All articles

The cheapest test is a better phenotype.

Yield follows the quality of the phenotypeThe observable clinical findings of a patient, recorded here as HPO terms with their status. more than the depth of sequencing. Recording what was examined, what was absent and what was never looked at.

A girl of nine is referred for exome sequencing with three terms on the request form: global developmental delay, hypotonia, dysmorphic features. Her clinic letters, read together, also mention nystagmus noted at eight months and never mentioned again, a heart murmur called innocent on one echocardiogram, feeding difficulty through infancy, and a growth chart that crossed two centile lines downward between three and six. None of that reached the laboratory.

The exome was negative.

What the field knows

The Human Phenotype Ontology was built for exactly this gap. Robinson and colleagues (2008) introduced it as a controlled vocabulary of clinical abnormalities arranged so that a computer can tell that a specific finding is a kind of a more general one; the 2021 update (Köhler et al.) describes a resource of many thousands of terms, with annotations linking them to diseases and genes. The phrase deep phenotyping (Robinson, 2012) names the practice the ontology was made for: recording the patient's abnormalities precisely and completely, rather than as a diagnosis-shaped summary.

Phenotype-driven prioritisation turns those terms into ranking. The Exomiser (Smedley et al., 2015) and the tools that followed it compare a patient's HPO terms with the terms annotated to each candidate gene's diseases, and move genes up or down the list accordingly. The pilot of the 100,000 Genomes Project (Smedley et al., 2021) built its diagnostic pipeline around HPO-coded phenotypes and this kind of prioritisation, so the terms a clinician entered were an input to the analysis, not an annotation on it.

Two properties of these methods matter for the clinic. First, specific terms carry more information than general ones: a rare, precise finding narrows the candidate list far more than a common, broad one, and the similarity calculations are built to reward exactly that. Second, the methods can only score what they are given. A finding that was examined and found absent is evidence against candidates that predict it. A finding that was never examined is nothing, and should be scored as nothing. The ontology and the Phenopacket standard both make room for an excluded phenotype, so that absent and unrecorded can be told apart. Most request forms do not.

There is also the patient with two diagnoses. Posey and colleagues (2017) found that a small but consistent fraction of patients with a molecular diagnosis from exome sequencing had two, and that the blended phenotype had often been read as one unusual presentation of a single condition. A phenotype recorded completely is what lets a second explanation be seen.

Where practice breaks

The phenotype that reaches the laboratory is a summary of a summary. The clinician writes a letter that condenses years of notes; the request form condenses the letter; a free-text box of three or four terms condenses the form. At each step the general survives and the specific is lost. Delay survives; the age at which the delay became apparent does not. Dysmorphic features survives; the particular features, which are what the annotations are made of, do not.

Negatives are lost first. A clinician who has examined the eyes and found them normal rarely writes that down for the laboratory, because it does not feel like a finding. For a prioritisation method it is one of the most useful findings there is: it argues against every candidate whose disease predicts a retinal or ocular abnormality. Its absence from the record is read as nothing, and the candidates it would have argued against stay on the list.

Onset and course go next. Nystagmus at eight months that resolved, a murmur on one echocardiogram, growth that faltered between three and six: these are temporal facts that distinguish conditions from each other, and a request form has no field for them. The same is true of what was not done. If the child has never had an eye examination, the honest record is not tested, and a candidate that predicts retinal findings should stay open, with the examination as the next step. A blank box cannot say that.

Finally, the phenotype is captured once, at the point of ordering, and then frozen. The child in the case above is nine. When she was tested at seven she had no seizures. If she has them now, the negative exome is being read against a phenotype that no longer describes her.

How Tulparus handles it

The workbench starts from the notes rather than from the form. Clinical text is read into structured phenotype terms, and each term keeps the sentence it came from, so a clinician can see that nystagmus came from a letter at eight months and judge whether it still belongs. Every finding carries one of four states: present, absent, unknown or not tested. The states are separate values throughout the system, and a term that is unknown or not tested is never scored as absent.

The phenotype ranker then compares the patient's recorded findings with what is known about each candidate gene's diseases. A present finding that matches raises a gene; an examined-and-absent finding that the disease predicts lowers it; an unknown or untested finding does neither. The ranker shows which terms drove each position, so a clinician who disagrees with a term can correct it and watch the list change.

What the system cannot score, it names. If a candidate predicts retinal findings and the record says not tested, that gap appears as missing evidence and as a proposed next step: an eye examination, with the candidate it would support or weaken. For the girl above, the next most useful action may be an afternoon in ophthalmology rather than a genome, and the record says so.

The phenotype stays live. New letters can be read in, terms can be added as the child grows, and any re-check against new knowledge runs on the phenotype as it is now, not as it was at ordering. A rank remains a rank. The workbench does not tell a clinician what the child has; it tells them which of the candidates the evidence fits, which it does not, and what has not been looked at.

What you can do tomorrow

Most of the gain is in habits that cost minutes.

  • Before ordering a genomic test, re-read the letters for three things the form will not ask for: findings examined and absent, age of onset for each finding, and anything that resolved. Put them on the request as text if there is no field.
  • Use the most specific HPO term you can defend, not the parent term. Bilateral sensorineural hearing loss with onset in infancy does more work than hearing impairment.
  • Write examined and absent as a finding, and not tested as a finding. Give the laboratory the chance to tell them apart.
  • Before any reanalysis, re-phenotype. The patient has changed since the test was ordered, and the new findings are the cheapest evidence the reanalysis will get.

Sources

  1. Robinson PN, Köhler S, Bauer S, et al. The Human Phenotype Ontology: a tool for annotating and analyzing human hereditary disease. American Journal of Human Genetics, 2008.
  2. Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research, 2021.
  3. Robinson PN. Deep phenotyping for precision medicine. Human Mutation, 2012.
  4. Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols, 2015.
  5. Smedley D, Smith KR, Martin A, et al. 100,000 Genomes Pilot on Rare-Disease Diagnosis in Health Care. New England Journal of Medicine, 2021.
  6. Posey JE, Harel T, Liu P, et al. Resolution of disease phenotypes resulting from multilocus genomic variation. New England Journal of Medicine, 2017.

One case, fully phenotyped.See what it changes.

A demo takes 20 minutes, screen shared, on synthetic cases. Bring the question from one of your own.