All articles

The ontology changed under the case.

A phenotypeThe observable clinical findings of a patient, recorded here as HPO terms with their status. is coded in terms, and the terms belong to a release. When the release moves, a case coded three years ago means something slightly different.

A child's phenotype was coded in 2021 with eleven terms from the ontology release of that spring. In 2024 the case was reanalysed. One of the eleven terms had been retired in the interval and split into two more specific terms; the pipeline mapped it to the nearest surviving ancestor, a broad term that matched a thousand conditions. Two other terms had been moved to a different branch of the hierarchy. The disease ontology had merged the leading candidate's disease entry with another.

The similarity scores in 2024 differed from the ones in 2021 for every candidate. The scientist assumed the patient had changed. The patient had not. The vocabulary had, in three places, and nothing in the system had said so.

What the field knows

The Human Phenotype Ontology is released on a regular cycle and each release changes it (Köhler et al., 2021). Terms are added, retired, split and merged; definitions are refined; the hierarchy is restructured when a branch is revised. A retired term carries a pointer to its replacement where one exists, and the ontology's maintainers document each change, which is what allows a careful user to follow a term across releases and what makes a careless one silently wrong.

Disease ontologies change the same way. The Mondo Disease Ontology (Vasilevsky et al., 2022) unifies disease identifiers across sources and, in doing so, merges entries that turn out to describe one disease and splits entries that turn out to describe two. A gene-disease annotation made against a disease entry in one release may point at a merged or split entry in the next. The OBO Foundry principles (Jackson et al., 2021) ask that ontologies version themselves, keep retired identifiers resolvable, and document changes, and the major biomedical ontologies do.

Semantic similarity depends on the structure it is computed over. The information content of a term, which drives the similarity measures used in phenotype-driven ranking (Köhler et al., 2009), is a function of how many diseases are annotated to the term and its descendants in a given release. Add annotations, restructure the branch or retire the term, and the same patient and the same gene produce a different number. The number was never a property of the patient alone. It was a property of the patient, the gene, the ontology release and the annotation release together, and only the first two are usually named.

Where practice breaks

Term identifiers are stored without the release. A case record holds HP identifiers and nothing that says which release they came from, so when a term is retired the system either fails to find it or maps it to something else without saying so. The mapping is usually to a parent, which is usually broader, which usually weakens the match.

Mapping is automatic and silent. A split term is a question for the clinician, since only the clinician knows which of the two new terms describes the patient, and the pipeline answers the question on the clinician's behalf by picking the ancestor. The patient's phenotype becomes less specific with each release, and nobody chose that. Over three releases a carefully coded case can drift into a vaguely coded one with no human action at all.

Scores are compared across releases. A similarity of 0.82 in 2021 and 0.71 in 2024 is read as a weakening of the evidence for the candidate, when the two numbers were computed over different ontologies and different annotation sets and are no more comparable than two prices in two currencies. A reanalysis that finds the leading candidate has fallen may be reporting nothing more than a restructured branch, and a reanalysis that finds it has risen may be reporting a new annotation to the gene and nothing about the patient.

The report names no release. A report that says the phenotype was assessed against the ontology, with no release, cannot be reproduced, because the ontology it was assessed against no longer exists in that form. The reader in 2027 cannot tell what the terms meant to the author in 2021, and the author, by then, cannot either.

How Tulparus handles it

Every term on a case carries the release it was coded in, and every analysis records the phenotype ontology release, the disease ontology release and the annotation release it ran against. A similarity score is displayed with its releases, and a comparison between scores computed over different releases is labelled as such and never presented as a change in the evidence.

An ontology release is diffed against open cases, like any other knowledge release. When a release retires, splits, merges or re-parents a term that is coded on an open case, the case receives an entry naming the term, the change and the affected cases, and the recoding is a clinician's decision. A term split into two is shown to the clinician with both successors and the patient's original description, and the clinician picks; a term retired with a single replacement is proposed with the replacement, and confirmed; nothing is remapped without a person. The original term and release are kept.

A disease merge in the disease ontology is handled the same way for hypotheses. A hypothesis that names a disease entry which has since been merged shows the merge, with both identifiers and the release, and the scientist decides whether the hypothesis still says what she meant. The annotation release behind a gene-disease link is shown the same way, so that a candidate whose score rose on a new annotation shows the annotation.

For the case above, the 2024 reanalysis would have opened with a list of three terms that had changed, the clinician would have recoded the split term to the specific successor in a minute, and the comparison of 2021 and 2024 scores would have been shown with the releases on each, so that a fall in the leading candidate was read as a vocabulary change first and a change in evidence only if it survived the recoding.

What you can do tomorrow

Recording the release is a one-field change in most systems, and it is the field that makes everything else recoverable.

  • Store the ontology release beside every coded term and in every report. If the system has no field for it, put it in the case note until it does.
  • Before any reanalysis, run the case's terms against the current release and list the retired, split and merged ones. Recode them with the clinician before the ranking runs.
  • Never compare a similarity score across two analyses without checking that the releases match. If they do not, say so in the note.
  • Keep the original terms when recoding. A term the clinician chose in 2021 is evidence of what the clinician saw in 2021, whatever the ontology later did with it.

Sources

  1. Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research, 2021.
  2. Vasilevsky NA, Matentzoglu NA, Toro S, et al. Mondo: unifying diseases for the world, by the world. medRxiv, 2022.
  3. Jackson R, Matentzoglu N, Overton JA, et al. OBO Foundry in 2021: operationalizing open data principles to evaluate ontologies. Database, 2021.
  4. Köhler S, Schulz MH, Krawitz P, et al. Clinical diagnostics in human genetics with semantic similarity searches in ontologies. American Journal of Human Genetics, 2009.
  5. Deans AR, Lewis SE, Huala E, et al. Finding our way through phenotypes. PLoS Biology, 2015.

Eleven terms, one release.See the vocabulary change shown as what it is.

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