Family-Based Genomics Benchmark Promises More Accurate Variant Detection

By Allison Proffitt 

August 20, 2025 | A multi-generational family tree is helping researchers dramatically improve the accuracy of genomic variant calling, according to a new study published in Nature Methods (DOI: 10.1038/s41592-025-02750-y). The research, led by scientists at Pacific Biosciences, introduces a comprehensive benchmarking dataset that could reshape how genomic technologies are trained and validated. 

The study uses the latest versions of several leading sequencing technologies—PacBio’s HiFi sequencing, Illumina, and Oxford Nanopore—to sequence a four-generation family pedigree from Utah—CEPH-1463. Combining various sequencer data with Mendelian inheritance rules gives the new benchmark unprecedented accuracy and coverage. Unlike existing genomic benchmarks that focus on highly confident but limited regions of the genome, the new benchmark tackles complex genomic territories that have historically been too difficult to assess accurately. 

“We’ve sequenced this pedigree with our collaborators using a variety of different technologies,” explained Mike Eberle, PacBio’s VP of Computational Biology and senior author on the paper, who has been working with this pedigree for years. In addition to PacBio’s HiFi sequencing, the benchmark also compared platforms from Illumina and Oxford Nanopore Technologies. PacBio worked with collaborators from the University of Washington School of Medicine, Google, National Institute of Standards and Technology (NIST), and other groups.  

“What we really wanted to do was concentrate on the parents and their children to build up a dataset where we could say, is this variant correct or is it not correct?” 

A Family Affair for Genomic Accuracy 

The pedigree offers key insight that sequencing alone can’t reveal. By examining how variants are passed from parents to their eight consented children, researchers can create predictable inheritance patterns across the genome. If a variant appears in a parent, the pattern of which children inherit it validates (or invalidates) the variant call. 

Samples (whole blood when available or cell line) were sequenced, and data were aligned to GRCh38 to generate small variant calls (SNVs and indels). Structural variant calls were made using the long-read data (PacBio and ONT) using both alignment-based genotypers and assembly-based callers.   

Combining the various sequencer data creates the most accurate truth set. “We can put all these together and say, ‘Ok, maybe one technology didn’t call this position correct, but another one did,’” Eberle said. “So that the truth set isn’t biased toward, for example, a particular technology.”  

This family-based validation system allows researchers to assess variants that would be impossible to validate using traditional single-sample or even trio-based approaches. The method is particularly powerful for structural variants and complex genomic regions that have been largely ignored by existing benchmarks. 

The pedigree itself has deep roots in genomics research. While this study used three generations, there are four in the pedigree now, numbering 28 family members. Members of this family tree have been part of landmark projects including HapMap and the 1000 Genomes Project. One individual, known as NA12878, “might be the most sequenced individual ever,” according to Eberle. 

Beyond Traditional Benchmarks 

Current genomic benchmarks, such as those developed most recently by the Genome in a Bottle (GIAB) consortium at NIST, take a highly conservative approach: focusing on variants they can call with extremely high confidence. While this creates reliable benchmarks for clinical validation, it leaves significant gaps for technology developers working on the most challenging parts of the genome. 

When compared with v4.2.1 of GIAB, the new Platinum Pedigree dataset identifies 11.6% more SNVs and 39.8% more indels in NA12878, the authors report in the paper.  

“They have very high confidence [in what variants they call], and part of that is because they filter away a lot of things around the complex regions,” Eberle explained. “I would say that for a tool developer, that’s not as good of a truth set. It’s very good for labs that are validating clinical pipelines and things like that.” 

Justin Zook, Co-Leader, Biomarker and Genomic Sciences Group at NIST and co-author on the Nature Methods paper, sees the Platinum Pedigree work as complementary rather than competitive to GIAB benchmarks. “When you’re evaluating methods, all benchmarks have strengths and weaknesses,” he said.  

“We tend to be a little more conservative in what we include in Genome in a Bottle,” Zook agrees. “We tend to make sure that when you’re comparing against the benchmark, the vast majority of the false positives and false negatives that you identify should really be false positives and false negatives. That causes us to exclude some of the more complex regions of the genome that can’t be benchmarked reliably right now.”   

The Genome in a Bottle benchmark is an assembly-based benchmark, Zook explains, focused on the HG002 sample from a genomic trio, so these benchmarks reflect different families. “It’s always good to have more benchmarks for more samples,” Zook said.  

The pedigree-based benchmark takes into account inheritance patterns to understand variant confidence. “They certainly also get down to a really highly confident set of variants in the end, but there are some edge cases that we would exclude from our assembly-based approach that don’t end up getting excluded from this pedigree-based approach,” he said.   

As far as Zook is concerned, it’s valuable to have a variety of benchmarks. GIAB is pursuing de novo benchmarks as well. “We’ve been doing in depth work trying to get a really, really accurate de novo assembly of the HG002 individual and polishing that assembly to make it as accurate as we can,” he added. Benchmarks from that work will soon be published. 

Improvements in AI Training 

Zook does note that the comprehensive coverage of the Platinum Pedigree benchmark makes it particularly valuable for training machine learning models and validating new variant calling technologies. 

And that’s what the researchers did. They used the new benchmark to retrain DeepVariant, Google’s widely-used AI system for variant calling. Retraining DeepVariant with the new data “reduced errors in small variant calls by 38.4% for SNVs and 19.3% for indels when evaluated on the Platinum Pedigree truth set, and by 18.2% for SNVs and 5.3% for indels when evaluated on GIAB,” the authors write in the paper.  

“This is primarily because there’s a lot of variants in here in complex regions that weren’t normally tested on a more specific truth set,” Eberle said. This new training will be integrated into DeepVariant v1.8.  

A Living Genomic Document 

Eberle describes the Platinum Pedigree as an evolving resource. The team continues to sequence samples with new technologies and refine their analysis methods, including visual inspection of complex calls using genome visualization tools. 

“I don’t think this is a finished product. This is like kind of a living document of this family’s genome that can actually continue to improve over time,” Eberle said. 

Future work will focus on incorporating more assembly-based methods to better characterize complex genomic regions and expanding coverage to approach complete genomic characterization. The researchers also plan to continue improving their variant calling algorithms, including PacBio’s structural variant caller called Sawfish. 

Broader Impact for the Genomics Community 

The entire dataset is publicly available, making it a resource that extends far beyond PacBio’s commercial interests. Academic groups, including Evan Eichler’s lab at the University of Washington, are already incorporating the benchmark into their research. 

For Eberle, who has been working with this pedigree since before joining PacBio, the dataset represents a debugging tool for complex genomic regions. “Let’s say we make a caller for a complex region of the genome ... as we develop it, we will test it out on this family and just say, okay, are the variant calls actually agreeing with what we expect?” 

The work addresses a critical need in genomics: how to validate accuracy in the most challenging regions of the human genome. As sequencing technologies continue to advance and researchers tackle increasingly complex genomic questions, benchmarks like the platinum pedigree provide essential infrastructure for ensuring that new methods actually work as intended.