De Novo Mutation Calling

Zetobit · Bioinformatics Insight Series Rare Variants & Families

Bioinformatics Insight Series

De Novo Mutation Calling

Why the Confirmation Step Does More Work Than the Caller

Kanna Nandakumar, PhD · Zetobit, LLC

De novo mutation calling has an arithmetic problem before it has a software problem. The event being detected occurs at roughly one in 10⁸ positions per generation. Sequencing, mapping and genotyping errors occur at rates several orders of magnitude higher. The signal is not merely buried; it is outnumbered by something like four orders of magnitude¹.

The rule for finding a de novo mutation is trivial to state: a variant present in the child and absent from both parents. Applied literally to a trio VCF, it returns a list dominated almost entirely by artifacts. Every subsequent step — depth filters, allele-balance filters, mappability masks, orthogonal confirmation — exists to reduce that list, and those steps, not the caller, determine what the final callset contains.

This matters because the callset is used for two very different purposes. In research, it estimates a mutation rate. In the clinic, it identifies the variant a family will be counselled about. Both are sensitive to the filtering, and in different ways.

The scale of the mismatch

The human germline mutation rate is approximately 1.2 × 10⁻⁸ per base per generation, giving on the order of 40 to 80 de novo single-nucleotide variants per child depending on the study and parental ages¹. A recent four-generation pedigree reference, assembled with five complementary sequencing technologies, put the total higher once repeats are included: 98 to 206 DNMs per transmission, of which 74.5 are de novo SNVs, 7.4 are non-tandem-repeat indels, 65.3 arise from tandem repeats and 4.4 are centromeric².

That last breakdown is worth pausing on: more than half the events are in tandem repeats, and the same study found the DNM rate varies by an order of magnitude with repeat content, length and sequence identity². Most published DNM counts describe SNVs in unique sequence, because that is the part short reads can adjudicate.

Against those few dozen true events, the raw Mendelian-inconsistency rate — sites where the child's genotype is incompatible with the parents' — sits about four orders of magnitude higher, so nearly all Mendelian inconsistencies are attributable to sequencing error or chromosomal anomalies rather than mutation¹.

A caller that finds every true de novo mutation and nothing else does not exist. What exists is a caller that finds them alongside ten thousand times as many artifacts, and a filtering stack that removes most but not all of both.

The filters are the method

Because the discriminating work happens after the caller, the filter parameters are the scientific parameters. This has been measured directly, and the result is uncomfortable.

In a coordinated exercise where multiple groups estimated the germline mutation rate from the same sequencing data, different filtering strategies produced a twofold variation in the resulting mutation-rate estimate³⁴. Same reads, same family, same question — a factor of two, from the choices made downstream of variant calling.

Two mechanisms drive that. The obvious one is the numerator: stricter filters remove more artifacts and also more true mutations. The less obvious one is the denominator. A mutation rate is events divided by callable genome, and the callable genome must itself be corrected for the false-negative rate of the very filters being applied — a quantity each group estimates separately⁴. Tighten the filters and both terms move, in the same direction, by amounts that must be estimated rather than observed.

The trade-off has been quantified where PCR-verified truth was available. Comparing seven approaches on family trios, the top-ranked method still showed a 4% false-positive rate alongside a 12% false-negative rate; across the remaining methods, false positives ran 4–12% and false negatives 8–19%⁵. The same study found that taking the intersection of candidates across all approaches yielded a set with no false positives, while any candidate called by only one or two approaches out of seven was a false positive⁵.

That is a useful heuristic and an uncomfortable one. Agreement across independent methods is strong evidence — and singleton calls, which is where novel biology would appear if it appeared anywhere, are where the artifacts live.

Why confirmation is load-bearing

In most fields, orthogonal validation is a check on a result. Here it is closer to being the result.

Published DNM sets are typically defined by what survived confirmation: a genotyping assay validating 82% of candidates in one study⁶, Sanger or amplicon confirmation of every candidate in others, and in the four-generation reference, refinement of putative DNMs through orthogonal sequencing and transmission². Transmission is the strongest evidence available: a variant that appears in the child and is then passed to a grandchild was in the germline, and no filter argument is needed.

The practical consequence is that a DNM callset without a confirmation step is a different kind of object from one with it, even if both were produced by the same software. When reading a paper, the validation method and the validated fraction tell you more about the callset's reliability than the caller's name does.

The machine-learning approaches now used in this space make the dependency explicit rather than removing it: candidates are called with deliberately permissive filters, a subset is validated experimentally, and that validated subset trains a classifier to score the remainder³. The confirmation step is not an audit of the pipeline. It is an input to it.

The third category

So far this is a signal-to-noise problem, and a candidate is either real or an artifact. There is a third possibility, and it is the one with clinical consequences.

An earlier piece in this series dealt with mosaicism as a detection problem in a single sample: given a low-allele-fraction variant, is it a genuine mosaic or an artifact? Here the question has a different shape. The variant in the child is unambiguous — a clean heterozygote at full allele fraction. What is uncertain is the parent. A mutation that arose during a parent's early development, or in their germline, can be present in a fraction of that parent's cells and absent from the reads that a standard-depth trio genome produces.

The trio genotypes look identical to a true de novo event. The recurrence risk does not.

Figure 1. The three-way classification. The single-read calculation in the first panel is illustrative arithmetic, not a study result; the allele fractions and tissue findings in the other two are drawn from the cited work.

How common, and how would you know

The most informative measurement comes from a study that phased de novo variants genome-wide and then went looking for them in the parents at extreme depth. Across five trios with an average of 85.6 DNMs each and the expected 80% paternal bias, targeted resequencing of parental blood and sperm at over 5,000× found 20 of 334 variants (6.0%) showing parental mosaicism, at allele fractions from 0.24% to 14.7%⁷.

Three details in that result do most of the work.

Of the 20 mosaics, 13 were detectable in blood and seven were found only in sperm⁷. Sampling the accessible tissue is not the same as sampling the transmitting one, and a negative result in parental blood does not exclude germline mosaicism.

Maternally phased variants were 3.4 times more likely to be mosaic in blood, a consequence of the paternal bias in DNM origin overall⁷. And of 244 paternally phased variants, 14 (5.7%) showed detectable sperm mosaicism while 230 showed none — which meant sperm sequencing could quantify the recurrence risk of most individual DNMs rather than assigning a population average⁷.

Most striking: every child in that study carried at least one DNM detectable in a parental sample⁷. Not every child had a clinically relevant one — these were not selected for disease — but the phenomenon is not rare at the genome level. It is rare at the level of any particular variant.

Published rates vary widely, and the variation is instructive. Studies report parental mosaicism in roughly 3.0% of neurodevelopmental-disorder trios, 3.7% in one deep-sequencing series, 4.4% in another, 6.0% genome-wide as above, and 6.6–10.2% in cohorts enriched for epileptic encephalopathies or sampling multiple tissues⁷⁸⁹. The spread tracks sequencing depth, which tissues were sampled, and cohort composition. One of those studies used 7,000× smMIP sequencing of parental blood across 189 NDD patients and found 3.7% of families mosaic, with four of the seven at allele fractions below 1%⁸ — a rate the authors explicitly flag as lower than others, raising the question of whether a biological mechanism explains the higher rates in epilepsy-focused series⁸.

What each result actually licenses

Evidence	Supports	Does not support
Trio genotypes: 0/0, 0/0, 0/1	A candidate worth investigating	That the variant is real
Passes depth, allele-balance, mappability filters	Enrichment for true events	A per-variant probability
Orthogonal confirmation in the child	The variant exists in the child	That it arose de novo
Absent from parents at 30–40×	Not a high-fraction parental mosaic	Absence of low-fraction mosaicism
Absent from parental blood at >5,000×	Strong evidence against blood mosaicism	Absence of germline-restricted mosaicism7
Absent from paternal sperm at depth	A quantified low recurrence risk7	Anything about maternally phased variants

Each row narrows the possibilities without closing them. The clinically important asymmetry is in the last three: negative results at increasing depth are increasingly informative, but none of them reaches zero, and the tissue that determines transmission is the one hardest to sample.

What to do about it

1. Treat the raw trio-inconsistency list as a screening output, not a result. At a ratio of roughly ten thousand to one¹, an unfiltered Mendelian-violation list is closer to a list of error-prone sites than a list of mutations.
2. Report the filter parameters as method, not as configuration. Filtering choices alone moved a mutation-rate estimate twofold on identical data³.
3. Report the denominator and how it was corrected. A rate needs a callable genome adjusted for the false-negative rate of the filters used⁴; without that, two rates are not comparable.
4. Confirm orthogonally and state the validated fraction. It is the most informative single number about a DNM callset, and more informative than the caller's identity.
5. Use agreement across methods as evidence, and treat singletons with suspicion⁵ — while remembering this biases against anything unusual.
6. Never report a clinically relevant de novo finding without examining the parental reads directly. A genotype of 0/0 in a parent can rest on a single low-quality read; look at the alignment rather than the call.
7. Where recurrence risk is the question, sequence the parents deeply at that position. Targeted deep sequencing detects mosaicism at fractions standard trio depth cannot reach⁷⁸, and this is a per-family decision rather than a pipeline one.
8. Where paternal transmission is at issue and sperm is obtainable, test it. Some mosaics appear in sperm and nowhere else⁷, and a negative blood result does not cover that case.
9. Phrase the counselling claim to match the evidence. “No mosaicism detected in parental blood at this depth, which reduces but does not exclude germline mosaicism” is a different statement from “de novo, recurrence risk approximately 1%” — and only one of them is supported by a standard trio.

The shape of the error

Two distinct failures share this territory, and they pull in opposite directions.

The first is ordinary and well managed: artifacts vastly outnumber true events, and a stack of filters plus confirmation removes nearly all of them. The field understands this problem and has built good tools for it. What is easy to lose is that the filters are doing the scientific work, and that their settings are not neutral defaults but the parameters on which the answer depends.

The second is not a signal-to-noise problem at all. When a parental mosaic sits at 1% allele fraction, a standard trio genome does not fail to detect it. There was nothing to detect — at 35×, the expected count of variant reads is well below one. The pipeline behaved correctly, the confirmation assay behaved correctly, and the resulting classification is still wrong in the specific way that matters to a family planning another pregnancy.

The label “de novo” carries an implicit claim about where a mutation came from, and the trio data supports a narrower one: the variant is in the child and was not seen in the parents at the depth used. Those two statements are close enough to be interchangeable most of the time, and far enough apart that in roughly one case in twenty, the difference is the entire clinical message.

References

1. Mendelian inconsistent signatures from 1,314 ancestrally diverse family trios distinguish biological variation from sequencing error. Journal of Computational Biology 2019. Cited for the ~1.2 × 10⁻⁸ per-generation rate and the roughly four-order-of-magnitude gap between Mendelian-inconsistency and true DNM rates. Author list not captured — verify before publication. ncbi.nlm.nih.gov/pmc/articles/PMC6533806
2. Human de novo mutation rates from a four-generation pedigree reference (CEPH 1463). Nature 2025. Author list not captured — verify before publication. nature.com/articles/s41586-025-08922-2
3. Burda K, et al. Validation of machine learning approach for direct mutation rate estimation. Molecular Ecology Resources 2023. Cited here for its summary of the twofold between-group variation reported by Bergeron et al. 2022; see ref. 4 for the primary. onlinelibrary.wiley.com — 1755-0998.13841
4. Bergeron LA, Besenbacher S, Bakker J, et al. The Mutationathon highlights the importance of reaching standardization in estimates of pedigree-based germline mutation rates. eLife 2022;11:e73577. pubmed.ncbi.nlm.nih.gov/35018888
5. How precise are mutation rate estimates? Comparison of different approaches to estimate de novo mutation rates. Heredity 2026. Non-human trios (flat-headed loach); cited for the filter trade-off structure rather than human-specific rates. Author list not captured — verify before publication. nature.com/articles/s41437-026-00852-7
6. Study reporting validation of 82% of 150 candidate DNMs by an independent genotyping assay, in a non-human pedigree. Cited for the role of orthogonal confirmation in defining a callset. Full citation to be confirmed before publication. researchgate.net — DNM detection in trio samples
7. Parental germline mosaicism in genome-wide phased de novo variants: recurrence risk assessment and implications for precision genetic counselling. PLOS Genetics 2025;21(3):e1011651. Author list not captured — verify before publication. journals.plos.org/plosgenetics — pgen.1011651
8. Assessment of parental mosaicism rates in neurodevelopmental disorders caused by apparent de novo pathogenic variants using deep sequencing. Scientific Reports 2024;14:s41598-024-53358-9. Author list not captured — verify before publication. nature.com/articles/s41598-024-53358-9
9. Revealing parental mosaicism: the hidden answer to the recurrence of apparent de novo variants. Human Genomics 2023;17:535. Cited for the multi-tissue findings, including gonadal mosaicism confined to semen. link.springer.com — s40246-023-00535-y

Zetobit, LLC · Bioinformatics consulting · Lexington, KY · zetobit.com