Size dependent loss of structural variant allele counts in pangenome derived VCFs

Abstract Background. Pangenome graphs represent structural variation as divergences between haplotype paths, but almost all downstream analysis consumes VCF, so graphs are projected onto a linear reference by a standard conversion chain. Whether structural variant allele counts survive that projection is unknown, because real data provides no ground truth. Results. We simulated haplotype panels on real GRCh38 sequence with known carriers for every structural variant, giving 4,056 variants across two chromosomes. Recovery of the true allele count falls with variant size, from 0.988 at 100 bp to 0.553 at 30 kb, with near identical coefficients on log₁₀ size on the discovery and held out chromosomes (− 2.173 and − 2.183). Recovery declines further with segmental duplication content, to 0.375 for 30 kb variants. The loss arises in bubble decomposition rather than post-processing: vcfbub changes nothing, and vcfwave raises readable recovery from 0.466 to 0.865 overall yet only to 0.553 at 30 kb. Graphs built from identical panels by PGGB and Minigraph-Cactus behave alike. Computing allele counts from haplotype path traversals instead recovers 0.992 overall and 0.986 at 30 kb, independent of size. In the HPRC Release 2 pangenome, a 62 kb deletion carried by 216 of 461 haplotypes appears as 84 alleles whose largest count is 25. Conclusions. Allele counts of large structural variants, especially in segmental duplications, are fragmented when pangenome graphs are decomposed into VCF records. The effect belongs to the representation, not to any one graph builder, and counting alleles from haplotype path traversals recovers what the conversion loses.