Reanalyse unresolved rare-disease cases

• Workload only: the tables count calls returned to experts per caller, not new diagnoses per caller. Per-caller detection of the 77 confirmed pathogenic CNVs is printed only in the Results text (claim on the protocol).
• A CNV found by more than one caller was returned once per caller (Supplementary Table 2, 'Tool' annotation), so per-caller counts overlap and their sum (7849) counts calls, not distinct CNVs.
• Primary call thresholds differ by caller. ExomeDepth's Bayes factor threshold is printed as BF > 15 (Methods 'Alignment and definition of capture regions of interest' paragraph 2) and as discarding BF < 0.15 ('ExomeDepth workflow' paragraph 2).
• ClinCNV was developed by the paper's first author (ref. 47; Author contributions), so its row is author_reported.
• CNV reanalysis only; no SNV or indel reanalysis method is compared. Expert time is reported only as a mean of 5 minutes per CNV, pooled over callers (Methods 'Clinical interpretation').

Protocol: CNV calls returned to clinical experts per caller after filtering, Solve-RD exome reanalysis (Demidov et al. 2024 Table 1, Supplementary Table 4)

What each column measures:

• ACG trio: Known diagnoses recovered and candidate variants returned per proband by Talos default and strict (phenotype-match) filtering, ACG trio cohort
• ACG singleton: Known diagnoses recovered and candidate variants returned per proband by Talos default and strict (phenotype-match) filtering, ACG singleton cohort
• ACG full: Known diagnoses recovered and candidate variants returned per proband by Talos default and strict (phenotype-match) filtering, ACG full cohort
• RGP trio: Known diagnoses recovered and candidate variants returned per proband by Talos default and strict (phenotype-match) filtering, RGP trio cohort
• RGP singleton: Known diagnoses recovered and candidate variants returned per proband by Talos default and strict (phenotype-match) filtering, RGP singleton cohort
• RGP full: Known diagnoses recovered and candidate variants returned per proband by Talos default and strict (phenotype-match) filtering, RGP full cohort

• Recall here is recovery of diagnoses that were already known. It is not new diagnostic yield.
• Only known diagnoses detectable by the standardised GATK, GATK-SV and GATK-gCNV pipeline form the recovery denominator (Table 1 footnote).
• Talos developers' own evaluation (author_reported).
• Candidates per proband is the ratio of table totals; the text calls such values medians (stored protocol limitation).
• The same evaluations are also mapped as proxy for use-case-rare-disease-candidate-ranking.

Protocols: Talos diagnosis recovery and workload: ACG trio, Talos diagnosis recovery and workload: ACG singleton, Talos diagnosis recovery and workload: ACG full, Talos diagnosis recovery and workload: RGP trio, Talos diagnosis recovery and workload: RGP singleton, Talos diagnosis recovery and workload: RGP full

• The 37 cases were already diagnosed, so this measures recovery of known later diagnoses. It is not new diagnostic yield.
• Run by the Exomiser developers on their own tool (author_reported).
• Supplementary Table 1 does not name the database releases compared. The 0.2/0.8 row (TP 31, FP 23) matches the Results comparison of the February 2019 and February 2022 databases; reading the whole table as that comparison is an inference (claim on the protocol).
• The thresholds were chosen by F2 on the same 37 cases, so the 0.2/0.8 row, which is the recommended setting and the highest F2, overstates expected performance on new cases. The other 80 rows are as printed.

Protocol: Flagging new reanalysis candidates from Exomiser score changes across database releases (Vestito et al. 2024 Supplementary Table 1)

What each column measures:

• Reanalysis of 219 unresolved cases: New P/LP and VUS case findings, case-category counts, manual review time and Talos candidates per case for manual and Talos reanalysis of the same 219 unresolved genomes
• Talos benchmark, singletons: Share of 35 known P/LP variants and count of 21 known VUS that Talos prioritised in 45 singleton cases, with failure modes
• Talos benchmark, trios: Share of 145 known P/LP variants Talos prioritised in 162 trio cases in trio mode and proband-only mode, VUS captured and failure modes

• Unequal inputs: the manual arm re-aligned and re-called the raw reads with DRAGEN v4.2.4 and used Emedgene; Talos used the archived DRAGEN v3.7.5 VCFs. Differences cannot be attributed to the method alone.
• This does not meet the question's condition of the same updated information: the arms had different variant calls and partly different knowledge dates, so a difference between them cannot be attributed to the method.
• The reference is the manual arm's own result: Talos is scored on whether it returned what manual reanalysis found, so findings that only Talos might have made are not counted. Talos can only match or miss what manual reanalysis found; candidates only Talos returned were not adjudicated. Talos found all three new P/LP cases from the archived v3.7.5 calls, so none of the P/LP gains depended on the manual arm's re-calling.
• Very small numbers: three new P/LP cases in each arm.
• The Talos total diagnostic yield is printed as 41.9% in a sentence that also gives 158 to 161 cases and calls it identical to the manual result; 161 of 377 is 42.7%, and 41.9% is the initial yield. That one result is disputed and not shown; the Talos count of 161 is stored.
• Talos output per case is printed as an average of three variants (Results; Discussion P3) and as a median of approximately three (Discussion P1); the stored value is the average. It is a rounded word, and the cases it averages over (219 or the whole cohort) are not stated.
• Manual time excludes report writing and preparation of evidence for secondary review; no hands-on time is printed for Talos.
• Single centre, high initial yield (41.9%), one reanalysis round after a mean of 660 days.
• A new candidate or classification is not a confirmed diagnosis; per-case confirmation details are in the supplement, which was not read.
• Review effort is measured in different units: mean analyst minutes per case for the manual arm, candidate variants per case for Talos. No analyst time is printed for reviewing the Talos output, so the two cannot be compared as effort.
• Manual reanalysis ran from October 2024 to December 2025 and Talos once in October 2025, so the knowledge available to each arm differed by case.
• Illumina, which makes DRAGEN and Emedgene, provided sequencing reagents, software licences and technical support and supported the study (Competing interests; Funding statement).
• Positive controls are known findings from the initial analysis; recovering them does not establish new diagnostic yield.
• Concordance is per variant against the initial report, not per case. Methods 'Benchmarking of Automated Reanalysis' P1 defines the detection rate per case; the Results report concordance per variant (for example 28 of 35 P/LP variants), which is what is stored.
• The pipeline conversion share for singletons is 8.6% in the text but 9.6% in the Figure 2 legend; the text value is stored (3 of 35 is 8.6%).
• Strict inheritance filtering in trio mode excluded 11 P/LP variants that proband-only mode recovered.

Protocols: Manual versus Talos reanalysis of 219 unresolved genomes after a mean 660 days (Kaschta et al. 2026), Talos recovery of known P/LP and VUS variants in 45 singleton cases (Kaschta et al. 2026), Talos recovery of known P/LP and VUS variants in 162 trio cases (Kaschta et al. 2026)

• No equal-input refreshed conventional control.
• No randomization/blinding; updated calls and knowledge contribute to yield.

Protocol: Talos programme reanalysis yield

• Recall here is recovery of diagnoses that were already known. It is not new diagnostic yield.
• The Results text gives 194 trio families with SNV or indel diagnoses, and every printed percentage matches 194 (172/194 = 88.7%, printed 89%; 127/194 = 65.5%, printed 65%). Only the Extended Data Fig. 1 legend gives N = 190, for the paired panels, which are not extracted.
• Talos Table 1 uses 196 known ACG trio diagnoses of all variant types, so Talos and Exomiser recall have different denominators and sit in separate comparison groups.
• Exomiser v14 with default configuration; data release, phenotype input and pathogenicity sources not stated.
• A rank budget (top 1, 5 or 10) is not measured analyst effort.

Protocol: Exomiser rank-budget recovery in ACG trios