Blended Genome-Exome Sequencing is now available for clinical applications! Visit our Clinical BGE page.

Search
Close this search box.
Search
Close this search box.

Seeking Truth: Solving CNV Evaluation Challenges with T2T Genome Assembly

By: Yueyao Gao and Mark Fleharty

Copy Number Variations (CNVs) are pivotal in shaping human genetic diversity and influencing disease susceptibility. These segments, either deleted (DEL) or duplicated (DUP), can dramatically alter gene dosage, thereby modulating phenotypic outcomes1. Given their impact, the precision of CNV calling is not just a technical requirement but a critical factor for robust downstream analyses and accurate genetic interpretation. We are interested in evaluating the CNV calling capabilities of the latest available DRAGEN™ version with an eye toward future validation and inclusion in Broad’s existing clinical WGS pipeline (which currently uses DRAGEN™ v3.10.4).

A significant enhancement in the DRAGEN™ 4.2.4 CNV calling pipeline is that CNV calls now come with SV support. In previous versions, DRAGEN™ CNV callers primarily rely on depth signals for detecting CNV events. The underlying concept behind this approach is that short reads are randomly sampled on the genome, and copy number can be inferred from the read depth of aligned short reads2. However, while depth signal is reliable for detecting large CNV events, its efficacy diminishes significantly when dealing with shorter CNV events (<10kbp) due to a lower signal to noise ratio. Consequently, many smaller CNV events were historically left to be identified by the DRAGEN™ structural variant (SV) caller, which utilizes junction signals. In practice, this presented a challenge as researchers needed to perform an extra processing step to merge CNV events from both DRAGEN™ SV and CNV callers’ outputs, potentially introducing inaccuracies. In the latest iteration, DRAGEN™ 4.2.4 introduced an additional integration step that leverages junction signals from the SV caller and depth signal from CNV caller to generate the cnv_sv.vcf. This newly introduced VCF includes a SVCLAIM field which indicates how each CNV event was detected – whether through depth signal D, or junction signals J, or a combination of both DJ. Illumina has indicated that this improvement streamlines the workflow for researchers and boosts the accuracy of CNV calls, especially with shorter CNV events (<10kbp)3.

This blog introduces our latest effort in evaluating the performance of the DRAGEN™ v4.2.4 CNV caller with SV support. We have used the conventional method of comparing a query VCF file to benchmarking a VCF file which helped us in identifying true positives (TPs) when query variants matching events in benchmark datasets. However, we have found limitations of currently available benchmarking VCFs, often omitting relevant events, which hinder our ability to accurately estimate false positive rates – a pivotal metric in our clinical validation. Recognizing these limitations, we have developed a more robust and reliable alignment-based evaluation method to address shortcomings in the conventional approach while enhancing our understanding of CNV events. 

Old School: Conventional CNV Caller Evaluation

Traditionally, we utilize benchmarking datasets to evaluate the performance of our variant calling pipelines. The National Institute of Standards and Technology (NIST) has provided a benchmark ‘truth’ dataset for the HG002 genome, generated using dipcall – a reference-based variant calling pipeline. Dipcall employs the HG002-T2T (Telomere-to-Telomere)  v0.9 assembly as its ground truth. The T2T assembly represents one of the most accurate genome assemblies4,5 and offers a comprehensive view of CNV events. The assembly is then compared against the hg38 reference genome to generate a robust and precise variant benchmarking dataset for HG002. However, it’s important to note because dipcall relies on the hg38 genome as a template for variant calling, the output may be biased against some unresolved regions in hg38. Because long-read technology has superior capabilities in detecting structural variants, we also use PBSV calls from long-read HiFi data of the hg38 genome as another means for benchmarking accuracy on HG0026. It’s important to recognize this dataset, like the NIST benchmarking dataset, may have its own pitfalls. HiFi has its own limitation in detecting extra-long CNV (>20kbp) events. Those benchmarking datasets allow us to measure the accuracy of DRAGEN™ CNV calling pipeline. We have conducted a comparative analysis using Witty.er by comparing the results of DRAGEN™ v3.10.4 CNV calls, DRAGEN™ v3.10.4 bcftools concat CNV SV calls, and DRAGEN™ 4.2.4 CNV_SV calls against the HG002 dipcall and HiFi PBSV benchmarking datasets within the high-confidence region7. 

As illustrated in Figure 1, DRAGEN™ 4.2.4 demonstrates a strong correlation with dipcall and HiFi benchmarking datasets, suggesting its proficiency in identifying CNV DEL events across various event lengths, especially those under 10 kbp threshold. On an event-level assessment, the performance of the 4.2.4 cnv_sv caller is closely aligned with 3.10.4 bcftools merged calls. However, it’s worth noting that performance outcomes exhibit variations depending on the choice of benchmarking datasets. For instance, when using dipcall as benchmarking reference, DRAGEN™’s precision in the 20-50kbp and >50 kbp is notably higher than when utilizing HiFi PBSV as the benchmarking dataset. One possible explanation for this is the limitation of HiFi in detecting extra-long CNV events. Our results highlighted DRAGEN™ CNV caller performance and the impact of selected benchmarking datasets on performance evaluation. 

FIGURE 1. Precision and Recall of DEL Events by DRAGEN™ CNV in HG002 (hg38) Using Dipcall and HiFi PBSV as Benchmarking Datasets

 

For CNV detection, the task of identifying DUP events has long been acknowledged as more challenging as compared with identifying DEL events. This challenge can be attributed to lack of sensitivity of both depth signal and junction signal in distinguishing a change in copy number accurately8. Consequently, neither of the benchmarking datasets we use differentiate between insertions and duplications, which poses a challenge in assessing the performance of DRAGEN™ caller for DUP events. To address this limitation, we used an external tool SVWIDEN to annotate DUP events within the benchmarking datasets. Given the complexity of DUP event detection, it came as no surprise that our DUP events evaluation yielded less ideal results (Figure 2). Moreover, we also observed performance discrepancies across different benchmarking datasets pointing to the incompleteness of our benchmarking datasets for DUP events. This prompted us to explore an alternative alignment-based method to evaluate CNV events.

FIGURE 2. Precision and Recall of DUP Events by DRAGEN™ CNV in HG002 (hg38) Using Dipcall and HiFi PBSV as Benchmarking Datasets

 

A New Approach: Alignment-Based CNV Evaluation

In this method, we use an alignment search tool (LAST9) to identify CNV event sequences relative to both the HG002-T2T v0.9 assembly and the hg38 reference genome. We align the DNA sequence representing a DUP variant called in hg38 to the HG002-T2T reference.  A TP DUP event is characterized by a higher copy number in HG002-T2T than in hg38, while a FP DUP event shows fewer copies in HG002-T2T than hg38. Since the hg38 assembly is haploid, and the HG002-T2T assembly is diploid, we assume a copy neutral event has one copy in hg38 and two copies (one for the maternal and paternal haplotypes) in HG002-T2T.  A copy is defined as a LAST alignment where the alignment length is at least 95% of the query length and the percent identity of the alignment is at least 95%. This method allows us to identify not only whether a CNV event is correct, but in the case of duplications also allows us to identify the precise number, locations, and genotype of duplication events, even if they occur on separate chromosomes from the original call on hg38.

By using this method, we scrutinized certain DUP events that were absent in both HG002 benchmarking datasets we used. Our objective was to figure out whether these events were indeed false positives (FPs) or if their classification was a byproduct of the limitations inherent in the benchmarking datasets we relied upon. As shown in Figure 3B, a specific DRAGEN™-called DUP event with depth signal on chromosome 10 (chr10) attracted our attention. We identified seven copies on HG002-T2T’s chr10 in the maternal side and one copy on the paternal side. However, only one copy exists in the corresponding region in the hg38 genome. Although this DUP event was not present in either the dipcall or HiFi PBSV VCF, our alignment results strongly suggest that the genomic region chr10:39364454-39376272 in hg38 indeed represents a genuine DUP event in HG002. Furthermore, our alignment-based evaluation method unveiled a valuable insight into the mechanism behind this DRAGEN™-called DUP event, revealing it as a heterozygous tandem duplication event.

FIGURE 3. Detailed Analysis of a Unique DUP Event. A. An Integrated Genomics Viewer (IGV) screenshot displays a genomic region of interest. The highlighted red region represents a DUP event uniquely called by DRAGEN™, and absent in both HG002 dipcall VCF and HiFi PBSV VCF benchmarking datasets. Additionally, two green-highlighted DEL events are observed, with their benchmarking variants present in HG002 dipcall VCF but not in HiFi PBSV VCF. B. The alignment search results using LAST for the red-highlighted DUP event in 3A. Copies of this event are shown in hg38, HG002 maternal, and HG002 paternal chromosomes.

 

Obtaining a precise depiction of the genomic landscape of CNV events is a challenging quest. Our alignment-based evaluation method has proven itself as a powerful tool, shedding light not only on individual DUP events but also on the intricate relationships between multiple CNV events. Figure 4A depicts two DRAGEN™-called DUP events that were positioned approximately 70 kbp apart from each other. Neither DUP events were found in the benchmarking datasets that we used. What amplifies the intrigue is our discovery regarding DUP event a (chr1:16605769-16645359) and DUP event b (chr1:16715827-16727637), challenging the conventional spatial arrangement depicted in hg38 coordinates. According to hg38 coordinates, DUP event a occurred before DUP event b. However, our alignment results reveal a different scenario: on both the maternal and paternal chromosomes of HG002-T2T DUP event b precedes event a in sequence (Figure 4B). This intriguing twist is accompanied by a consistent presence of two copies on the maternal side and three copies on the paternal side, suggesting both DUP events are TPs and might be located on the same haplotype. Furthermore, our method reveals these copies to be dispersed, suggesting both events are homozygous dispersed DUP events. Our alignment-based CNV evaluation method not only unveils the intricate details of DUP events but also illuminates their spatial distribution, expanding our understanding of the complexity of CNV events.

FIGURE 4. Investigation of Two Adjacent DUP Events Identified by DRAGEN™. A. An IGV screenshot displays the genomic region containing DRAGEN™-called DUP event a (chr1:16605769-16645359) and DUP event b (chr1:16715827-16727637). Three genomic tracks from top to bottom include DRAGEN™ 4.2.4 output cnv_sv.vcf, HG002-T2T v0.9 dipcall VCF, HG002 HiFi PBSV VCF. B. Alignment results for DUP event a and DUP event b. Copies of DUP event a and DUP event b are shown in hg38, HG002 maternal, and HG002 paternal chromosomes.

 

In addition, we confronted an interesting case involving a DUP event located at chr17:22130212-22156606, a genomic region that also lacked any support from the benchmarking datasets we relied upon (Figure 5A). Unfortunately, we were unable to find supporting evidence for this 26kbp DUP event using our alignment-based method, which only had depth-only signal (SVCLAIM=D). As shown in Figure 5B, we observed only one copy of this sequence within the HG002 maternal chromosome, conspicuously absent on the HG002-T2T’s paternal chromosome. This observation defied our hypothesis, as we anticipated a DUP event sequence should have two or more copies of this sequence on either maternal or paternal chromosome. The presence of just a single copy on the HG002-T2T maternal chromosome strongly points to a putative false positive DUP event. It’s worth noting that our alignment-based evaluation method imposes stringent requirements on the precision of genomic coordinates. Thus, while it remains plausible that this genomic region harbors a genuine DUP event, its event length might be smaller than what’s indicated by DRAGEN™ CNV calling algorithm.

Figure 5. A Detailed Investigation of a DUP Events Identified by DRAGEN™ CNV. A. An IGV screenshot displays the genomic region harboring one DRAGEN™-called DUP event chr17:22130212-22156606. Three genomic tracks from top to bottom include DRAGEN™ 4.2.4 output cnv_sv.vcf, HG002-T2T v0.9 dipcall VCF, HG002 HiFi PBSV VCF. B. Alignment results for DRAGEN™-called DUP Event chr17:22130212-22156606. Copies of this DUP event are shown in hg38, HG002 maternal, and HG002 paternal chromosomes.

 

Overall, our alignment-based CNV evaluation method represents a pioneering approach in the intricate realm of CNV events. By harnessing the power of the HG002-T2T genome assembly, we’ve unlocked the ability to not only unveil CNV events in finer resolution but also delve deeper into spatial relationships and characteristics of adjacent CNV events.

Conclusion and Future Work

In this work, we evaluated the performance of the DRAGEN™ v4.2.4 CNV caller, utilizing dipcall events from HG002-T2T v0.9 assembly and HiFi PBSV data as our benchmarking datasets. It’s worth noting that some of the FPs initially identified using this method ended up revealing limitations within the benchmarking datasets. The limitations inherent in conventional evaluation approaches prompted us to embark on a journey towards developing a new alignment-based methodology by harnessing the power of the HG002-T2T genome assembly. Using this novel approach, we found some of the query FPs suggested by benchmarking datasets were actually TP events, highlighting the room for improvements of benchmarking datasets. Conversely, some FPs were indeed true FPs providing valuable insights that can inform methods or filtering strategies in future CNV callers.

While we have made significant strides in understanding CNV events within the HG002 sample, we acknowledge that a full validation across all DRAGENTM called CNV events leveraging our alignment-based approach is yet to be completed. As we anticipate a growing number of completed genomes emerging from the T2T effort, we hope to streamline and refine this alignment-based CNV evaluation method, providing a transformative means of evaluating CNV events across a broader spectrum of genomes.

References

  1. Ruderfer, et al. Patterns of genic intolerance of rare copy number variation in 59,898 human exomes. Nat. Genet, 2016
  2. Yoon S, et al. Sensitive and accurate detection of copy number variants using read depth of coverage. Genome Res, 2009
  3. https://www.illumina.com.cn/content/illumina-marketing/spac/en_AU/destination/Webinar-Illumina-DRAGEN-4-2-Enhanced-machine-learning-new-targeted-callers-and-more.html
  4. Wang T, et al. The Human Pangenome Project: a global source to map genomic diversity. Nature, 2022
  5. Rautiainen M, et al. Verkko: telomere-to-telomere assembly of diploid chromosomes. Nature, 2022
  6. Zook J, et al. A robust benchmark for detection of germline large deletions and insertions. Nature Biotech, 2022
  7. Ebert, Peter, et al. Haplotype-resolved diverse human genomes and integrated analysis of structural variation. Science, 2021
  8. Teo SM, et al. Statistical challenges associated with detecting copy number variations with next-generation sequencing. Bioinformatics, 2012
  9. Kielbasa, et al. Adaptive seeds tame genome sequence comparison. Genome Res, 2011
Important note for this blog: Posts do not equal endorsements. Opinions expressed in this blog are those of the author, on behalf of the genomics group at Broad. We make every effort to ensure the accuracy of data/figures presented here but these are not peer reviewed and errors may occur from time to time.

Return to the Blog

Services

Resources

Copyright © 2023 Broad Institute. All Rights Reserved. Privacy Policy | Terms & Conditions

Sean Hofherr

Chief of Clinical Strategy and Product Development, Broad Clinical Labs

Sean Hofherr is dual board certified by ABMGG in Clinical Biochemical Genetics and Clinical Molecular Genetics. Sean serves as the Chief of Clinical Strategy and Product Development at Broad Clinical Labs. In this role at BCL, Sean is able to leverage his extensive experience to guide the clinical vision and delivery across the organization. Sean most recently served as the Chief Operating Office at Fabric Genomics, which focuses on the use of AI and Bioinformatics for Clinical Interpretation of whole genome sequencing. Prior to Fabric, Sean was the Chief Scientific Officer and CLIA Director at the commercial reference laboratory, GeneDx.

Sean received his B.S. degree in Microbiology and Cell Sciences from the University of Florida before earning his Ph.D. in Molecular and Human Genetics from Baylor College of Medicine. Sean completed clinical fellowships in Clinical Biochemical Genetics and Clinical Molecular Genetics at the Mayo Clinic.

Danielle Perrin

Chief of Staff, Broad Clinical Labs

As Broad Clinical Labs’ Chief of Staff, Danielle Perrin advises and supports colleagues on the executive leadership team in BCL’s strategic planning and execution. She builds and leads new organizational functions and processes and leads critical projects, as well as driving effective information flow, decision making, and execution throughout the organization. An operations leader with a business, engineering, and biology background and 20+ years of experience in the genomics field, Perrin has a track record of driving operational excellence and building and scaling both physical and business processes. During her career at Broad, which started in 2003 at the tail end of the Human Genome Project, Perrin has led laboratory operations and R&D teams in Broad’s Genomics Platform, as well as fulfilling senior advisory and leadership roles in the Broad Institute’s COO and CFO offices.

Perrin received her B.S. in Biology and M.E. in Biotechnology Engineering from Tufts University and her M.B.A. from the MIT Sloan School of Management.

Tim De Smet

Chief Commercial Officer, Broad Clinical Labs

As Chief Commercial Officer of Broad Clinical Labs, Tim De Smet leads BCL’s business development, alliance management, external project management, and customer support teams. A Broad Institute employee since 2008, De Smet has held leadership roles and managed teams of various sizes in Broad’s Genomics Platform and clinical lab, spanning laboratory operations, finance, and informatics, and has expertise in work design, financial modeling, and high scale laboratory and business operations.

De Smet received his B.S. in Biochemistry and M.B.A. from Northeastern University.

Jim Meldrim

Chief Technology Officer, Broad Clinical Labs

As Chief Technology Officer, Jim Meldrim sets the vision for Broad Clinical Labs’ informatics systems, including the hardware and software used for sample intake and tracking, data production, analysis, and delivery. Having held a variety of laboratory and informatics-focused leadership roles at Broad, spanning R&D and production operations, Meldrim has been a leader and innovator in the generation, management, and analysis of genomic data since 1999, beginning with sequencing data generation for the Human Genome Project.

Meldrim received his B.S. in Biology from Cornell University.

Sheila Dodge

Chief Operating Officer, Broad Clinical Labs

As Chief Operating Officer, Sheila Dodge leads Broad Clinical Labs’ process development and implementation activities, as well as lab operations, financial planning and operations, quality & compliance, and core business processes. A Six Sigma Black Belt with extensive experience in process development and high throughput genomics operations, Dodge is an expert in work design and in collaborating with a range of collaborators, scientists, engineers, and technology partners to rapidly integrate new technologies and operationalize innovations. A member of the Broad Institute since 2001, Dodge is an Institute Scientist and lectures at the MIT Sloan School of Management on operations, dynamic work design, and visual management techniques.

Dodge received her B.A. in biochemistry and molecular biology from Boston University and her master’s degree in biology from Harvard University. She earned her M.B.A. from MIT Sloan School of Management.

Heidi Rehm, Ph.D., FACMG

Chief Medical Officer and Clinical Laboratory Director, Broad Clinical Labs

Heidi Rehm is board-certified by ABMGG in Clinical Molecular Genetics and Genomics and serves as BCL’s Chief Medical Officer and Clinical Laboratory Director. She oversees BCL’s regulatory requirements, leads the clinical team performing genomic interpretation and variant analysis, and guides BCL’s efforts in genomic testing for clinical and research use. She is also an Institute Member of the Broad and co-director of the Medical and Population Genetics Program. Rehm is also the Chief Genomics Officer in the Department of Medicine and Genomic Medicine Unit Director at the Center for Genomic Medicine at Massachusetts General Hospital, working to integrate genomics into medical practice. She is a principal investigator of ClinGen, providing free and publicly accessible resources to support the interpretation of genes and variants. She co-leads both the Broad Center for Mendelian Genomics, focused on discovering novel rare disease genes, and the Matchmaker Exchange, which aids in gene discovery. She is Chair of the Global Alliance for Genomics and Health, a principal investigator of the Broad-LMM-Color All of Us Genome Center, co-leader of the Genome Aggregation Database (gnomAD), and a Board Member and Vice President of Laboratory Genetics for the American College of Medical Genetics and Genomics.

Rehm received her B.A. degree in molecular biology and biochemistry from Middlebury College before earning her M.S. in biomedical science from Harvard Medical School and Ph.D. in genetics from Harvard University. She completed her post-doctoral training with David Corey in neurobiology and a fellowship in clinical molecular genetics at Harvard Medical School.

Niall Lennon, Ph.D.

Chair and Chief Scientific Officer, Broad Clinical Labs

As Chair and Chief Scientific Officer of Broad Clinical Labs, Niall Lennon leads the team and sets the scientific and clinical vision for the organization. Dr. Lennon joined the Broad Institute in 2006 and has since contributed to the development of applications for every major massively parallel sequencing platform across a range of fields. In 2013 Dr. Lennon led the effort to establish a CLIA licensed, CAP-accredited clinical laboratory at the Broad Institute to facilitate return of results to patients and to support clinical trials. More recently, he has led efforts to achieve FDA approval for large-scale genomics projects (NIH’s All of Us Research Program) and for Broad’s own clinical diagnostic for COVID-19 testing operation, which returned 37+ million results to patients. Dr. Lennon is a principal investigator of the eMerge and All of Us projects, an Institute Scientist at Broad, Associate Director of Broad’s Gerstner Center for Cancer Diagnostics, and an adjunct professor of biomedical engineering at Tufts University, where he teaches Molecular Biotechnology.

Dr. Lennon received a Ph.D. in pharmacology from University College Dublin and completed his postdoctoral studies at Harvard Medical School and Massachusetts General Hospital. He holds an executive certificate in management from the MIT Sloan School of Management.