Building complete and diverse reference genomes

Assembly reconstructs chromosomes from reads, and the regions that matter most are the hardest to place: segmental duplications, centromeres, ribosomal DNA arrays, and structurally variable loci. Accurate long reads and graph-based assemblers such as hifiasm (Cheng et al., 2021) changed what is possible. The Telomere-to-Telomere Consortium completed a human genome end to end (Nurk et al., 2022), and Verkko (Rautiainen et al., 2023) automated telomere-to-telomere assembly of both haplotypes.

During my Ph.D., I worked on what comes after a complete assembly: turning it into a reference others can use. That takes provenance, annotation, and diversity as much as contiguity.

Questions that shaped this work

What does “complete” mean? A gapless assembly can still contain sequence that did not come from the individual. A useful reference says which bases were assembled from that person’s reads and which were borrowed.

Whose genomes become references? GRCh38 is a mosaic of several individuals, and T2T-CHM13 is one genome. Neither represents human diversity, which is why the Human Pangenome Reference (Liao et al., 2023) assembles many haplotype-resolved genomes.

Why must sequence and annotation advance together? Without genes, two assemblies are hard to compare biologically; with a careless annotation, assembly errors become apparent gene differences.

What I built

Han1: a complete Southern Han Chinese genome

Han1 (Chao et al., 2023) was my first Ph.D. project. After T2T-CHM13, Steven Salzberg proposed completing a genome from another ancestry, and I chose HG00621, a Southern Han Chinese individual with deep HiFi and ultralong Nanopore data. We built a hifiasm backbone, scaffolded it to chromosomes, closed gaps with HiFi and Nanopore contigs, resolved the last gaps by hand, and polished the result. Han1 is 3.1 Gb in 25 sequences, every chromosome a single gapless contig. About 120 Mb in the hardest repeat-rich regions was filled with T2T-CHM13 sequence, so Han1 is mostly, but not purely, HG00621.

We lifted over 60,708 genes, 20,003 of them protein-coding, with a two-pass Liftoff workflow that handled ribosomal DNA arrays separately. Gene by gene, Han1 and T2T-CHM13 are largely collinear. Of 235 protein-coding genes with major predicted coding differences, 46 carried homozygous non-SNP coding changes once heterozygous calls were excluded; the two genomes also differ by a 4.1 Mb inversion over the chromosome 8 β-defensin cluster and by copy number in 94 near-identical gene clusters.

Annotating a complete diploid genome

The complete diploid HG002 benchmark (Hansen et al., 2026) represents the maternal and paternal haplotypes separately. I joined the annotation effort, whose pipeline extended the two-pass approach from Han1, and contributed an extra-copy search that recovered 51 gene copies across the two haplotypes. A diploid reference keeps differences that a single consensus erases: GSTT1 appears only on the maternal haplotype and GSTM1 only on the paternal.

What remains open

One genome is not a reference for everyone. The open problem is making many complete, haplotype-resolved genomes comparable, each with provenance precise enough to say where every base came from and each annotated on both haplotypes rather than one.