Genomics
How a genetic testing company traces every piece of variant evidence back to the paper
Talk to our team
- Every item traceable
- each case links back to its place in the paper, or its row in a supplementary table
- Draft, then review
- Micius drafts the evidence summary; scientists judge case independence and evidence strength
- Every edit logged
- each correction a scientist makes goes into a review log that can be checked later
Gathering published evidence for one variant means searching paper by paper, downloading full texts, digging through supplementary tables, copying cases into a spreadsheet and then judging how strong the evidence is. At this genetic testing company, that work is still mostly done by hand. In the workflow it is building with Zhiman, Micius reads the papers first and drafts the evidence, every item linked to its source, and variant scientists check it and make the call.
Challenge
Combing full texts and tables for one variant
Take a rare missense variant in a gene linked to a recessive disorder. To apply the ACMG/AMP PM3 criterion, which asks whether affected patients carry the variant together with a pathogenic variant on the other copy of the gene, a variant scientist has to find every published patient who carries it and check what is present on the other allele.
That is rarely one search. Some papers give the c. (DNA-level) change, some only the protein change, and older ones use legacy names. Patients are seldom described in the abstract, so each hit means downloading the full text and finding them in a results paragraph or a row of a supplementary table. The scientist copies patient, zygosity and inheritance into a spreadsheet, checks whether two papers describe the same family, and only then weighs the evidence. That is one criterion for one variant, and a report can carry several candidate variants.
The company holds tens of thousands of disease and gene records, hundreds of thousands of variants and literature in the millions. Searching it still depends mostly on people, and judging whether cases are independent takes experienced scientists. Two scientists given the same variant can also come back with different papers and different cases.
Extraction
Micius reads the full text first
In this project, the workflow runs like this. A keyword search finds new papers in public literature databases, and Micius screens the abstracts and lists the ones worth reading. Once a scientist confirms the list, they upload the full-text PDFs with their supplementary tables, and a task list shows whether each paper is queued, in progress or done. The workflow runs on the customer’s own servers, and the PDFs are parsed there.
In one case series, the abstract says only that three patients carry the variant. The details are further in: one sentence of the results gives the first patient’s two variants, shown by parental testing to be in trans, and the sentences on the other two point to a supplementary table. Micius pulls out the case-level evidence for each patient (condition, variants, homozygous or compound heterozygous) and attaches a source. An item from a table points to its row, and an item from the text points to its sentence. Anything the paper doesn’t state, or that couldn’t be extracted, is marked “not stated”.
So this paper yields three items. Patient 1 is compound heterozygous, sourced to a paragraph of the text. Patient 2 is homozygous, sourced to a table row. For patient 3 the paper gives only “heterozygous”, so the other allele reads “not stated”.
Review
A scientist checks every case against the paper
A scientist reviews each extracted case in a two-pane workbench: structured results on the left, the paper on the right. Opening a case moves the paper to where it came from, so nobody has to reread the whole paper, and when a paper reports several candidate variants the patients can be filtered by variant. The scientist corrects or adds fields as needed, and every edit goes into a review log. For patient 1, the variants and zygosity point to paragraph 3, which is highlighted on the right; the scientist reads it and marks the case reviewed.
Illustrative flowOpen a case and the paper jumps to its source; each check is logged.
Reviewed evidence can be exported as a spreadsheet.
Interpretation
A draft summary by gene and variant
At interpretation time, the scientist enters a gene and a variant in c. or p. notation. Micius gathers the case-level evidence from the parsed papers that mention it and drafts a summary for PM3 and for PS4, the criterion on whether the variant is more common in affected people than in controls. Every sentence cites its paper and links back to it. The summary can only reach as far as the papers already processed, and it is marked as a draft: case independence, evidence strength and the final classification stay with the scientists and experts.
For one variant, four parsed papers cover six patients. Two carry it in trans with a variant reported as pathogenic, two are homozygous, and for two the other allele isn’t stated. The PS4 line counts carriers only, notes that they haven’t been compared with controls, and leaves the criterion to the reviewer.
Illustrative flowSix patients’ evidence becomes one draft; independence and strength stay with a person.
Results
Starting from evidence that is already sourced
The old workflow had five steps, from searching and downloading to copying out cases, judging and writing up the call, and people did all of them a paper at a time. Now Micius reads, extracts and drafts first. People decide at two points: the review against the source, and the expert call at the end.
Illustrative flowBefore, people did all five steps; now Micius reads and drafts first, and people decide at two points.
Extraction has already been run in batches on more than 90 published papers. The genes, conditions and patient counts for each paper sit in one table that can be searched by paper, with the original PDF a click away. Not every paper extracts cleanly; where no gene or condition could be pulled out, the table says “not stated” for a person to fill in.
What’s next
Training the company’s own model on reviewed evidence
The project’s second track is to turn reviewed evidence into training data for the company’s own inherited-disease model, which is meant to take a patient’s clinical features and suggest likely conditions and candidate genes, with the evidence behind them. That work is still in progress. The training data is published evidence that the lab’s scientists have reviewed, not patients’ genetic data. Training runs inside the customer’s lab, and the model belongs to the customer. Zhiman does not use this data to train its own models.
The same approach also works for
-
Reclassification alerts
When new papers mention variants you’ve already reported, gather the new evidence and prompt a scientist to check whether a classification should change.
-
Functional and segregation evidence
Extract functional study results and family segregation data from papers, traced to the source in the same way, for review under criteria such as PS3 and PP1.
-
Somatic variants in cancer
Gather published evidence on therapy response and prognosis for a somatic variant and draft a summary against the tiering guidelines, for the interpretation team to judge.
-
Pharmacogenomics
Pull together evidence linking genotype to dosing changes from papers and prescribing guidelines, each item cited, for pharmacy and clinical teams to check.
Still combing supplementary tables by hand?
Tell us which variants you check most and where the evidence comes from. We’ll show which steps Micius can take on, and we can talk through running it on site.