HPRC2: A human pangenome reference with near-complete coverage of common genetic variation

Key results

1. 01

   ShownGreedy MaxVar sample selection lets 460 haplotypes cover 99.43% of common (AF ≥ 1%) small variants found in All of Us v8, up from 91.92% for HPRC1. Coverage of rarer variants (AF ≥ 0.1%) is still far from saturated.

   Of 7,274,300 common small variants in All of Us v8 (N≈440,000), HPRC2 covers 7,233,162 (99.43%) vs 6,686,383 (91.92%) for HPRC1 (88 haplotypes). At AF ≥ 0.1%: 83% vs 62%. Extrapolation projects ≥99.9% common coverage at 932 haplotypes. An incidence-based estimate gives a 98.77% chance that a random variant from a new sample is already in the graph. · Capturing greater than 99% of common variation; An Expanded Map of Haplotype Resolved Variation

2. 02

   ShownHifiasm assembly from HiFi plus ultra-long ONT reads, then DeepPolisher, gives more complete and accurate genomes than HPRC1. Automated T2T assembly is now possible (59% of chromosomes are T2T contigs or scaffolds), the 8p23.1 inversion region assembles without gaps, and Hi-C phasing performs close to trio phasing.

   3,666 chromosomes (35%) assembled as gapless T2T contigs and 2,490 (24%) as T2T scaffolds; mean k-mer QV 55 vs 52.5 for HPRC1; unreliable windows per haplotype 3.1% vs 6.9%. At 8p23.1 every HPRC2 assembly is gapless vs gaps in 45.7% of HPRC1 haplotypes. Complete, accurate active centromere arrays: 66% vs 15%. Trio vs Hi-C phasing: QV 57.09 vs 56.04. · Assembling thousands of complete, T2T…; Discussion

3. 03

   ShownThe base-level Minigraph-Cactus graph holds more variant sites than HPRC1's and agrees better with consensus callsets. The new Pantree and GRef coordinate systems let variants inside non-reference sequence be written in VCF.

   Small-variant sites: 34.49 million vs 20.78 million; SV sites: 109.16 thousand vs 65.90 thousand. F1 vs truth sets backed by ≥2 of 14 callers: 99.16% vs 97.87% (small) and 92.64% vs 88.29% (SV); SV precision drops in tandem repeats. Pantree: 2.6 million SNPs, 2.3 million indels and 80 thousand SVs off reference; GRef: median 143,275 off-reference variants per sample (2.6%). · An Expanded Map of Haplotype Resolved Variation; Mapping pangenome coordinates; Discussion

4. 04

   ShownUsing the HPRC2 graph in short-read analyses cuts small-variant calling errors compared with GRCh38-only pipelines, genotypes more SVs per genome and strengthens eQTL associations.

   DeepVariant and Sentieon on HG002 across four platforms (GIAB v5.0): errors fall 36% to 52.4% overall and 50.1% to 67.7% for SNPs. PanGenie on 3,202 1000G genomes: median 27,428 SVs per sample, 9,339 more than with the HPRC1 graph and 2,742 more than with HGSVC3+HPRC1. eQTLs in 430 samples: pangenome-specific variants are 20.3% of lead markers at 10,169 eQTLs; 893 eGenes (8.8%) gain ≥20% in -log10 p. · Pangenome applications; Pangenome based read mapping improves small…; Pangenome-based variant calling improves trait…

5. 05

   ShownThe assemblies add CpGs absent from the linear reference, and matched long-read methylation, chromatin and RNA data from the cell lines find methylation QTLs, accessible elements and transcripts that a linear reference cannot represent.

   17.6 million CpGs absent from T2T-CHM13 (a 51.9% increase over its 33.8 million). 80,854 promoter mQTLs, with SV lead variants 1.71-fold enriched in centromere/satellite regions (empirical P = 2.0 × 10^-4). Fiber-seq in 38 samples: 51,111 of 568,430 consensus accessible peaks (9.0%) lie in loci absent from GRCh38. Kinnex RNA in 206 samples: 68,681 reference-divergent transcripts across 578 genes. · A matched long-read pantranscriptome resource; Non-reference CpG methylation across the…; Mapping methylation quantitative trait loci…; A paired long-read chromatin epigenome map of…