lifton2: a first-party, refinement-first genome-annotation lift-over engine

LiftOn (Chao et al., 2025) was built on a simple observation: DNA alignment and protein alignment fail in different ways. DNA lift-over with Liftoff (Shumate and Salzberg, 2021) preserves gene structure well when two genomes are close, but introduces frameshifts and premature stops as sequence diverges. Protein-to-genome alignment with miniprot (Li, 2023) survives divergence, but is fragmentary and does not preserve the reference transcript structure. LiftOn fuses the two and keeps, per gene, whichever signal is locally better.

lifton2 is a research re-implementation of that engine. It exists to answer two questions LiftOn’s architecture does not isolate. First: can the Liftoff dependency — which LiftOn ships as a vendored fork — be replaced by first-party code without changing its output? Second: instead of merging two finished gene models after the fact, how far can you get by using the protein evidence to refine the structure of the genes the DNA lift already places?

What lifton2 is

lifton2 has three parts.

The first is a de-vendored DNA-lift core. lifton2 re-implements the minimap2-driven Liftoff lift-over as first-party code, and pins it byte-for-byte to a freshly regenerated Liftoff result with a regression gate: same feature counts, coordinates, and CDS phases, zero differences. The gate is green on five reference genomes spanning plants, insects, and mammals. This makes the de-vendoring a verified equality, not an approximation — and it frees the engine to evolve independently of the upstream fork.

The second is refinement-first chaining. For each transcript the DNA lift places, lifton2 also has miniprot’s protein-to-genome alignment of the same reference protein. It translates both the DNA-lift CDS and the miniprot CDS, aligns each back to the reference protein, and adopts miniprot’s splice structure only when it scores strictly higher. Because the swap is gated on a strict improvement against a fixed reference, it can only raise — never lower — the protein identity of a transcript the DNA lift already placed. This is the opposite of a post-hoc merge: rather than stitching two models together, lifton2 rewrites splice junctions where the protein evidence proves they are wrong.

The third is a miniprot-only rescue pass with a divergence-adaptive floor. Refinement improves genes the DNA lift places; it does nothing for genes it misses. On divergent pairs the DNA lift misses many, so lifton2 runs a separate pass after the main lift that emits a gene model for each reference coding gene with no overlapping lifted locus, provided its protein alignment clears an identity floor. A reference-id dedup guard means the pass only ever adds genes at empty loci — it never displaces or duplicates a lifted model. The floor itself is adaptive: it lowers from 0.5 toward 0.3 as DNA-lift recall drops, because on a eudicot-to-monocot or mammal-to-fish lift, 0.5-identity is the conserved core.

A note on framing, which the rest of this post follows. miniprot is the protein evidence lifton2 fuses — not a rival tool. The honest comparison for a refinement engine is against the DNA baseline it extends: Liftoff. That is the comparison the figures below lead with.

What the benchmark shows

I evaluated lifton2 against Liftoff (the DNA baseline), LiftOn v1.0.8 and LiftOn devel (the dual-evidence pipeline it re-implements), and miniprot (the protein evidence), on a four-tier divergence ladder — same-species, close, distant, and very-distant cross-species lift-overs. Every output is re-scored by one tool-neutral evaluator: for each lifted coding transcript it translates the CDS, aligns it to the reference protein, and computes protein identity. No tool’s self-reported numbers are used.

lifton2 refines the DNA lift, and the gain grows with divergence. Against Liftoff — the DNA baseline it extends, scored on the same metric over the same tool family — lifton2’s mean protein identity is higher on every tier, and the margin widens as the genomes diverge: from a fraction of a percent on same-species data to about +0.21 on the divergent tiers, where the DNA model frameshifts and the protein evidence most often supplies a strictly better splice structure. This is the precision side of lifton2, and it is a guaranteed per-transcript non-regression by construction. lifton2 tracks the dual-evidence LiftOn it re-implements closely — marginally behind on raw identity — while delivering that as a first-party engine with cleaner output and more robust whole-genome behavior, below.

⤢ Enlarge

✕

lifton2 lifts the features the coding-only baseline drops. A first-party DNA lift that stops at protein-coding genes inherits a narrower genome map than the reference provides. lifton2 auto-detects every reference parent type that behaves like a gene — pseudogenes, long non-coding RNA genes, tRNAs, and more — and lifts them in an additive second pass that leaves the coding output byte-for-byte unchanged. Across the dataset ladder this adds thousands of gene-like features (for example, hundreds of pseudogenes on a single Arabidopsis chromosome), raising lifton2’s feature breadth above the coding-only Liftoff on every tier and matching LiftOn’s coverage on the same- and close-species genomes, with zero coding genes lost.

⤢ Enlarge

✕

lifton2 emits standards-clean GFF3. Re-running every tool’s output through one GFF3-specification validator, lifton2 is the only one with zero errors — it repairs the duplicate-exon-ID and out-of-bounds-CDS problems that lift-over chaining tends to introduce, without touching any coding sequence, so protein identity is unchanged. The other tools each carry tens of validation errors per genome.

⤢ Enlarge

✕

lifton2 finishes whole genomes other tools crash on. Running its own de-vendored Liftoff under a recursion-limit guard, lifton2 produced scorable output on every whole genome tested, including extreme-divergence pairs where standalone Liftoff and LiftOn v1.0.8 abort mid-run. Robustness — completing the run at all — is a precondition for every accuracy number, and it is where a carefully re-engineered engine pays off.

⤢ Enlarge

✕

The honest reading of recall. lifton2’s protein-identity lead on the divergent tiers is measured over the genes it recovers, and on those pairs its recovered set is the DNA lift’s set plus the rescue — smaller than what the protein-only miniprot recovers. miniprot searches with proteins directly and is not anchored to a DNA lift, so on the hardest pairs it recovers more genes, at lower average identity. lifton2 is, by design, a high-precision refinement of the DNA-anchored core rather than a recall-maximizing tool; the rescue pass with the adaptive floor narrows that gap at zero redundancy, but does not close it. Which posture you want depends on whether you need the most genes or the most reliable models for the genes you do recover.

What it means

lifton2 makes two points. A lift-over engine’s DNA core can be de-vendored into first-party code without changing its output — the regression gate makes that a verified equality. And a refinement-first posture — rewrite splice structure only when the protein evidence strictly improves it — produces the highest-precision gene models of any tool tested on divergent pairs, on the genes the DNA lift recovers, while keeping the output standards-clean and the engine robust enough to finish whole genomes.

lifton2 is a research engine, open-source at github.com/Kuanhao-Chao/lifton2, built over minimap2, miniprot, and the parasail kernel. It is the first half of a two-part evaluation; the whole-genome benchmark with the same verification protocol is the planned follow-up. The method behind it all is still the original LiftOn idea — DNA plus protein, each catching what the other misses — now with a first-party engine and a discipline that every change prove it did not alter the science.