A boy of eight with a progressive cerebellar ataxia and no family history has an exome. The variant-level ranking puts a rare homozygous missense change in a gene with a plausible name near the top: it is predicted damaging, it is absent from population data, and the gene is in a pathway the scientist associates with neurodegeneration. Two hours go into it.
The gene is expressed in liver and kidney and, in every available dataset, barely at all in the cerebellum at any age. The candidate that fits his tissue is at position fourteen. The expression information was publicly available and not in the pipeline.
What the field knows
Where and when a gene is expressed constrains what its loss can do. The GTEx consortium (GTEx Consortium, 2020) measured expression across dozens of adult human tissues from hundreds of donors, and the Human Protein Atlas (Uhlén et al., 2015) mapped protein and transcript levels across tissues and cell types. Developmental atlases cover the periods adult tissue cannot: Cardoso-Moreira and colleagues (2019) profiled gene expression across organs through human development, which matters for conditions whose damage is done before birth or in infancy. Single-cell atlases now resolve expression to the cell type, so that a gene can be placed in a Purkinje cell rather than in the cerebellum as a whole.
Prioritisation methods have used expression for years, with care. Tissue-specific expression raises a candidate for a condition in that tissue and lowers one with no expression there, and methods that combine expression with other signals, such as phenotype similarity, constraint and protein interaction, perform better in benchmarks than those that do not. Exomiser's hiPHIVE scoring and successors draw on interaction and model organism data alongside phenotype; dedicated tissue-aware methods go further and weight candidates by expression in the tissue the phenotype points to.
The limits are well described. Bulk expression in an adult tissue says little about a cell type that makes up a small fraction of it, or about a developmental window that has closed. Some genes are expressed at low levels everywhere and cause tissue-specific disease anyway, because the tissue depends on them more. Expression can therefore lower a candidate's standing and should rarely eliminate it, and the strength of the argument depends on which dataset, which tissue and which age were consulted.
Expression also helps in the other direction, when the phenotype is broad. A child with multisystem involvement implicates several tissues at once, and a gene expressed in all of them is a better fit than one expressed in a single organ. The same data that lower a liver gene for an ataxia raise a ubiquitously expressed gene for a syndrome, provided the method asks the question tissue by tissue.
Where practice breaks
Expression is looked up by hand, for the candidate the scientist already likes. It is a confirmation step rather than a ranking input, so it finds support for the favourite and never lowers a gene the scientist did not reach. The check also tends to be generous: a scientist who has invested in a candidate reads a modest signal in a related tissue as enough, where the same signal on an unfamiliar gene would have been read as absence. The candidate at position fourteen, which expression would have raised, is never looked up.
The tissue is chosen loosely. A cerebellar phenotype is checked against 'brain', a cardiac conduction defect against 'heart', and the expression of the gene in the relevant cell type is unknown because the question was never asked that precisely. A gene expressed throughout the cortex and absent from the cerebellum looks fine under 'brain'. The precision of the question sets the value of the answer, and a browser with a dozen tissue labels invites the coarse question.
Age is ignored. A child's neurodevelopmental condition is checked against adult expression data, because those are the data in the browser. A gene expressed strongly in the foetal brain and silent after birth looks like a poor candidate when it may be the right one.
When expression is used, its provenance is lost. The report says the gene is 'expressed in the relevant tissue', without the dataset, the tissue label, the measure or the threshold, so a reader cannot tell whether that means a strong signal in the right cell type or a trace in the wrong organ.
How Tulparus handles it
Expression is a candidate-generation and scoring channel on every case rather than a lookup. From the phenotype terms, the workbench derives the tissues and, where the ontology allows, the cell types the phenotype points to, and the developmental window the age of onset implies. Each candidate gene is scored against expression in those tissues at that window, from named datasets at named releases, and the score is shown with the dataset and tissue it came from.
The channel lowers and raises, and never removes. A gene with no measurable expression in the implicated tissue is moved down with the reason stated, and a reviewer who knows the gene is expressed in a rare cell type that bulk data miss can override the score and record why. A gene with strong, specific expression in the right place and time is moved up, which is how the candidate at position fourteen in the opening case would have reached the first page.
Because the expression channel is one of several, its disagreement with the others is visible. A variant channel that likes a gene and an expression channel that does not is a flag for a person, shown on the fused list rather than averaged away. The page says what each channel measured, and no score is presented as a chance that the gene is the diagnosis.
The report carries the provenance. If expression evidence contributed to a hypothesis, the rendered report names the dataset, tissue, age window and measure, so that the next reader knows exactly how strong the argument was and can revisit it when a better atlas is released. For the boy above, the liver gene would have carried a visible mark against it from the first screen, and the two hours would have gone to the candidate at position fourteen.
What you can do tomorrow
Until expression is in the pipeline, make the manual check systematic.
- For each case, write down the tissue and cell type the phenotype implicates, and the age window, before looking at candidates. Check every reviewed gene against them, in that order.
- Use a developmental atlas for paediatric cases. Adult expression is the wrong question for a condition that began before birth.
- Record which dataset and tissue you consulted, in the case note, every time expression is cited.
- Look at the genes just below your review threshold for strong, specific expression in the right place. That is where the missed candidate usually sits.
Sources
- GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science, 2020.
- Uhlén M, Fagerberg L, Hallström BM, et al. Tissue-based map of the human proteome. Science, 2015.
- Cardoso-Moreira M, Halbert J, Valloton D, et al. Gene expression across mammalian organ development. Nature, 2019.
- Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols, 2015.
- Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research, 2021.