Near-perfect diploid genome benchmark advances personalized medicine

For decades, scientists have relied on a single reference genome to understand human genetic variation—a shortcut that has worked well enough for many purposes but leaves blind spots in the most medically important places. A team of researchers has now built something fundamentally different: a complete map of both copies of a human genome, maternal and paternal, accurate to 99.4% across its entire length. The benchmark, called T2T-HG002, represents a shift in how genomics might work in the clinic, moving away from comparing individual genomes against a fixed standard and toward building personalized maps from the ground up.

The problem with conventional genome sequencing is structural. When researchers want to find disease-causing mutations, they typically take a patient's genetic data and compare it to an existing reference sequence. But that reference is itself incomplete and biased—it represents only one haplotype, one set of inherited chromosomes, and it contains gaps and errors in the regions that matter most for human health. Short-read sequencing technology struggles with repetitive DNA, duplicated genes, and the complex structural variations that cluster in these hard-to-sequence zones. Researchers end up missing variants in places like immunoglobulin loci and copy-number variable genes, the very regions where genetic disease often hides.

T2T-HG002 was built using a different architecture. Instead of starting with an existing reference and calling variants against it, the researchers sequenced a single individual—HG002, a well-studied cell line—using multiple long-read technologies: PacBio HiFi for raw sequence data, Oxford Nanopore for additional coverage, and supplementary Illumina reads for phasing. They used Hi-C and Strand-seq data to scaffold the assembly correctly in three-dimensional space, and fluorescence microscopy to validate the most difficult regions. Over successive rounds of error correction and validation, they refined the assembly until it reached a quality score of Q68.9, meaning fewer than one error per 68 billion base pairs. The result contains both the maternal and paternal copies of every chromosome, including the sex chromosomes—a true diploid genome, not a haploid abstraction.

The benchmark adds 918.2 megabases of high-confidence sequence that earlier reference genomes simply lacked: 701.4 megabases of autosomal DNA and 216.8 megabases of sex chromosome sequences. These are not trivial additions. They include entire genes that exist in only one parental copy, like DUSP22 and the complement factor H-related genes, which are known to vary between individuals and affect immune function. The team identified 13 maternal-only and 12 paternal-only autosomal genes—variants that would be invisible in a haploid reference. Within the regions both benchmarks cover, the new standard reduced discrepancies with the older Genome in a Bottle benchmark from nearly 6,000 variants down to 219, suggesting that most of the remaining differences are errors in the older benchmark itself.

What makes T2T-HG002 transformative is not just its accuracy but its framework. The researchers developed software called GQC that can evaluate any new genome assembly by comparing it directly to the diploid benchmark, distinguishing between the two parental copies and assessing accuracy at the base level, the haplotype level, and across large structural changes. When they applied this tool to five years of HG002 assemblies generated with different technologies, they found that de novo genome reconstruction—building a complete genome from scratch rather than calling variants against a reference—recovered two to seven percent more sequence and achieved roughly tenfold greater accuracy. This suggests that the future of personalized genomics may not be variant calling at all, but rather complete genome assembly for each patient.

The benchmark is now a National Institute of Standards and Technology reference material, meaning it can serve as a standard for evaluating sequencing platforms, assembly algorithms, and clinical interpretation methods. But significant work remains. Some regions, particularly the ribosomal DNA arrays, remain unfinished even in this near-complete benchmark. Clinical genomics still relies primarily on variant calling, and no standardized methods yet exist for translating whole-genome benchmarking performance into actionable clinical interpretation. The authors note that HG002 cells contain only low levels of somatic variation, so the benchmark may not fully capture the complexity of tumor genomes or tissues with high mutation rates. Still, the framework is in place. As sequencing technology continues to improve and researchers apply the same approach to diverse populations, the path toward truly personalized genomic medicine—one that does not force every patient's genome through the lens of a single reference—has become visible.