Background: Why stored genomes need to be revisited

A study using the automated genome reanalysis tool Talos identified 241 new diagnoses in 238 of 4,735 people with rare diseases who had not received a complete diagnosis from prior testing. Published in Nature Medicine, the study assessed whether new genetic knowledge could be applied repeatedly to existing test data while narrowing the candidates experts need to review.

The value of genomic data does not disappear once testing is complete. Diagnostic clues may emerge from stored data when previously unknown gene–disease relationships are established or assessments of variant pathogenicity change. Yet manual reanalysis is constrained by limited staff and complex clinical and laboratory workflows. To ease this bottleneck, the researchers developed Talos, an open-source tool.

Key findings: Repeated analysis narrows candidates using current evidence

Talos combines basic annotations, such as variant location and predicted effect, with ClinVar variant classifications and PanelApp Australia evidence for gene–disease relationships. Rule-based filters select candidates and assess inheritance patterns within families; when symptom information is available, the tool also flags relevant candidates. By default, it does not exclude a candidate solely because it does not match the reported symptoms. In repeat runs, Talos presents only candidates whose supporting evidence has changed since the previous analysis, reducing the burden of reviewing the same variant again.

Performance was evaluated using data from 401 participants in an Australian acute-care genomics cohort and 688 in a US rare-genomics project, for a total of 1,089. In the Australian cohort, when the patient and both parents were analyzed together as a trio, the tool recovered 177 of 196 existing diagnoses detectable by the standard variant-calling pipeline, about 90%. It returned about 1.3 candidates per patient. This is not the diagnostic rate for all people with rare diseases; it measures how many detectable existing diagnoses were recovered.

When Talos was applied to a separate group of 4,735 people with undiagnosed or partially diagnosed conditions, expert review and variant classification by clinical laboratories confirmed 241 new diagnoses in 238 people. Three people each received two diagnoses. The paper reported an additional diagnostic yield of 5.1%. The new diagnoses comprised 78 involving new gene–disease relationships, 54 supported by new variant evidence, and 109 resulting from improvements in analysis strategies. Accumulating knowledge contributed, as did revisiting variants missed in earlier analyses.

The researchers conducted 29 monthly reanalysis runs, and each cohort was analyzed at least three times, depending on when it joined. The first reanalysis yielded 221 new diagnoses, with 20 more identified in later cycles. Each subsequent cycle returned an average of five new candidate variants per 1,000 people. This is a measure of review burden and should be distinguished from the number of diagnoses.

Meaning and outlook: Combining automated selection with expert judgment

In a comparison using single-nucleotide variants and short insertions and deletions in trio data from the Australian cohort, Talos and the existing variant-prioritization tool Exomiser had similar sensitivity. There was no significant difference when all Exomiser candidates or its top 10 were reviewed; when review was limited to its top five or top one, results favored Talos. The key distinction lies less in an overall performance advantage than in Talos’s design to present a small number of candidates.

The limitations are clear. Conservative filters may exclude some causal variants, and variant types missed during the earlier variant-calling stage also constrain the analysis. The current tool uses clinical information from the time of the initial test and does not automatically incorporate subsequent changes in symptoms. This was a benchmarking study in observational cohorts, so its findings should also be distinguished from those of a randomized comparative trial. The 5.1% additional diagnostic yield may vary with the composition of the study population, the scope of prior testing, and elapsed time. Expert variant review and clinicians’ judgment remain necessary for a final diagnosis.

The study’s methods and results are available in the original paper.