The diagnosis may already be in old genomic data. The missing ingredient is rechecking it

A genomic test can be technically finished while its interpretation remains unfinished. The open-source Talos system repeatedly re-examines stored sequencing data as evidence about genes, diseases and variants changes. Its central proposition is unusually practical: some people who remain undiagnosed do not necessarily need another sample first; they need yesterday’s data compared with today’s knowledge.[1][2][3]

That is not the same as asking an AI system to diagnose a patient from raw DNA. Talos filters and prioritises candidate variants for expert review. In the reported deployment, accredited laboratories subsequently classified every likely causative variant that became a diagnosis as pathogenic or likely pathogenic. The automation narrows the review queue; the clinical laboratory still establishes and reports the result.[2]

Why old data can contain a new answer

Sequencing records the letters observed in a person’s DNA. Interpretation asks a changing set of questions: which genes are credibly connected to a condition, which variants have enough evidence to matter, how inheritance fits the family, and whether the original analysis included the relevant variant class. The stored sequence may stay the same while the evidence around it changes.[2][4][5]

PanelApp Australia curates gene–disease relationships and versioned diagnostic panels; ClinVar aggregates submitted classifications of variants and the evidence attached to them. Neither resource is a magic answer key. PanelApp records expert assertions and their evidence, while ClinVar makes submissions and disagreements visible rather than independently replacing every submitter’s judgement. Reanalysis depends on how those changing resources are interpreted and governed.[4][5]

The third category is important because it prevents a simplistic story in which all progress came from newly published science. Sixty-nine diagnoses involved copy-number or structural variants not covered by the earlier analysis; other cases had been narrowed too aggressively by phenotype or filtering. Reanalysis can therefore capture both new knowledge and a better use of knowledge that already existed.[2]

The real optimisation target is human attention

A system that returns every vaguely plausible variant would maximise recall at the cost of exhausting the laboratory. Talos was designed for high specificity: its validation result paired 90% recovery of in-scope trio diagnoses with 1.3 candidates per case. That pairing matters more operationally than either number alone. A diagnostic service needs to know both what the filter may miss and how much qualified review it creates.[2]

The monthly passes make the same point. Ninety-two per cent of the programme’s diagnoses appeared on a cohort’s first Talos pass. That does not make later reanalysis pointless: the first pass collected the accumulated benefit of years of updated evidence and broader analysis. Once a cohort was current, later cycles surfaced an average of five new candidates per 1,000 cases—one per 200—because unchanged findings were not repeatedly placed back in front of reviewers.[1][2]

What the 5.1% figure does—and does not—mean

The paper reports 241 diagnoses among 238 of 4,735 people. Those denominators answer different questions: 241 divided by 4,735 is 5.09 diagnoses per 100 participants, while 238 divided by 4,735 means about 5.03% of people received a diagnosis. Neither rate is a promise for another programme. Yield depends on prior testing, variant coverage, time since analysis and the evidence resources used.[2]

The study itself shows that dependence. The additional yield was 6.1% in genome cases and 4.8% in exome cases. Among people originally tested in 2019 it was 8.3%, compared with 3.2% for the 2021 group. Those comparisons are descriptive rather than randomised estimates of the benefit caused by one sequencing method or waiting period, but they reveal why a single headline percentage is not a portable forecast.[2]

The next version of the programme is a learning system, not an autonomous clinic

A mature reanalysis service would record the version of every evidence source, the exact filters used, the candidate queue shown to reviewers, the laboratory decision and any later correction. Talos supports reproducibility through fixed software environments and versioned annotation inputs. That audit trail is the foundation for measuring sensitivity, reviewer burden, turnaround time and differences between patient groups.[2][3]

One current limitation is that Talos reuses the clinical features supplied with the original request. A patient’s phenotype can evolve, and the paper identifies automated updating from clinical records as a possible future improvement. That would add useful evidence, but it would also create a new responsibility: explaining which patient information was extracted, how it changed prioritisation and when a clinician must correct it.[2]

The broader lesson is about directing scarce expert attention. Automation created value by moving newly relevant evidence into a repeatable, clinically governed review process. The pipeline prioritised candidates; expert review and laboratory classification turned those candidates into recognised diagnoses.

Report an error or suggest a correction