The request form reads: seven-year-old, global developmental delay, seizures, please consider genetic testing. The laboratory codes two terms and runs the exome. The neurology letter from the same hospital, written four years earlier, describes spasms beginning at five months with a characteristic EEG, regression of hand use after the first year, repetitive hand movements, and a head circumference that fell from the fiftieth centile to below the second. None of it reached the laboratory.
The letter describes a recognisable condition. The form describes half the children in the clinic, and the exome was ranked for the form.
What the field knows
Phenotype-driven ranking depends on the phenotype it is given. The methods that compare a patient's terms against disease and gene annotations (Robinson et al., 2014; Smedley et al., 2015) were evaluated on well-described patients, and their performance falls as the terms become fewer and less specific. A term such as global developmental delay sits high in the ontology and matches thousands of conditions. A term such as infantile spasms with hypsarrhythmia, onset at five months, matches a short list. The difference between them is the difference between a candidate list that can be read in an afternoon and one that cannot be read at all, because every gene in it fits a little and none fits well.
The information usually exists. Son and colleagues (2018) showed that phenotype terms extracted from the electronic health record, where clinicians had already written the patient down over years of visits, were more numerous and more specific than the terms on the test request, and that ranking on the record terms placed the causal gene higher. Deisseroth and colleagues (2019) built a tool to extract HPO terms from free-text notes for exactly this reason: the notes hold a phenotype that the form does not, written by the people who examined the patient, at the time they did.
The Human Phenotype Ontology itself (Köhler et al., 2021) is built so that specificity counts. A deep term implies its ancestors, so coding the specific finding loses nothing and gains discrimination. The ontology also distinguishes a finding examined and absent from a finding never recorded, a distinction the field has learned to care about because a confident absence changes a differential and a missing entry does not.
Where practice breaks
The request form is designed for the laboratory's administration, with a box for the indication and a box for the test. It was never designed to carry a phenotype, and it does not. The person who fills it in is often a trainee, working from the most recent letter, under time pressure, and writes the headline problem. The neurologist who could describe the seizures in a sentence is not the person filling in the box.
The clinic letters are in the hospital record and the laboratory does not have access, or has access and no time to read six years of correspondence. Where a laboratory does read them, the terms extracted are written in a free-text comment on the case, where the ranking software cannot see them, and are lost when the case is handed over. A second scientist reanalysing the case two years later starts again from the form.
What is on the form is treated as the phenotype. A ranking run on two terms produces a long, flat list, and the list is read as evidence that there is no good candidate, when it is evidence only that the search was underpowered. The report says no causative variant identified. The letter that would have narrowed the search sits in the same building, and sometimes in the same electronic record, one tab away from the request.
What is missing from the form is treated as absent. Two terms, and the tools assume the patient has no other findings, so a gene whose disease includes microcephaly scores lower, when the microcephaly is in the letter. The pipeline cannot tell a finding nobody looked for from a finding that was looked for and ruled out, and it treats both as if the second.
How Tulparus handles it
The case starts from the clinical question and from whatever documents the clinic has, and the request form is one document among them. Uploaded clinic letters are read for phenotype terms, and every term found is proposed to the clinician with the sentence it came from, the letter's date and author, so that the clinician confirms, corrects or declines it. Nothing enters the phenotype without a person saying so, and every term that does enter carries where it came from.
A term the letters do not mention is recorded as not provided, and the workbench says so. Beside the candidate list it shows which findings, if known, would change the ranking most, and asks for them: age at seizure onset, seizure type, head circumference trajectory, whether hand use was lost. The questions go to the referring clinician as a short list, and the answers come back as terms with a source and a date.
The ranking shows its inputs. Each candidate displays which of the patient's terms it matched, which it did not, and which disease features were never assessed in this patient, so that a scientist can see whether a low rank reflects the patient or the paperwork. A gene whose disease includes a finding recorded as not provided is scored on what is known, and the gap is shown as a gap. A search run on two terms is labelled as such, and the report records the phenotype the analysis used, term by term, with the source of each.
For the child above, the letter would have been read on the day the case opened, the clinician would have confirmed six terms in a few minutes, and the ranking would have been run on the patient the neurologist described. The short list that resulted would have had the right gene near the top, with the spasms, the regression and the head circumference each named as the reason.
What you can do tomorrow
The letters already exist, and the gain comes from getting them into the search in a form the tools can use. None of this needs new software; it needs the phenotype to be treated as part of the test.
- Ask for the most recent clinic letter with every request, and refuse to accept a form with fewer than a handful of specific terms without one.
- Code the terms from the letter into the case before the analysis runs, in the field the ranking reads rather than in a comment.
- Record what was looked for and not found as a separate list from what was never recorded, and keep the two apart in every system that stores them.
- When a ranking is flat, report that the phenotype was thin before concluding that the exome was negative, and say what would narrow it.
Sources
- Robinson PN, Köhler S, Oellrich A, et al. Improved exome prioritization of disease genes through cross-species phenotype comparison. Genome Research, 2014.
- Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols, 2015.
- Son JH, Xie G, Yuan C, et al. Deep phenotyping on electronic health records facilitates genetic diagnosis by clinical exomes. American Journal of Human Genetics, 2018.
- Deisseroth CA, Birgmeier J, Bodle EE, et al. ClinPhen extracts and prioritizes patient phenotypes directly from medical records to expedite genetic disease diagnosis. Genetics in Medicine, 2019.
- Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research, 2021.