A boy of three with a sensorineural hearing loss, a structural kidney anomaly and a distinctive facial appearance has a trio exome. A de novo loss-of-function variant sits in a gene with no human disease association. The pipeline ranks by human disease annotations and the gene scores nothing, because there are none. It is at position ninety.
The knockout mouse for the same gene, published eight years earlier, is deaf, has malformed kidneys and has craniofacial abnormalities. The match is close enough that a clinician reading the mouse paper would have recognised the child. Nobody read it, because no step in the analysis asked.
What the field knows
Model organism phenotypes are a long-standing source of evidence for human gene function, and for prioritisation they were made computable by aligning the species' phenotype ontologies. The PhenoDigm method (Smedley et al., 2013) compares a human phenotype profile with mouse and zebrafish phenotypes through cross-species ontology mappings and scores the similarity, so that a gene with no human disease can still be ranked by how much its animal model resembles the patient. Exomiser's hiPHIVE score (Smedley et al., 2015) builds on this, combining human disease, mouse and fish phenotype similarity with protein interaction data, and the published evaluations show that the model organism channel finds genes the human-only channel cannot.
The data behind it are systematic. The International Mouse Phenotyping Consortium (Dickinson et al., 2016; Meehan et al., 2017) is knocking out every protein-coding gene in the mouse and phenotyping each line through a standard battery, which means the mouse evidence for a gene no longer depends on whether anyone happened to study it. The uber-anatomy and phenotype ontologies (Mungall et al., 2012) make the comparison between a mouse hearing test and a human audiogram tractable, with the losses in translation documented.
The limits are real and well characterised. Many human disease genes have no mouse phenotype at the standard battery's sensitivity; some mouse phenotypes have no human equivalent; a knockout models loss of function and says little about a missense change; and the match between a mouse feature and a human one is a semantic approximation. The model organism channel is therefore a generator of hypotheses, a strong one for the novel gene, and a weak basis for a classification on its own.
Zebrafish and fly add what the mouse cannot. A zebrafish morphant or mutant can be made and phenotyped in weeks for a gene the mouse consortium has not reached, and fly genetics has mapped pathways that the vertebrate data only sample. The cross-species ontologies cover these too, with more translation and more loss, and a channel that reads all of them finds candidates that any one would miss.
Where practice breaks
Clinical pipelines rank by human disease annotations only. The model organism channel exists in the tools and is switched off, or its results are folded into a single score where a strong mouse match cannot lift a gene the human annotations ignore. A de novo loss-of-function variant in a gene with a matching knockout ranks with the noise.
The mouse is consulted after the fact. A scientist who has already chosen a candidate looks up the knockout to support it; the gene at position ninety is never looked up. The channel is used to confirm and never to find. A confirmation step inherits the biases of the choice it confirms, and the mouse that matches the wrong candidate is read more charitably than the mouse that matches the right one would have been.
The evidence is misweighted in both directions. In some hands a matching mouse phenotype is read as proof, and a novel gene is reported with more confidence than the standards allow. In others it is dismissed as research data and never enters the clinical reasoning. Neither treats it as what it is: a specific, sourced, moderate piece of evidence about a gene, with a known distance from the patient.
The novel gene is then lost to the matchmaking networks, which depend on someone having identified the candidate in the first place. The second patient who would confirm the gene is never sought, because the first patient's gene was never seen. Years later, when another laboratory publishes the gene from its own patients, the first family is still filed as negative.
How Tulparus handles it
Model organism phenotype similarity is one of the candidate channels on every case, scored against the patient's phenotype through the cross-species ontologies, from named releases of the mouse and fish phenotype resources. Its output is shown as its own column beside the human disease channel, so that a gene with no human annotations and a strong mouse match is visible exactly as that, rather than averaged into obscurity. The boy's gene would have been near the top of the model organism column on the first day, and the column is the first thing a reviewer sees when the human column is empty.
The channel's score is labelled with its distance. The page shows which mouse features matched which patient findings, through which ontology mappings, and which patient findings had no counterpart in the model. A reviewer can see that the hearing loss and kidney anomaly matched and the facial features were an approximation, and weigh accordingly.
A candidate raised mainly by the model organism channel is handled under the novel-gene standard: a hypothesis with a stated evidence bar, the mouse match recorded as one item towards it, and the second patient and functional data recorded as the items outstanding. The matchmaking submission is prepared from the case where the laboratory has opted in, and the report describes the candidate as a research lead with its evidence, never as a diagnosis.
When the mouse and fish resources release, the channel is re-run and its changes listed, so that a gene whose knockout was phenotyped last year reaches the cases it touches. The case does not depend on a scientist remembering to read a paper, and the reviewer sees the new match as a change with a date rather than as a list to re-read.
What you can do tomorrow
The channel is available in free tools and most laboratories can switch it on.
- Enable model organism phenotype scoring in your prioritisation tool and look at its top ten as a separate list beside the combined score.
- For every de novo loss-of-function variant in a gene with no human disease, look up the knockout phenotype before dismissing it. It takes two minutes per gene.
- Record the mouse evidence on the case as evidence, with the source and the matched features, and weight it under your novel-gene standard.
- Submit the candidates that the model organism channel raises to a matchmaking network, and keep the submission on the case.
Sources
- Smedley D, Oellrich A, Köhler S, et al. PhenoDigm: analyzing curated annotations to associate animal models with human diseases. Database, 2013.
- Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols, 2015.
- Dickinson ME, Flenniken AM, Ji X, et al. High-throughput discovery of novel developmental phenotypes. Nature, 2016.
- Meehan TF, Conte N, West DB, et al. Disease model discovery from 3,328 gene knockouts by The International Mouse Phenotyping Consortium. Nature Genetics, 2017.
- Mungall CJ, Torniai C, Gkoutos GV, Lewis SE, Haendel MA. Uberon, an integrative multi-species anatomy ontology. Genome Biology, 2012.