequencing data to perform quality control, detect variants, determine tumor mutational burden (TMB), microsatellite instability (MSI) status, and genomic instability score (GIS), and report results. The following sections describe the analysis methods used in DRAGEN TruSight Oncology 500 Analysis Software.
DRAGEN TruSight Oncology 500 Analysis Software uses the following workflows to analyze sequencing data.
FASTQ Generation
DNA Analysis
DNA Alignment and Realignment
Read Collapsing
Indel Realignment and Read Stitching
Small Variant Calling
Small Variant Filtering
Copy Number Variant (CNV) Calling
Phased Variant Calling
Variant Merging
Annotation
Tumor Mutational Burden (TMB) Scoring
Microsatellite Instability (MSI) Status
Contamination Detection
RNA Analysis
Downsampling
Read Trimming
Alignment
Duplicate Marking
Fusion Calling
RNA Fusion Filtering
Splice Variant Calling
Annotation
Fusion Merging
Quality Control
Run QC
DNA Sample QC
RNA Sample QC
Refer to RNA Output for more information.
Each sample is downsampled to 30 million RNA reads. This number represents the total number of single reads (eg, R1 + R2, from all lanes). When using the recommended sequencing configurations or plexity, the samples can have fewer reads than the downsampling limit. In these cases, the FASTQ files are left as-is.
Reads are trimmed to 76 base pairs for further processing.
RNA alignment and fusion detection uses trimmed reads in FASTQ format as input. The outputs include a BAM file that contains duplicate-marked read alignments, an SJ.out.tab file that contains unannotated splice junctions, and a CSV file that contains fusion candidates.
DRAGEN aligns RNA reads in a transcript-aware mode using the human hg19 genome containing unplaced contigs (ie, chrUn_gl regions) and uses GENCODEv19 transcript annotations to identify splice sites. DRAGEN identifies and marks duplicate read alignments using start and end coordinates of alignments, which are adjusted for soft clipped reads.
Fusion and splice variant calling only use deduped fragments to score variants. DRAGEN identifies fusion candidates using chimeric split read alignments (pairs of primary and supplementary alignments) against multiple genes. DRAGEN scores and filters reads based on the various features of each candidate such as the number of supporting reads, mapping quality of supporting reads, and sequence homology between parent genes.
The DRAGEN RNA Fusion caller identifies gene fusions by searching for chimeric reads spanning two distinct parent genes. Based on the chimeric reads, DRAGEN first creates a list of fusion candidates, then scores the candidates to report the list of high confidence fusion calls from the candidate pool.
DRAGEN RNA Fusion caller performs the following steps:
Generates fusion candidate generation based on split read alignment.
Recruits additional evidence from fusion supporting discordant read pairs and soft-clipped reads.
Computes fusion candidate features such as gene coverage, read mapping quality, alternate allele frequency, gene homology, alignment anchor length, and breakpoint distance from exon boundary.
Scores and ranks the fusion candidates using a logistic regression model.
Selects a final list of fusion calls based on score and other filters including number of supporting reads, unique read alignment count, read through transcripts, and fusions matching the enriched regions.
RNA splice variant calling is performed for RNA sample libraries. Candidate splice variants (junctions) from RNA Alignment are compared against a database of known transcripts and a splice variant baseline of non-tumor junctions generated from a set of normal FFPE samples from different tissue types. Any splice variants that match the database or baseline are filtered out unless they are in a set of junctions with known oncological function. If there is sufficient read support, the candidate splice variant is kept. This process also identifies candidate RNA fusions.
Fusions identified during RNA fusion calling are merged with fusions from proximal genes identified during RNA splice variant calling. These are then annotated with gene symbols or names with respect to a static database of transcripts (GENCODE Release 19). The result of this process is a set of fusion calls that are eligible for reporting
The Illumina Annotation Engine annotates detected RNA splice variant calls with transcript-level changes (eg, affected exons in the transcript of a gene) with respect to RefSeq. This RefSeq database is the same RefSeq database used by the small variant annotation process.
The software calculates several quality control metrics for runs and samples.
The Run Metrics section of the metrics output report provides sequencing run quality metrics along with suggested values to determine if they are within an acceptable range. The overall percentage of reads passing filter is compared to a minimum threshold. For Read 1 and Read 2, the average percentage of bases ≥ Q30, which gives a prediction of the probability of an incorrect base call (Q‑score), are also compared to a minimum threshold. The following tables show run metric and quality threshold information for different systems.
The values in the Run Metrics section are listed as NA in the following situations:
If the analysis was started from FASTQ files.
If the analysis was started from BCL files and the InterOp files are missing or corrupt.
PCT_PF_READS (%)
Total percentage of reads passing filter.
≥80.0
All
PCT_Q30_R1 (%)
Percentage of Read 1 reads with quality score ≥ 30.
≥80.0
All
PCT_Q30_R2 (%)
Percentage of Read 2 reads with quality score ≥ 30.
≥80.0
All
PCT_PF_READS (%)
Total percentage of reads passing filter.
≥55.0
All
PCT_Q30_R1 (%)
Percentage of Read 1 reads with quality score ≥ 30.
≥80.0
All
PCT_Q30_R2 (%)
Percentage of Read 2 reads with quality score ≥ 30.
≥80.0
All
PCT_PF_READS (%)
Total percentage of reads passing filter.
≥85.0
All
PCT_Q30_R1 (%)
Percentage of Read 1 reads with quality score ≥ 30.
≥85.0
All
PCT_Q30_R2 (%)
Percentage of Read 2 reads with quality score ≥ 30.
≥85.0
All
PCT_Q30_R1 (%)
Percentage of Read 1 reads with quality score ≥ 30.
≥85.0
All
PCT_Q30_R2 (%)
Percentage of Read 2 reads with quality score ≥ 30.
≥85.0
All
DRAGEN TruSight Oncology 500 uses QC metrics to assess the validity of analysis for DNA libraries that pass contamination quality control. If the library fails one or more quality metrics, then the corresponding variant type or biomarker is not reported, and the associated QC category in the report header displays FAIL. Additionally, a companion diagnostic result may not be available if it relies on QC passing for one or more of the following QC categories.
DNA library QC results are available in the MetricsOutput.tsv
file. Refer to Metrics Output for details.
CONTAMINATION_SCORE
The contamination score is based on VAF distribution of SNPs.
Contamination Score ≤
All
MEDIAN_EXON_COVERAGE
Median exon fragment coverage across all exon bases.
≥ 150
Small variant TMB
PCT_EXON_50X
Percent exon bases with 50x fragment coverage.
≥ 90.0
Small variant TMB
MEDIAN_INSERT_SIZE
The median fragment length in the sample.
≥ 70
Small variant TMB
USABLE_MSI_SITES
The number of MSI sites usable for MSI calling.
≥ 40
MSI
MEDIAN_BIN_COUNT_CNV_TARGET
The median raw bin count per CNV target.
≥ 1.0
CNV
The input for RNA Library QC is RNA alignment. Metrics and guideline thresholds can be found in the MetricsOutput.tsv
file. Refer to Metrics Output for details.
MEDIAN_CV_GENE_500X
The median CV for all genes with median coverage > 500x. Genes with median coverage > 500x are likely to be highly expressed. Higher CV median > 500x indicates an issue with library preparation (poor sample input and/or probes pulldown issue).
Fusion Splice
MEDIAN_INSERT_SIZE
The median fragment length in the sample.
≥ 80
Fusion Splice
TOTAL_ON_TARGET_READS
The total number of reads that map to the target regions.
≥ 9000000
Fusion Splice
GENE_MEDIAN_COVERAGE
The median deduped coverage across all genes in the RNA panel (55 genes).
N/A*
Fusion Splice
To avoid failing RNA samples unnecessarily, Illumina does not recommend a universal threshold to determine RNA sample quality. RNA expression varies significantly across tissue types and a small panel size (55 genes), which makes normalization challenging. Tissue-specific thresholds could be considered for normalization.
DNA alignment and error correction involves aligning sequencing reads derived from DNA libraries to a reference genome and correcting errors in the sequencing reads prior to variant calling.
DRAGEN unique molecular identifier (UMI) error correction comprises three main steps:
DRAGEN UMI uses its HW accelerated mapper (based on a hash table implementation) to align DNA sequences in FASTQ files to the hg19 reference genome. These alignments are not written to a BAM.
The raw alignments are processed to remove errors, including errors introduced during FFPE preservation, PCR amplification, and sequencing. Reads from the same original DNA molecule are tagged with the same UMI during library preparation. The UMI allows DRAGEN to compare related reads, remove outlier signals, and collapse multiple reads into a single high-quality sequence. Read collapsing adds the following BAM tags:
RX/XU—UMI.
XV—Number of reads in the family.
XW—Number of reads in the duplex-family or 0 if not a duplex family.
DRAGEN performs a final alignment step on the UMI-collapsed reads. These final alignments are then written to a BAM file and a corresponding BAM index file is created.
DRAGEN continues to use these final alignments as input for gene amplification (copy number) calling, small variant calling (SNV, indel, MNV, delin), microsatellite instability (MSI) status determination, and DNA library quality control.
DRAGEN supports calling SNVs, indels, MNVs, and delins in tumor-only samples by using mapped and aligned DNA reads from a tumor sample as input. Variants are detected via both column wise pileup analysis and local de novo assembly of haplotypes. The de novo haplotypes allow the detection of much larger insertions and deletions than possible through column wise pileup analysis only. DRAGEN insertions and deletions are validated with lengths of at least 0–25 bp and more than 25 bp can be supported. In addition, DRAGEN also uses the de novo assembly to detect SNVs, insertions, and deletions that are co-phased and part of the same haplotypes. Any such co-phased variants that are within a window of 15 bp can then be reassembled into complex variants (MNVs and delins). The tumor-only pipeline produces a VCF file containing both germline and somatic variants that can be further analyzed to identify tumor mutations. Variant calling extends ± 10 bp into introns; details of the regions covered can be found in the assay manifest file. The pipeline makes no ploidy assumptions, enabling detection of low-frequency alleles.
DRAGEN small variant calling includes the following steps:
Detects regions with sufficient read coverage (callable regions).
Detects regions where the reads deviate from the reference and there is a possibility of a germline or somatic call (active regions).
Assembles de novograph haplotypes are assembled from reads (haplotype assembly).
Extracts possible somatic or germline calls (events) from column wise pileup analysis.
Calibrates read base qualities to account for FFPE noise.
Computes read likelihoods for each read/haplotype pair.
Performs mutation calling by summing the genotype probabilities across all reads/haplotype pairs.
Performs additional filtering to improve variant calling accuracy, including using a systematic noise file. The systematic noise file indicates the statistical probability of noise at specific positions in the genome. This noise file is constructed using clean (normal) samples. Regions where noise is common (eg, difficult to map regions) have higher noise values. The small variant caller penalizes those regions to reduce the probability of making false positive calls.
The DRAGEN copy number variant caller performs amplification, reference, and deletion calling for CNV targets within the assay. It counts the coverage of each target interval on the panel, uses a preprocessed panel of normal samples to normalize target counts, corrects for GC coverage bias, and calculates scores of a CNV event from observed coverage and makes copy number calls.
The BRCA large rearrangement step generates segmentation of the BRCA1 and BRCA2 genes for exon-level CNV detection from the BAM file. Using the same method as CNV calling, the large rearrangement component counts coverage of each target interval of the panel, performs normalization, and calculates the fold change values for each probe across the BRCA genes. Normalization includes GC bias correction, sequencing depth, and probe efficiency using a collection of normal FFPE and genomic DNA samples. Initial segmentation is performed for each gene with circular binary segmentation. The merging of segments is then determined by amplitude, noise, and variance at adjacent segments using thresholds established with in silico data. A large rearrangement is reported for genes with more than one segment. Coordinates of the exon-level CNV and the log2 mean fold change for each of the BRCA gene segments are found in the *_DragenExonCNV.json
file.
The Illumina Annotation Engine performs annotation of small variants, CNVs, and exon-level CNVs. The inputs are gVCF files and the outputs are annotated JSON files.
The Illumina Annotation Engine processes each variant entry and annotates with available information from databases such as dbSNP, gnomAD genome and exome, 1000 genomes, ClinVar, COSMIC, RefSeq, and Ensembl. The header includes version information and general details. Each annotated variant is included as a nested dictionary structure in separate lines following the header.
The following table shows version information for each annotation database:
DRAGEN is used to compute tumor mutational burden (TMB) in coding regions where there is sufficient coverage.
The following variants are excluded from the TMB calculation:
Non-PASS variants.
Mitochondrial variants.
MNVs.
Variants that do not meet a minimum depth threshold.
Variants that do not meet the minimum variant allele threshold.
Variants that fall outside the eligible regions.
Tumor driver mutations. Variants with a population allele count ≥ 50 are treated as tumor driver mutations. Germline variants are not counted towards TMB. Variants are determined as germline based on a database and a proxy filter.
Variants with a population allele count ≥ 10 that are observed in either the 1000 Genomes or gnomAD databases are marked as germline. MNVs, which do not count towards TMB, may be marked as germline when all their component small variants are marked as germline. The proxy filter scans the variants surrounding a specific variant and identifies those variants with similar variant allele frequencies (VAF). If the majority of surrounding variants of similar VAF are germline, then the variant is also marked as germline.
The formula for TMB calculation is:
Outputs are captured in a _TMB_Trace.tsv
file that contains information on variants used in the TMB calculation and a .tmb.json
file that contains the TMB score calculation and configuration details.
DRAGEN can determine the MSI status of a sample. It uses a normal reference file, which was created from a set of normal samples. During sequencing, normal reference files are generated by tabulating read counts for each microsatellite site. The normal file contains the read count distribution for each microsatellite.
MSI calling for a tumor-only sample is performed by first tabulating tumor counts from the read alignments for each microsatellite site. Then, the Jensen-Shannon distance (JSD) is calculated between each pair of tumor and normal baseline samples. DRAGEN determines unstable sites by performing Chi-square testing of tumor JSD and normal JSD distributions. Unstable sites are called if the mean distance difference of the two JSD distributions is ≥to the distance threshold and Chi-square p-value is ≤ to the p-value threshold. Lastly, DRAGEN produces an MSI status given assessed site count, unstable site count, the percentage of unstable sites in all assessed sites, and the sum of the Jensen-Shannon distance of all the unstable sites.
Requires HRD add-on assay
Genomic instability score (GIS) is a whole genome signature for homologous recombination deficiency. The GIS is composed of the sum of three components: loss of heterozygosity, telomeric allele imbalance, and large-scale state transition. These components are estimated using the GIS algorithm contracted from Myriad Genetics, which uses an input of the b-allele frequency and coverage across a genome-wide single nucleotide panel. A panel of normal samples is used for both bias reduction and normalization prior to GIS estimation. Final GIS results can be found in the *.gis.json
file.
The contamination analysis step detects foreign human DNA contamination using the SNP error file and pileup file that are generated during the small variant calling and the TMB trace file. The software determines whether a sample has foreign DNA using the contamination score. In contaminated samples, the variant allele frequencies in SNPs shift from the expected values of 0%, 50%, or 100%. The algorithm collects all positions that overlap with common SNPs that have variant allele frequencies of < 25% or > 75%. Then, the algorithm computes the likelihood that the positions are an error or a real mutation. The contamination score is the sum of all the log likelihood scores across the predefined SNP positions with minor allele frequency < 25% in the sample and are not likely due to CNV events.
The larger the contamination score, the more likely there is foreign DNA contamination. A sample is considered to be contaminated if the contamination score is above predefined quality threshold. The contamination score was found to be high in samples with highly rearranged genomes or HRD samples. 1% of HRD samples found to be above the threshold with no evidence for actual contamination.
This is a beta feature. Beta feature results are included in the Combined Variant Output file and other files. However, disclaimers that the results are generated by beta features are only provided in the Combined Variant Output file. Requires HRD add-on assay.
Tumor fraction is calculated as described in the User Guide, section “HRD Metrics Report” and leverages the Myriad Genetics algorithm. Tumor fraction is output in the Logs_Intermediates/Gis/SAMPLE/SAMPLE.gis.json and Combined Variant Output file.
This is a beta feature. Beta feature results are included in the Combined Variant Output file and other files. However, disclaimers that the results are generated by beta features are only provided in the Combined Variant Output file. Requires HRD add-on assay.
Ploidy is calculated as described in the User Guide, section “HRD Metrics Report” and leverages the Myriad Genetics algorithm. Ploidy is output in the in the Logs_Intermediates/Gis/SAMPLE/SAMPLE.gis.json and Combined Variant Output file.
This is a beta feature. Beta feature results are included in the Combined Variant Output file and other files. However, disclaimers that the results are generated by beta features are only provided in the Combined Variant Output file. Requires HRD add-on assay.
Absolute copy numbers are calculated by leveraging the Myriad Genetics algorithm. The algorithm segments the entire genome using the HRD panel and provides an A and B allele estimate for each segment. After the TSO 500 pipeline determines CNV calls (using the TSO 500 panel), the segment covering the gene is identified, and the A and B allele numbers of the segment overlapping the gene are reported. If the gene is within 300 kbases from the segment boundary, the estimate is unreliable and “-1” is output. Absolute copy numbers are output in the Logs_Intermediates/Gis/SAMPLE/SAMPLE.abcn_annotated.vc f, Logs_Intermediates/Gis/SAMPLE/SAMPLE.abcn_genes.tsv and Combined Variant Output file.
This is a beta feature. Beta feature results are included in the Combined Variant Output file and other files. However, disclaimers that the results are generated by beta features are only provided in the Combined Variant Output file. Requires HRD add-on assay.
Gene-level loss of heterozygosity is calculated based on the minor copy number reported in the abcn_annotated.vc f. If the minor copy number is 0 then the gene is assumed to have a loss of heterozygosity. Gene-level loss of heterozygosity is output in the Logs_Intermediates/Gis/SAMPLE/SAMPLE.abcn_genes.tsv and Combined Variant Output file.
gnomeAD
2.1
COSMIC
v84
ClinVar
2019-02-04
dbSNP
v151
1000 Genomes Project
Phase 3 v5a
RefSeq
NCBI Homo sapiens Annotation Release 105.20201022