BGPT: Paper Review: Targeted sequencing and iterative assembly of near-complete genomes

Explore by Goal

Quick Explanation Copied

Paper focus:

Cornetto integrates ONT ReadUntil/ReadFish targeted adaptive sequencing with iterative reassembly to enrich sequencing effort onto the “unsolved” parts of a draft assembly, yielding near-complete diploid genomes (including saliva-derived samples) and resolving extreme repetitive disease loci (FSHD D4Z4 and MUC1 VNTR).

Long Explanation

Targeted sequencing and iterative assembly of near-complete genomes — Visual & Skeptical Review

DOI: 10.1038/s41467-025-65410-x • Received 28 May 2025; Accepted 15 Oct 2025 • Publication: Nature Communications (2025)

Skim-first roadmap

What Cornetto claims (adaptive targeting on unresolved assembly regions)
Key results (hg002 contiguity/phasing; saliva viability; repetitive disease loci)
Mechanistic plausibility (why “reject boring bits” could work; where it could fail)
Critical quality audit (metrics chosen; benchmarking ground truth; sample-size caveats)

1) What Cornetto is (and what it is not)

Core idea: start from a moderate-quality primary assembly, then define genomic regions that are already “confidently resolved” (“boring bits”) and use ONT adaptive/ReadUntil (via ReadFish) to reject reads originating from those regions during subsequent sequencing runs. The sequencing is paused periodically; a new assembly is computed; the “boring bits” list is updated; then sequencing resumes, progressively enriching the remaining unsolved fraction.

Selection mechanism: the study explicitly uses ReadFish for targeted nanopore sequencing.

Assembly evaluation: the pipeline is evaluated with widely used assembly benchmarks: Nx, telomere/L90 chromosome completeness, BUSCO (via Compleasm), QV and sequence error rates (via Yak), and genome-aligned contiguity metrics (NGA via QUAST), plus assembly-guided SV evaluation (dipcall + truvari).

2) Results audit: hg002 contiguity, diploidy, and accuracy

Contiguity gains with Cornetto (primary assemblies): The excerpt gives a detailed numeric example for hg002-Cornetto-1: final primary assembly with 130 contigs, N50 = 134.1 Mb, and N90 = 50.4 Mb, and chromosome-scale contiguity improved from none in the starting assemblies to 15 chromosomes as single primary contigs.

ONT-only variant: for hg002-Cornetto-3 (ONT flow cell simplex base + one Cornetto run with 3× cycles), the excerpt reports N50 = 154.4 Mb and N90 = 79.1 Mb.

Diploid (partially phased vs fully phased with Hi-C): The study reports that hg002-Cornetto-4 (partially phased diploid via exclusion of unphased regions from boring-bits) contains 33 partially-phased T2T chromosomes out of 46, and hg002-Cornetto-4-HiC (fully phased with Hi-C) contains 32/46 chromosomes as fully-phased T2T contigs or scaffolds.

Accuracy comparisons (against a fully phased ONT assembly): The excerpt says hg002-Cornetto-4 is comparable/superior on many standard metrics but shows higher hamming errors than the published T2TC-ONT fully phased assembly; however, hg002-Cornetto-4-HiC is reported as more complete/accurate/contiguous than T2TC-ONT across most metrics, with only a small metric where T2TC-ONT outperforms.

3) Mechanistic plausibility (and skeptical failure modes)

Why this could work: assembly uncertainty is commonly concentrated in (i) low-mappability/repetitive regions, (ii) contig ends/locally broken structure, and (iii) segments with poor coverage/quality. If the method can reduce wasteful reads from already-resolved regions, the effective depth and coverage uniformity over the remaining unsolved fraction should improve, which can increase contiguity. Cornetto’s boring-bit definitions explicitly include short contigs, contig-end-adjacent windows, coverage/mappability outliers, and low-quality regions from hifiasm output.

Where it could fail (important):

Initial assembly bias: if the starting “solved” regions are misclassified (e.g., misassemblies or under-resolved repeats), rejecting reads there could lock in errors. Cornetto tries to mitigate by using a moderate-quality base assembly and multiple signals, but the approach is still bootstrapped on the correctness of that base.
Read mapping / target coordinate errors: the rejection decision is alignment-based in real time. If mapping thresholds or coordinate buffers cause systematic off-target rejection/acceptance, the enrichment could be distorted. (The excerpt describes realignment and the boring-bits update process; it does not quantify the sensitivity to coordinate/threshold perturbations in the provided text.)
Iterative pause-point lag: the study reports reassembly at experimental pause points (hours-scale), not continuous real-time updating; this can limit how quickly “still unsolved” regions are targeted as assembly improves.
Phasing limits for non-Hi-C samples: saliva assemblies are reported as sometimes only partially phased; fully phased chromosome-scale contigs/scaffolds required Hi-C integration.

4) Clinically/biologically motivated validation: D4Z4 (FSHD) and MUC1 VNTR (ADTKD)

FSHD D4Z4 macrosatellite repeat (4q): The excerpt describes extracting sub-telomeric chr4q regions from both haplotypes in hg002-Cornetto-4 and comparing to the T2T-hg002 ground truth. It reports identifying D4Z4 haplotypes at 42 copies (4qA) (~139 kb) and 26 copies (4qB) (~86 kb), with 99.99% and 99.97% sequence concordance across entire D4Z4 regions (differences concentrated in small indels).

MUC1 VNTR (ADTKD): The excerpt reports that hg002-Cornetto-4 contains two VNTR haplotypes: 65 × 60 bp copies (~3.9 kb) and 78 × 60 bp copies (~4.7 kb) with perfect sequence concordance when aligned to T2T-hg002. In ten diagnosed MUC1-ADTKD patients (plus hg002), the paper states 20 unique VNTR haplotypes ranging 40–83 copies, with a dupC frameshift variant detected on a single haplotype in each patient and observed to be frequent across independently formed haplotypes (8/10 pathogenic haplotypes unique).

5) Resource/accessibility claim: saliva and non-human vertebrates

Saliva viability: saliva is described as more convenient for collection/storage but less amenable to HMW extraction and long-range methods, and often contaminated with non-human DNA. Cornetto is presented as enabling high-quality assemblies from saliva, including ONT-only diploid assemblies with reported contig contiguity improvements (e.g., N90 values and complete chromosome counts for saliva-A-Cornetto-2 and saliva-B-Cornetto-2).

Non-human vertebrates: the method is argued to be sequence-agnostic and applied to a set of species spanning diverse lineages: orange-bellied parrot, western saw-shelled turtle, Gould’s petrel, and a redstriped eartheater cichlid, with reported improvements in assembly contiguity and chromosome completeness for multiple cases.

6) Reproducibility & transparency check

Data availability: raw sequencing data for animals and reference samples are deposited in ENA under accession PRJEB86853; genome assemblies are available on Dryad at DOI 10.5061/dryad.kkwh70sfr; human participant data are restricted.

Code availability: Cornetto software is open source on GitHub and original code is deposited at Zenodo.

Key technical dependency points: Reproducibility likely depends on (i) ReadFish/ReadUntil configuration templates, (ii) base assembly quality and the boring-bits coordinate generation rules, (iii) pause-point reassembly time, and (iv) exact assembly/evaluation tool versions. The excerpt indicates wrapper scripts and versioned tooling, but the full reproducibility score cannot be fully audited from this text-only excerpt alone.

BGPT science actions (bespoke follow-ups)

Author reviews on BGPT

If any author name is missing or mis-parsed from the provided text, re-run BGPT’s author-list extraction for full accuracy.

Feedback:

Updated: March 30, 2026