A trio exome is run for an infant with a severe epileptic encephalopathy. The de novo filter returns two variants, neither convincing, and the case is reported as negative. Months later a scientist runs a relatedness estimate across the laboratory's recent samples for a different reason and finds that the man recorded as the father shares no more of his genome with the child than a stranger would.
Every de novo call in the case had been filtered against the wrong genome. Variants the child inherited from the biological father had been removed as inherited whenever the recorded father happened to carry them, and the true inheritance of the remaining candidates was unknown. The analysis had to be done again from the start, as a duo, with a conversation first about what the family had and had not been told.
What the field knows
Estimating how related two samples are from their genotypes is a solved problem. Methods based on identity by descent, such as those in PLINK (Purcell et al., 2007), compare allele sharing across many common variants with what is expected for a parent and child, full siblings, second-degree relatives or unrelated people. KING (Manichaikul et al., 2010) made the estimate insensitive to population structure, so that two people of the same ancestry are not mistaken for cousins. Peddy (Pedersen and Quinlan, 2017) packaged the checks for sequencing studies: it reads a pedigree file and the called genotypes and reports every pair whose inferred relationship disagrees with the stated one, the sex inferred from the X chromosome against the sex recorded, and an estimate of ancestry.
The checks matter because the errors they catch are common enough to plan for. Sample swaps, mislabelled tubes, undisclosed non-paternity and unrecorded consanguinity each occur at a rate that is small per sample and large per year. The Deciphering Developmental Disorders study (Wright et al., 2015) ran relatedness and sex checks across thousands of trios as a standing quality step and corrected or removed the samples that failed. Pedersen and Quinlan found pedigree and sex discrepancies in the cohorts they examined and built the tool so that a study could find its own before analysis.
The consequence of a wrong pedigree follows from how a trio is analysed. The power of a trio comes from subtracting the parents: a variant in the child and in neither parent is de novo; a variant in the child and in one parent is inherited and, for a dominant condition with a well parent, probably not the cause. Both deductions depend on the parents being the parents. With the wrong father, every variant the child inherited from the real father is called de novo and floods the list, and every variant the child shares by chance with the stated father is called inherited and dismissed. Both error directions act at once.
Sex is the second check. A recorded sex that disagrees with coverage of the X and Y chromosomes points to a swap or a labelling error, and it also changes the analysis directly: X-linked filters depend on it, and so does the reading of a hemizygous call. The interpretation framework in use across laboratories (Richards et al., 2015) assumes throughout that segregation and de novo status are known, and neither is known until the pedigree has been confirmed by the data.
Where practice breaks
The check is cheap and widely available, and it is still skipped, for ordinary reasons. It is not on the request form; the form carries a pedigree drawn by a clinician from what the family said, and nobody is asked to verify it. It is not in the pipeline of a laboratory that was built for singletons and added trio filtering later; the filter was added, and the relatedness step was never put in front of it. And it is uncomfortable. A laboratory that finds non-paternity has found something the family may not know, the policy for what to do with that is often unwritten, and a finding with no policy is easier not to make.
When the check is run, it is often run once and late. The scientist discovers the discrepancy during review, after the filter has been applied and the candidate list has been read. Everything done so far has to be done again, at the point in the case where time is scarcest.
A subtler failure is the pedigree that is right about parentage and wrong about everything else. Parents who are second cousins and do not know it, or know it and were not asked, change the expected pattern of homozygosity and the weight a recessive candidate should carry. A sibling listed as unaffected who was never examined changes a segregation argument. The kinship estimate catches the first of these. The second is a phenotype problem wearing a pedigree's clothes, and it needs the same discipline about what was examined and when.
Finally, the result is rarely stored where the case is. It lives in a quality log, in a file on a share, in a scientist's memory. The next person to open the case, at reanalysis a year on, does not know whether the pedigree in front of them was ever confirmed by the data.
How Tulparus handles it
The workbench runs relatedness and sex inference on every multi-sample case when the files arrive, before any filter can be applied. The result is written to the case as a finding: each stated relationship with the relationship the data support, each stated sex with the sex the coverage supports, and an estimate of relatedness between the parents. Where everything agrees, the case carries a confirmed pedigree with the date of confirmation, and the trio filters are unlocked. Where something disagrees, the case stops at a gate.
The gate is a decision for a person, with the options laid out. A sample swap is corrected by re-labelling and re-running, and the record keeps the correction. Non-paternity is handled under the laboratory's own policy; the workbench records the decision taken, lets the analysis proceed as a duo with the mother, and keeps the finding out of anything the family will read unless the laboratory and the clinician decide it should be disclosed. Nothing is inserted into a report because a calculation produced it.
Unrecorded consanguinity is surfaced the same way. If the parents' estimated relatedness is higher than the pedigree implies, the case shows it and the recessive candidate channel is told to expect runs of homozygosity. The clinician is asked to confirm with the family rather than having the information silently written into the pedigree.
Because the check is a finding on the case rather than a line in a log, it stays with the case. A reanalysis a year later opens on a pedigree the data confirmed, and says when. If a new sample is added, the check runs again across all of them.
What you can do tomorrow
None of this needs a product. The checks exist as free software and run in minutes on called genotypes.
- Put a relatedness and sex check in front of the trio filter, as a step the pipeline cannot skip.
- Write down what the laboratory does when it finds non-paternity before it finds any. The policy is easier to write without a family waiting.
- Store the result in the case record, with the date, where the next analyst will see it.
- Ask the clinician for the pedigree as it was taken, including who was examined and who was only reported on, so that the pedigree can be checked against the data and the phenotype against the pedigree.
Sources
- Manichaikul A, Mychaleckyj JC, Rich SS, Daly K, Sale M, Chen WM. Robust relationship inference in genome-wide association studies. Bioinformatics, 2010.
- Pedersen BS, Quinlan AR. Who's who? Detecting and resolving sample anomalies in human DNA sequencing studies with Peddy. American Journal of Human Genetics, 2017.
- Purcell S, Neale B, Todd-Brown K, et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. American Journal of Human Genetics, 2007.
- Wright CF, Fitzgerald TW, Jones WD, et al. Genetic diagnosis of developmental disorders in the DDD study: a scalable analysis of genome-wide research data. Lancet, 2015.
- Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine, 2015.