A scientist assessing a missense variant searches a literature database for the gene and the variant, reads six abstracts and two full papers, decides that two case reports describe patients with a similar phenotype and comparable variants, and records 'supported by the literature' with two citations. Eighteen months later the case is reanalysed by a colleague, who searches again, finds the same two papers and four more that have appeared since, and cannot tell which of the four are new and which the first scientist saw and set aside.
The first search took an afternoon. It left behind two citations and no trace of the reasoning.
What the field knows
The literature is evidence under the interpretation framework. Several ACMG and AMP criteria (Richards et al., 2015) rest on published observations: previously reported patients with the variant, functional studies, the established mechanism of disease for the gene. Each requires that a specific paper be found, read and judged, and the judgement, whether a reported patient's phenotype really matches, whether a functional assay is well validated, is where the expertise sits.
The volume makes the search a real task. Tens of thousands of papers a year touch human genetic variation, and a gene with an established disease association may have hundreds of relevant reports spread across decades, nomenclatures and journals. Tools exist to help: AMELIE (Birgmeier et al., 2020) reads the primary literature and ranks papers by their fit to a patient's phenotype and candidate genes, and LitVar (Allot et al., 2018) indexes variant mentions across the literature under their normalised identifiers so that the same variant can be found whatever notation a paper used. Variant databases such as ClinVar (Landrum et al., 2018) carry citations attached to submissions, which is a partial, curated record of what others found.
Evidence-based medicine solved the repeatability problem for systematic reviews decades ago: a review records its search strategy, databases, dates and inclusion criteria so that another team can reproduce it and update it. Clinical variant interpretation makes a smaller search for every case and almost never records it to that standard. The reanalysis guidance (Deignan et al., 2019) assumes the earlier evidence review can be compared with the current one, which in practice it cannot.
The judgement is the expensive part and the recordable part. Deciding that a reported patient matches this one means comparing phenotypes, ages, inheritance and variant class, and it takes a clinician's or a scientist's time. Once made, it is a small structured fact: this paper, this patient, this claim, supports or does not, because. The fact fits in a sentence, and it is the sentence nobody writes.
Where practice breaks
The search is unrecorded. Which terms, which database, which date, which filters: none of it is written down, so the search cannot be repeated or extended. A colleague who finds a relevant paper the first scientist missed cannot tell whether it was missed, found and rejected, or published since.
Rejections are lost. A scientist who reads a paper and decides its patient does not match has done useful work that protects the next reader from repeating it. The decision lives nowhere. The next reader opens the same paper and does the same reading. Across a laboratory's caseload the same handful of papers for a common gene are read and rejected dozens of times a year by different people, with no one the wiser.
Nomenclature hides papers. A variant reported under a legacy transcript, a protein-level description or an older gene symbol is missed by a search on the current name, and the omission is invisible because nobody knows what the search should have returned.
At reanalysis the whole search is redone, badly or not at all. Redoing it costs an afternoon per variant, which a backlog cannot afford; skipping it means the reanalysis is a re-run of the annotations with the human evidence left stale. Either way the new papers, which are the ones most likely to change the answer, are the least likely to be found.
Functional studies are the hardest to track. A single assay paper can carry strong weight under the framework, its validity depends on details a reader has to assess, and it is often published in a journal the variant search does not reach. A scientist who finds and weighs one has done something the laboratory should keep.
How Tulparus handles it
The literature is an indexed evidence source with a release, like every other source the workbench uses. The index resolves variant and gene mentions to normalised identifiers, so that a paper that used an older notation is found, and each investigation records the release of the index it searched, so that a later reanalysis can list what has been published since rather than searching from scratch.
The scientist's reading becomes evidence on the case. Each paper surfaced for a candidate can be marked as supporting, against, or not relevant, with a reason and the specific claim it supports, and each mark is kept with the scientist's name and the date. A paper rejected once stays rejected, visibly, with the reason, until someone chooses to revisit it. Nothing in this is done by a model on the scientist's behalf; the index finds, the person judges, and the judgement is what is kept.
When the literature index is updated, the difference is computed against each open and closed case: new papers mentioning a variant or gene under consideration appear as a queue entry with the case and the candidate they touch, so that reanalysis of the literature is a review of what is new. For the colleague in the opening case, the four later papers would have been presented as new, the two earlier ones as already read, with the first scientist's notes beside them.
The report cites what the case cites, with identifiers, and the citation list is rendered from the evidence record rather than typed. A reader can follow each claim in the interpretation back to the paper and to the judgement made about it, and a laboratory can answer, for any report, exactly which literature it rested on at the time.
What you can do tomorrow
The systematic review's habits transfer to the single variant with little effort.
- Record the search: database, terms, date, and any filters, in the case note. One line per search.
- Record rejections as well as citations. 'Smith 2019: patient phenotype does not match, no neurological features' is more useful to the next reader than the two papers you kept.
- Search under the legacy names too: previous gene symbols, protein-level notation, older transcripts. Write down which you tried.
- At reanalysis, limit the first pass to papers published after the recorded search date. If there was no recorded date, make one now.
Sources
- Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine, 2015.
- Birgmeier J, Haeussler M, Deisseroth CA, et al. AMELIE speeds Mendelian diagnosis by matching patient phenotype and genotype to primary literature. Science Translational Medicine, 2020.
- Allot A, Peng Y, Wei CH, Lee K, Phan L, Lu Z. LitVar: a semantic search engine for linking genomic variant data in PubMed and PMC. Nucleic Acids Research, 2018.
- Landrum MJ, Lee JM, Benson M, et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Research, 2018.
- Deignan JL, Chung WK, Kearney HM, Monaghan KG, Rehder CW, Chao EC. Points to consider in the reevaluation and reanalysis of genomic test results: a statement of the American College of Medical Genetics and Genomics (ACMG). Genetics in Medicine, 2019.