When a child with a severe developmental disorder has their genome sequenced alongside both parents (a "trio"), diagnostic software throws away any candidate variant that an unaffected parent also carries. The logic is that a healthy parent with the same variant shows it cannot be the cause. This filter is the main reason trio sequencing works. It also carries a hidden assumption: that a parent's blood sample reports the parent's whole body. It does not. A parent can be mosaic, meaning the variant arose after conception and sits in only a fraction of their cells. That parent can look healthy, show a faint signal in blood, and still pass the variant through their eggs or sperm. A variant can also be truly de novo, absent from both parents, yet arise after conception in the child and appear in blood at a fraction well below the expected half of reads.
In both cases, averaging DNA across many cells hides the variant's true distribution. This is the genomic version of the "cellular heterogeneity masking" that single-cell biology has documented for gene expression. Standard trio filters then turn that hidden signal into a confident wrong call. We propose to replace the yes/no inheritance filter with a model of what fraction of cells carry each variant, built on the allele fractions the trio already measured. We then audit how many "incomplete penetrance" misses and unsolved families in public trio cohorts are explained by mosaicism. Finally, we test whether the corrected calls change diagnostic rankings, and whether those changes hold up under orthogonal laboratory confirmation.
Core Claim
The trio filter treats parental genotype as discrete: absent, heterozygous, or homozygous. We claim a measurable share of its known failures comes from treating a continuous quantity, the fraction of a parent's sampled cells carrying the allele, as discrete. Modeling that fraction explicitly should do three things. It should recover causal variants that current pipelines discard because a "carrier" parent looks unaffected. It should flag apparent de novo variants that are really low-level parental mosaics, which matters for how likely the condition is to recur in a sibling. And it should produce a short, ranked list of variants worth sending for confirmation by deep targeted sequencing, rather than the 50-candidate lists that current tools produce.
The prior work frames the problem. Refined preferences of prioritizers improve intelligent diagnosis for Mendelian diseases benchmarked four diagnostic ranking tools (Exomiser, PhenIX, AMELIE, LIRICAL) on DDD and KGD trios. It found that switching to trio mode was the single most impactful setting. It also found that causal variants with low penetrance were routinely discarded because a healthy parent carried them. That paper motivates our proposal, and its tools are our baselines. Huang et al.'s study of postzygotic single-nucleotide mosaicism in clinically unremarkable individuals established that such mosaic mutations are present in healthy people and that bulk sequencing cannot resolve which cell lineages carry them. It supplies the biological premise. Single-nucleotide variant calling in single-cell sequencing data with Monopogen calls variants from sparse single-cell data, but notes it cannot separate germline variants from early embryonic somatic ones when allele fractions approach 50%. That limitation is exactly the ambiguity we target from the bulk side. It also supplies a tool for one of our validation routes. None of these papers models parental cell fraction inside the trio filter.
Research Program
The core object is a per-variant, per-individual likelihood over the fraction of sampled cells carrying the alternate allele. For a parent with $a$ alternate reads out of $d$ total reads at a site, we model
$$ a \sim \text{Binomial}\big(d,\; \epsilon + (1-2\epsilon)\, f\big), \quad f \in \{0\} \cup (0, 0.5) \cup \{0.5\}, $$
Here $f$ is the fraction of sampled DNA copies carrying the variant: zero for absent, one half for a constitutional heterozygote, and anything in between for mosaic. The term $\epsilon$ is a site-specific sequencing error rate. We estimate $\epsilon$ from unrelated samples in the same cohort at the same position, which keeps systematic artifacts from looking like mosaicism. Clustered reads from duplicated or poorly mapped regions spread the counts more than a binomial predicts. We will therefore allow a beta-binomial extension and decide between the two forms on held-out sites, not by assumption.
The same model applied to the child separates a constitutional de novo variant (near 0.5) from a postzygotic one in the child (well below 0.5). The joint posterior over the three family members' fractions replaces the binary Mendelian label. Each candidate then falls into one of three states. A variant is "constitutionally inherited" when the parent's fraction is near 0.5; the current filter is correct here. It is "parental mosaic" when the parent shows a small but nonzero fraction; the variant becomes eligible again as a candidate cause with reduced expressivity in the parent. It is "proband postzygotic" when it is absent in both parents and sits at a low fraction in the child.
The second component is an audit, which we expect to be the most durable contribution even if the model itself turns out to be simple. Using read-level allele counts from DDD trios, the KGD cohort, and the RGP families evaluated in Critical assessment of variant prioritization methods for rare disease diagnosis within the rare genomes project, we will count how often a solved family's causal variant was observed at low fraction in a parent or the child. We will also count how often the current filter would have discarded or down-ranked it. In unsolved families, we will enumerate variants that were filtered for parental presence but have high posterior probability of parental mosaicism. The audit answers a question the file leaves open: how much of the "incomplete penetrance" problem documented by Yuan et al. is actually mosaicism that the filter cannot see?
The third component plugs the posterior into existing ranking tools. We do not build a new phenotype matcher. Instead, we replace the hard trio filter inside Exomiser and LIRICAL with a soft weight equal to the posterior probability that the variant is compatible with the child's phenotype given each parent's estimated fraction. Every other setting stays at Yuan et al.'s optimized configuration. That isolates our ingredient.
The fourth component addresses the gap between computational prediction and laboratory confirmation. It is deliberately designed so validation is cheap and decisive. Mosaic calls make physical predictions that ordinary pathogenicity scores do not. A specific parent should show a specific low fraction on deep targeted resequencing of stored DNA. Where a second tissue such as saliva or buccal cells is available, the fraction should differ between tissues if the mosaicism is confined to particular cell lineages. If the parental mosaicism reaches the germline, the same variant should recur in affected siblings. Where single-cell data from a parent exists, Monopogen-style genotyping can show which cell types carry the allele. We rank variants for confirmation by expected diagnostic value per assay. This turns the model into a tool for allocating confirmation experiments, not just another score.
Baselines
The naive reference point is the standard binary trio filter as configured in Yuan et al.'s optimized Exomiser and LIRICAL settings. Any variant with alternate reads above the caller's threshold in an unaffected parent is treated as inherited and deprioritized. On DDD trios, Exomiser in this configuration ranked the causal gene first in 61.6% of cases, and the four tools placed it in the top ten in 83.3–91.8% of cases (Refined preferences of prioritizers improve intelligent diagnosis for Mendelian diseases). We will rerun these exact configurations on the same families and the same variant calls, so the only difference is the inheritance component.
The strongest competitor on mosaic detection itself is a simple per-site statistical test. It asks whether the parent's alternate read count exceeds what the site-specific error rate predicts, and it uses the same error estimates our model uses. This baseline is dangerous because it may capture most of the signal. We give it identical inputs and tune its significance threshold on the same training families our model is tuned on. No reported result for this baseline or any mosaic-aware trio method was found on Lacuna. Beating it means higher recall of orthogonally confirmed mosaic variants at a matched false-positive rate, and a larger downstream gain in top-1 or top-5 diagnostic rank.
The strongest end-to-end diagnostic competitor on RGP is the top CAGI6-RGP model, Invitae Moon. It identified the causal variant within the top five in 13 of 14 solved families (Critical assessment of variant prioritization methods for rare disease diagnosis within the rare genomes project). With only 14 solved test families, we cannot claim to beat it on that set. The meaningful comparison is on the 16 unsolved families: does the mosaic-aware pipeline surface candidates that its predecessors filtered out, and do any of those candidates survive confirmation?
The proband-only competitor is popEVE (Proteome-wide model for human disease genetics). It calibrates evolutionary and population-constraint scores for missense variants across all proteins and reports prioritizing the causal mutation from the child's exome alone in 98% of cases on its severe developmental disorder cohort. We will run popEVE-ranked candidates with no parental information as a third arm.
The key ablation keeps everything in our pipeline but collapses the fraction posterior back to three discrete genotypes. Any gain that survives this ablation cannot come from modeling the fraction, and we will report it as such.
Why not just…?
Why not just drop parents and use a proband-only classifier like popEVE? If parents were truly dispensable, the mosaicism question would vanish along with the trio filter. Two things in the file argue against this. Yuan et al. measured that trio mode was the single largest improvement across four tools on 457 families. popEVE also covers only missense variants, so it cannot see the deep intronic and near-splice causal variants that CAGI6-RGP confirmed in ASNS and TCF4. Still, popEVE's 98% figure is a real tension. We will settle it empirically with a three-arm comparison: proband-only, binary trio, and fraction-aware trio, on the same DDD families. Our prediction is that binary trio filtering loses to proband-only ranking precisely in the families where a parent is mosaic, and that fraction-aware trio analysis recovers the advantage of having parental data.
Why not just lower the variant caller's allele-fraction threshold? This is the most serious threat. A low threshold on parental reads, or the site-error test described above, might recover most mosaic variants. If it matches our model, we will say so, and the contribution shrinks to the audit plus the soft integration into ranking tools. We expect a threshold to fail in two places. First, at low-depth exome sites, the same read count can mean mosaic or noise depending on depth, so a fixed threshold either floods candidates or misses them. Second, it gives a call but no probability to weight a ranking or set confirmation priority. The matched-false-positive comparison against confirmation data decides this.
Why not just add single-cell or tissue expression priors to the ranking tools? This is the obvious way to bring cellular heterogeneity into trio diagnosis, and it is a reasonable idea. But it addresses a different failure. Expression priors help choose among variants that survive filtering. They cannot restore a causal variant the filter already deleted because a mosaic parent looked like a carrier. Atlas-based approaches such as Cross-modal transfer learning for mapping bulk transcriptomes at cellular level also depend on the reference atlas containing the relevant cell states, which is doubtful for transient embryonic progenitors. The two ideas combine naturally, and we treat expression priors as a later layer rather than a rival.
Why not just sequence parents' cells individually? Single-cell genotyping would show mosaicism directly. However, standard variant callers recover under 8% of variants from such data, and even Monopogen needs population reference panels and still struggles near 50% allele fraction. That cost and sparsity make it a confirmation tool, not a screening tool. We use it in that role for a small number of high-posterior cases.
Proposed Experiments
The first experiment is the retrospective audit on DDD and KGD trios with known diagnoses. Its primary measure is the fraction of solved cases whose causal variant shows parental fraction posteriors between 0.02 and 0.4, together with the rank change under fraction-aware filtering. If this fraction is negligible, the program stops here with a useful negative result.
The second experiment is the diagnostic comparison. We report top-1, top-5, and top-10 causal-gene rates for binary trio Exomiser and LIRICAL, the site-error test, popEVE proband-only ranking, our full model, and the collapsed-genotype ablation, all on identical families and calls. Confidence intervals come from bootstrapping over families, since a handful of cases will drive any difference.
The third experiment is calibration against natural labels. We use apparent de novo variants recurring in affected siblings in the multi-disorder trio collection studied by Quantifying concordant genetic effects of de novo mutations on multiple disorders, where family-level data permits. Recurrence implies parental germline mosaicism, so variants with higher posterior parental fractions should recur more often.
The decisive check is prospective. For the unsolved RGP families, we will send the top-ranked reclassified variants to deep targeted resequencing, using a second parental tissue where stored. We pre-register the predicted fraction for each call. The proposal succeeds if confirmed fractions fall within the predicted intervals at a clearly higher rate than for calls from the site-error test, and if at least some confirmed variants pass clinical review as candidate diagnoses. It fails, informatively, if confirmation shows the low-fraction parental signals are mostly technical artifacts.
References
- Refined preferences of prioritizers improve intelligent diagnosis for Mendelian diseases
- Single-nucleotide variant calling in single-cell sequencing data with Monopogen
- Critical assessment of variant prioritization methods for rare disease diagnosis within the rare genomes project
- Proteome-wide model for human disease genetics
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Quantifying concordant genetic effects of de novo mutations on multiple disorders