Building accurate genome annotations

A genome assembly provides sequence; an annotation provides the biological map that makes the sequence useful. It identifies genes, transcripts, exons, coding regions, and their relationships. Resources such as NCBI RefSeq (O’Leary et al., 2016) provide annotations for widely studied organisms, but assemblies are produced faster than annotations can be curated.

One response is annotation lift-over: transfer a trusted annotation from a reference genome to a new assembly. I work on this problem through LiftOn, an algorithm that combines DNA and protein alignments. DNA preserves locus and transcript structure between related genomes; protein conservation helps recover coding structure when nucleotide sequences diverge. LiftOn turns those complementary signals into one annotation rather than treating either alignment as sufficient on its own.

Questions that drive my work

The first question is how to combine complementary evidence explicitly. Liftoff (Shumate and Salzberg, 2021) maps gene models through DNA alignment and preserves their hierarchy. Protein-to-genome aligners such as miniprot (Li, 2023) remain informative across greater evolutionary distance, but they emit coding structure alone — no UTRs, no non-coding features — and their models can be fragmented. A useful algorithm needs a clear rule for choosing supported coding segments while retaining the broader annotation structure.

The second is what the algorithm should optimize and validate. A transferred model may map to the expected locus yet contain a frameshift, premature stop, or incorrect coding boundary. Protein identity, open-reading-frame integrity, overlap resolution, and gene-copy recovery measure different aspects of quality. The objective should be explicit, and the emitted GFF should remain structurally valid.

The third is scale and reproducibility. A method must finish whole genomes, handle unusually large and repetitive genes, preserve non-coding features, and produce deterministic output. Accuracy on a small benchmark is not enough if the implementation cannot complete the assemblies researchers need to annotate. General annotation systems such as BRAKER3 (Gabriel et al., 2024) integrate experimental and comparative evidence for de novo annotation; lift-over addresses the complementary case where a trusted reference annotation already exists.

Directions I have explored

LiftOn: combining DNA and protein alignments

I developed LiftOn (Chao et al., 2025) to improve annotation lift-over across both assemblies and species. It starts with DNA-based models from Liftoff and protein-based candidates from miniprot. LiftOn translates each candidate, aligns it with the reference protein, and uses a protein-maximization chaining algorithm to select the better-supported coding segments. An open-reading-frame search can then repair remaining coding disruptions. The algorithm also resolves overlapping loci and searches for additional gene copies that a one-to-one transfer can miss.

LiftOn does not average DNA and protein evidence. It preserves their distinct strengths and reconciles disagreements through a protein-centered objective. In the published evaluation, LiftOn transferred annotations within species and across more divergent species pairs, where DNA-only mapping becomes less reliable. My continuing work focuses on whole-genome robustness, bounded-memory alignment, output validation, and broader feature support so the same algorithmic principle remains practical at scale.

Applying lift-over to complete human genomes

Annotating complete human genomes is what led me to LiftOn, and it is where LiftOn is now used. For CHESS 3 (Varabyou et al., 2023), my contribution focused on lifting the CHESS 3 annotation from GRCh38 to T2T-CHM13. In the earlier Han1 project (Chao et al., 2023), we used Liftoff in a two-stage workflow to transfer annotation to a gapless Southern Han Chinese genome, with separate handling for repetitive ribosomal DNA arrays. Han1 predates LiftOn, but it exposed the practical demands of annotating a newly completed genome.

The complete diploid HG002 benchmark (Hansen et al., 2026) provided a later application. The project generated annotations for the maternal and paternal haplotypes, and my contribution was to that effort: LiftOn, together with miniprot, helped identify additional gene copies from translated MANE transcripts. This is the setting LiftOn is designed for: carrying trusted gene knowledge onto a more complete genome while preserving differences that may represent real copy-number variation.

Future directions

LiftOn’s remaining work is as much engineering as algorithm design: bounded-memory alignment across whole genomes, support for every feature type a full annotation carries, and a validation trail that records why each correction was made. What that buys is scientific. As complete genomes and pangenomes multiply, a lift-over that can be trusted at genome scale is what decides whether a gene-content difference between two individuals is real biology or an artifact of the transfer.