I work on the parts of a genome a single reference cannot represent.

I am Liu Yang, a postdoctoral researcher in the Department of Animal and Avian Sciences at the University of Maryland. From 2023 to 2025 I held a joint appointment at the Animal Genomics and Improvement Laboratory at USDA-ARS in Beltsville, where most of the cattle pangenome work was done. My work sits between three things that are usually done by different people: assembling and graphing genomes, genotyping variants across large cohorts, and mapping those variants to traits.

Most of what I do concerns structural variants — insertions, deletions, inversions and more complicated rearrangements above about fifty base pairs. They are a large fraction of the sequence that differs between two animals, they are enriched exactly where short reads fail, and until recently they were mostly excluded from the analyses that decide which bulls get used. Pangenome graphs changed that, and a good part of my work has been finding out how much they changed it in practice rather than in principle.

Three questions I keep returning to

How many haplotypes do you actually need? A breed-specific graph built from phased assemblies of twenty animals recovers over ten thousand more structural variants per sample than read-based long-read calling on the same data. A graph built across breeds, or from fewer or unphased assemblies, does measurably worse. But the curve saturates, and knowing roughly where it saturates is the difference between a tractable project and an expensive one.

How do you get from a graph to fifty thousand animals? Almost no population of interest has long reads. The route that works is graph-based genotyping on the few thousand samples that have short-read whole genomes, then imputation from SNP array data onto everything else. In dairy cattle roughly 87% of structural variants come through that second step at Beagle R² ≥ 0.8 — which also means 13% do not, and being honest about which ones is part of the job.

Do the variants explain anything? Structural variants show stronger genome-wide enrichment than SNPs across most complex traits we have looked at, which is a real result. They also add only a fraction of a percent to the genetic variance explained, which is also a real result. Both belong in the same sentence.

The short record

• PhD, 2021Animal genetics, breeding and reproduction, Sichuan Agricultural University. Postdoc at the Agricultural Genomics Institute at Shenzhen, CAAS, 2021–2023, before Maryland.
• 38 papersFifteen as first, co-first or corresponding author, including Nature Communications and Genome Biology.
• Principal investigatorNSFC Youth Science Fund 32302699, on a transposable element affecting reproductive traits in European pigs (2024–2026).
• Lead writerNRSP-8 Animal Breeding RCN fellowship proposal, 2025. Contributor on USDA NIFA and U.S.–Israel BARD applications.
• ReviewerFor Nature Communications, Advanced Science, Genetics Selection Evolution, Animal Genetics and others.
• Invited talkAuburn University, Department of Pathobiology seminar, March 2026, on phased-assembly-driven pangenome graphs.

About the name

This site is under my own name, Liu Yang — that is where the papers, author lists and consortium contributions live, and always will. OXGGI is the name I use for tools, writing and code meant to outlive any one position. The two are linked so neither has to be looked up separately.