Genome Assembly Metrics That Mislead: Why a High N50 Can Coexist With Structural Misassemblies

Zetobit · Bioinformatics Insight Series

Genome Assembly Metrics That Mislead: Why a High N50 Can Coexist With Structural Misassemblies

N50 is the number everyone quotes to prove an assembly is good: the bigger it is, the more contiguous the genome. But N50 measures only how long the pieces are — not whether they're joined correctly. And the fastest way to raise it is to make exactly the kind of aggressive join that produces a chimera.

By Kanna Nandakumar, PhD Zetobit LLC

Assemble a genome and the first number anyone asks for is N50 — the length such that contigs (or scaffolds) of that size or longer make up half the total assembly.¹ It's a natural shorthand for contiguity: a higher N50 means fewer, longer pieces, and an assembly in a handful of chromosome-scale scaffolds feels far closer to "done" than one in thousands of fragments. In the long-read era a contig N50 above 1 Mb is generally considered good, and the number gets top billing in nearly every assembly paper and deliverables table.²

The problem is that N50 answers exactly one question — how long are the pieces? — and is routinely read as if it answered a different one: is the genome right? Those are not the same question, and the gap between them is where structural misassemblies live, invisible to the metric that's supposed to signal quality.

The three Cs, and why N50 only covers one

Assembly quality has three distinct dimensions, often called the three Cs: contiguity (how fragmented the assembly is), completeness (how much of the genome and its gene content is present), and correctness (whether the sequence is actually assembled in the right order and orientation, free of structural errors).³ The most commonly reported metrics cover only two of the three: N50 for contiguity, BUSCO for gene-space completeness — and correctness, the one that determines whether the assembly is true, is frequently the one left out.³

These dimensions do not move together. A meta-analysis of 611 eukaryotic genome papers found that while high-N50 assemblies reliably had high BUSCO completeness, a high BUSCO score could also come from an assembly with a low N50 — the metrics measure genuinely different things and don't reduce to one another.⁴ Contiguity and completeness can each look excellent while correctness quietly fails, because none of the popular single numbers is designed to detect a join in the wrong place.

How a misassembly makes N50 go up

Here is the mechanism that turns N50 from merely incomplete into actively misleading. A misjoin — stitching together two sequences that don't actually belong adjacent — produces a longer contig than either input piece.⁵ Because N50 is a length statistic, that erroneously long contig raises N50. The structural error and the metric improvement are the same event. An assembler can be tuned to join more aggressively, and doing so reliably increases N50 while introducing more misassemblies; a wrongly fused genome full of chimeras can post a better N50 than a correct but more fragmented one.⁶

This creates a genuine and well-documented tradeoff: at repeat regions too long for a read to span, an assembler must guess how to extend, and aggressive assemblers win contiguity at the cost of correctness, while conservative ones stop and stay fragmented but accurate.⁷ It also creates a perverse incentive — a developer or lab can make an assembler look superior simply by making it more prone to join, attracting users pleased with higher N50 even though the assemblies are more chimeric.⁶ The number that's supposed to certify quality can be optimized directly at quality's expense.

Assembly 1 is fragmented but every contig is correct. Assembly 2 fused pieces of two different chromosomes into one long contig — a misjoin — which gives it the higher N50. Ranked by N50 alone, the wrong assembly wins.

Reference-based correctness metrics — and their own limits

The standard fix, when a reference is available, is to stop counting raw contig length and count aligned block length instead. Metrics like NGA50 and NA50 (from QUAST) align contigs to a reference, break them at misassembly points, and recompute N50 on the resulting fragments — so a misjoined contig is split back apart and no longer earns its inflated length.⁸ The gap between NG50 and NGA50 is itself a signal: the larger the drop, the more misassembly the raw contiguity number was hiding.

But these corrected metrics carry caveats. They require a high-quality reference of the same organism — often unavailable for the novel genomes assembly is most used for — and they penalize genuine structural differences between your sample and the reference as if they were errors, so NGA50 is a lower bound and true contiguity sits somewhere between it and NG50.⁹ Even a corrected N50 isn't fully immune: a few misassemblies inside large contigs can still inflate it.⁵ Correctness assessment is better than raw N50, not a solved problem.

Reference-free doesn't mean structure-free. Tools like Merqury evaluate assemblies without a reference by comparing k-mer content between the reads and the assembly, giving excellent base-level accuracy (QV) and completeness estimates.¹⁰ But k-mer consensus accuracy is not structural accuracy: a high QV says the bases are right, not that they're in the right order — and some structural misassemblies, such as inversions, can go unnoticed by these methods.¹⁰ An assembly can be simultaneously high-N50, high-BUSCO, high-QV, and structurally wrong.

What actually reveals correctness

Because no single number captures correctness, verifying it means triangulating with evidence the contiguity metric can't see:

• Report all three Cs, not one. Pair N50 with a completeness metric (BUSCO/gene-space) and an explicit correctness assessment; treat agreement among them as the signal, disagreement as a flag.³
• Show the corrected contiguity. When a reference exists, report NGA50/NA50 alongside NG50; a large NG50→NGA50 drop reveals hidden misassembly.⁸
• Map the reads back to the assembly. Consistent coverage and read-pair orientation across a join support it; discordant long-range evidence at a junction is a misjoin signature that summary metrics never show.⁷
• Inspect the assembly graph and use orthogonal long-range data. Ambiguous graph structure flags repeat-driven uncertainty, and independent data (Hi-C contact maps, optical maps, linked reads) validates or corrects chromosome-scale order rather than assuming it.⁷
• Match validation to the failure mode. A suspiciously long scaffold calls for structural validation, not more aggressive scaffolding; repeat-rich regions call for longer-range spanning evidence.⁷

As with the other blind spots in this series, the trustworthy assembly reports what a single number hides: contiguity and completeness and correctness, how correctness was assessed, and where structural confidence is weakest. An assembly summarized by N50 alone looks finished, but the one statistic offered is precisely the one that a misassembly can improve.

The takeaway

N50 is a useful contiguity statistic and a terrible correctness statistic, and the trouble is that it's routinely asked to be both. Because a misjoin produces a longer contig, structural errors can raise the very number meant to signal quality — so a more chimeric assembly can outscore a correct one, and an assembler can be tuned to exploit exactly that. Contiguity, completeness, and correctness are three different questions; the popular metrics answer the first two and too often leave the third unasked. The discipline is to stop treating a big N50 as a finished genome and to demand the evidence contiguity can't provide: that the long pieces are not just long, but right.

References

1. Assembly metrics that matter — N50 defined; N50 measures contiguity only; high N50 can coexist with misassemblies. CD Genomics. cd-genomics.com
2. Beyond contiguity — the 3 Cs; contig N50 >1 Mb generally considered good in long-read era; BUSCO complete >95% good; common metrics tackle only two of three Cs. PacBio. 2022. pacb.com
3. From contiguity to accuracy — 3C criteria (contiguity, completeness, correctness); N50 does not directly reflect accuracy; high-N50 assemblies can contain misassemblies/collapsed repeats. J Microbiol. 2026. jmicrobiol.or.kr
4. Assessing genome assembly quality: N50 versus BUSCO — meta-analysis of 611 eukaryotic genomes; high BUSCO obtainable even from low-N50 assemblies; metrics measure different dimensions. Mol Biol Evol. 2021. PMID 33629477
5. PDR metric — misassembly (misjoin/structural error) produces larger contigs and thus inflated NG50; even corrected N50 can be inflated by a few misassemblies in large contigs. Bioinformatics. 2021. academic.oup.com
6. The N50 misassembly problem — making an assembler more prone to join raises N50 but yields chimeras; a wrongly assembled genome is not better than a fragmented correct one. The Molecular Ecologist. 2017. molecularecologist.com
7. De novo assembly guide — contiguity/correctness tradeoff at repeats; chimeric misjoin increases apparent continuity while reducing structural truth; validate against failure modes (read mapping, graph inspection, Hi-C). CD Genomics. cd-genomics.com
8. Solutions to the N50 misassembly problem — NGA50/NA50 (QUAST) align contigs to reference and break at misassembly points before recomputing N50. The Molecular Ecologist / Gurevich et al. 2013. molecularecologist.com
9. Tigmint — NGA50 summarizes contiguity and correctness but penalizes true sample-vs-reference structural variation; true contiguity lies between NGA50 (lower bound) and NG50 (upper bound). BMC Bioinformatics. 2018. PMC6204047
10. Rhie et al. Merqury — reference-free k-mer evaluation of base accuracy (QV) and completeness; does not directly validate structural accuracy and some misassemblies (e.g. inversions) can go unnoticed. Genome Biol. 2020. via CD Genomics

Zetobit LLC · Bioinformatics Insight Series · Lexington, KY · zetobit.com