Building accurate genome annotations

An assembly gives a genome’s sequence; an annotation says what that sequence encodes. Assemblies now arrive far faster than curated annotations such as NCBI RefSeq (O’Leary et al., 2016) can follow, and the newest are complete or haplotype-resolved. Lift-over closes part of the gap by carrying a trusted annotation from a reference onto a new assembly.

The difficulty is that a transferred gene can land in the right place and still be wrong: a shifted splice site, a frameshift, a premature stop, a missing copy. I work on lift-over that checks what it transfers against the protein it is supposed to encode.

Questions I’m pursuing

How should DNA and protein evidence be combined? Liftoff (Shumate and Salzberg, 2021) maps whole gene models by DNA alignment and keeps their structure, but it degrades as sequences diverge. Protein aligners such as miniprot (Li, 2023) reach farther yet emit coding exons only, with no UTRs or noncoding genes. The two fail in different places, which is what makes them worth combining.

What counts as a correct gene model? Mapping to the expected locus is not enough. I judge a transferred model by the protein it encodes, the integrity of its reading frame, and whether the copies present in the target genome are found.

Does it hold at genome scale? A method is useful only if it finishes whole genomes, including very large and repetitive genes, and produces valid, reproducible output. De novo pipelines such as BRAKER3 (Gabriel et al., 2024) serve genomes without a close reference; lift-over serves the many that have one.

What I have built

LiftOn

LiftOn (Chao et al., 2025) runs Liftoff and miniprot, pairs their models locus by locus, and aligns both translations to the reference protein. A protein-maximization chaining algorithm splits the two models where they cover the same span of the reference protein and keeps, for each segment, the version with higher identity. An open-reading-frame search repairs remaining frameshifts and premature stops, and a copy search adds gene copies that one-to-one transfer misses; noncoding features stay on the DNA path. Lifting RefSeq from GRCh38 to T2T-CHM13, LiftOn mapped 99.6% of genes, improved 866 protein-coding transcripts over Liftoff, and found extra copies of 86 genes at 320 new loci. The largest gains came in the more divergent pairs: from mouse to rat, it improved 15,420 transcripts over Liftoff.

Annotating complete human genomes

Complete genomes test lift-over hardest. Annotating Han1 (Chao et al., 2023) was one of the threads that led me to LiftOn, and for CHESS 3 (Varabyou et al., 2023) I lifted the catalog from GRCh38 to T2T-CHM13. For the complete diploid HG002 benchmark (Hansen et al., 2026), my contribution was an extra-copy search: aligning MANE proteins to each haplotype with miniprot and filtering them with LiftOn’s copy logic recovered 51 gene copies (30 maternal, 21 paternal) that DNA-based lift-over had missed.

Future directions

As complete genomes and pangenomes multiply, the interesting questions compare individuals: which genes one person carries and another lacks, and which copies differ. Lift-over will produce most of those annotations, so it has to make its errors visible. I want transfer methods that keep working as evolutionary distance grows, record the evidence behind each change, and let a reader tell real gene-content variation from an artifact of the transfer.