Automated reanalysis of genomic data for rare disease diagnostics at scale Authors: Matthew J. Welland, K. D. Ahlquist, Paul De Fazio, Christina Austin-Tse, Lynn Pais, Laura Wedd, Samantha Bryen, Rocio Rius, Michael Franklin, Caitlin Morrison, Giles Hall, Laura Gauthier, Alex Bloemendal, David I. Francis, Andrew J. Mallett, Amali Mallawaarachchi, Paul J. Lockhart, Richard Leventer, Ingrid E. Scheffer, Katherine B. Howell, Karin S. Kassahn, Hamish S. Scott, Julie McGaughran, John Christodoulou, David R. Thorburn, Bryony A. Thompson, Chirag V. Patel, Greg Smith, Anne O’Donnell-Luria, Simon Sadedin, Heidi L. Rehm, Sebastian Lunke, Jeremiah Wander, Kaitlin E. Samocha, Cas Simons, Daniel G. MacArthur, Zornitza Stark Published: 2026 ABSTRACT Reanalysis of genomic data in rare disease is highly effective in increasing diagnostic yields but remains limited by manual approaches. Automation and optimization for high specificity will be necessary to ensure scalability, adoption and sustainability of iterative reanalysis. We developed Talos, an open-source tool that automates variant prioritization by integrating dynamically updated gene−disease and variant-level evidence with inheritance-aware filtering and validated its performance using data from 1,089 individuals with rare disease. Trio-based analysis identified 90% of known diagnoses, returning 1.3 variants per case on average. Variant burden reduced to one variant per 200 cases on iterative monthly reanalysis. Application to an unselected cohort of 4,735 undiagnosed individuals identified 241 diagnoses (5.1% yield): 78 (32%) due to new gene−disease relationships, 54 (22%) due to new variant-level evidence and 109 (45%) due to improved analysis strategies. Our automated, iterative reanalysis model demonstrates the feasibility of delivering frequent, systematic reanalysis at scale. Main Genomic testing has transformed rare disease diagnosis, with many more families receiving a timely and accurate molecular diagnosis to guide clinical care1–3. However, more than 50% of individuals remain undiagnosed after the initial test4. Unlike most other diagnostic investigations, genomic data can be stored and repeatedly reanalyzed, and such reanalysis has proven highly effective in delivering new diagnoses over time. A meta-analysis of 29 studies comprising 9,419 undiagnosed patients with Mendelian disorders reported an increase in diagnostic yield of 10% (95% confidence interval: 6−13%) after a median time of approximately 24 months5. Several research studies have delivered reanalysis in large cohorts, using different models. The Deciphering Developmental Disorders study performed cohort-wide, automated reanalysis between 2014 and 2018, incorporating updated annotations and pipeline improvements, resulting in an additional 13% increase in diagnostic yield in a cohort of 1,133 individuals with neurodevelopmental disorders6. The Exomiser variant priorization tool has been applied to 24,015 unsolved rare disease cases from the 100,000 Genomes Project, increasing diagnostic yield by 2% but requiring extensive manual review and tool optimization to increase efficiency7. Most recently, the Solve-RD Consortium demonstrated the feasibility of scaling manual case-based reanalysis in a cohort of 6,004 families, using a collaborative network approach combined with dedicated infrastructure, with an 8.4% increase in diagnostic yield8. These studies have illuminated the principal sources of new diagnoses: emerging knowledge of disease-associated genes and variants and improvements in analysis processes9. The knowledge base for genomic interpretation is constantly evolving. Each year, hundreds of novel gene−disease associations and thousands of new disease-associated variants are identified and deposited in public databases10–12, directly affecting diagnostic yields. The bioinformatic tools used in analysis also continue to mature. Diagnostic pipelines increasingly incorporate analysis for a broader range of variant types, including copy number variants (CNVs), structural variants (SVs) and short tandem repeats (STRs), resulting in more diagnoses. Reanalysis is supported at a policy level13–15 yet, in practice, is limited by complex clinical and laboratory processes, lack of reimbursement and workforce shortages16,17. A very low proportion of undiagnosed individuals currently benefit from reanalysis, and access is inequitable, primarily driven by highly motivated clinicians, laboratories and patients16,18. Without these inconsistent factors to drive reanalysis, patients remain undiagnosed, which, in turn, may directly impact the quality and cost of clinical care. Meanwhile, the volume of stored genomic data that would benefit from reanalysis continues to grow year on year. Automation has been proposed as a possible solution, potentially enabling reanalysis to be scaled equitably to thousands of datasets and to be performed frequently. Clinical and laboratory workforce attitudes toward automation are largely positive16, with models incorporating automation starting to emerge19–24. These efforts have used a mix of proprietary tools19, in-house solutions22–24 or case-by-case submission to external analysis portals20, all of which limit broader adoption. Key questions remain about how to implement reanalysis at scale in both research and clinical practice. In any model, tradeoffs around sensitivity, specificity, acceptable variant burden, frequency of reanalysis and degree of automation would be required to balance the added strain on resources while optimizing diagnostic outcomes25,26. Hence, evaluations of the performance and workload implications of different models in large, unselected cohorts are needed to inform future policy and practice. Results Talos is an open-source tool designed for iterative reanalysis of genomic data in rare disease (Fig. 1), intended to be run periodically as new gene−disease associations and variant classifications emerge, returning only variants with newly actionable evidence. We assessed the performance of Talos in validation cohorts that had previously undergone manual analysis in clinical and research settings. We then deployed Talos to perform iterative reanalysis in an unselected cohort of patients with undiagnosed rare disease to evaluate diagnostic yield and variant burden.Fig. 1Overview of the Talos reanalysis workflow.Overview of the Talos reanalysis workflow including required inputs (genomic data, metadata and annotation resources); key components of the variant prioritization algorithm; and outputs for manual review and result return. The variant filtering and prioritization stage uses a variant tagging and filtering process to select variants likely to be classified as disease causing using ACMG/AMP criteria. When data are available, the family and phenotype modules filter based on concordant MOI and can filter or prioritize variants based on phenotype. ‘PM5’ refers to ACMG/AMP criteria for evidence derived from amino acid substitutions at the same position. Datasets enter the Talos workflow after uninformative primary analysis, and the variant filtering, prioritization, review and reporting steps are repeated monthly as databases recording variant−disease and gene−disease associations are dynamically updated. Only new information is displayed during iterative cycles. Icons were sourced from The Noun Project (https://thenounproject.com/) under an Icon Pro license. Overview of the Talos reanalysis workflow. Overview of the Talos reanalysis workflow including required inputs (genomic data, metadata and annotation resources); key components of the variant prioritization algorithm; and outputs for manual review and result return. The variant filtering and prioritization stage uses a variant tagging and filtering process to select variants likely to be classified as disease causing using ACMG/AMP criteria. When data are available, the family and phenotype modules filter based on concordant MOI and can filter or prioritize variants based on phenotype. ‘PM5’ refers to ACMG/AMP criteria for evidence derived from amino acid substitutions at the same position. Datasets enter the Talos workflow after uninformative primary analysis, and the variant filtering, prioritization, review and reporting steps are repeated monthly as databases recording variant−disease and gene−disease associations are dynamically updated. Only new information is displayed during iterative cycles. Icons were sourced from The Noun Project (https://thenounproject.com/) under an Icon Pro license. Assessing Talos performance We assessed the performance of Talos using two previously analyzed cohorts that included both diagnosed and undiagnosed individuals. These cohorts were selected to enable benchmarking against manual analysis and to evaluate performance with and without parental data. The cohorts had previously undergone a range of clinical and research analyses, which together contributed to the set of known diagnoses used for benchmarking. Each cohort was reprocessed from raw sequencing data using a standardized pipeline incorporating Genome Analysis Toolkit (GATK) for short variant calling, GATK-SV for SV detection in genomes and GATK-gCNV for CNV analysis in exomes (Methods). The performance of Talos was evaluated against the set of variants detected by this pipeline. Acute Care Genomics cohort We first applied Talos to the Acute Care Genomics (ACG) cohort, in which ultra-rapid clinical genomic testing had been performed as a first-tier test in 401 critically ill infants and children between 2018 and 2022 as part of a national Australian study1,27. Sequencing was performed as exomes (n = 115) or genomes (n = 286), with most cases analyzed as trios (378/401, 94%; Fig. 2). The most common reasons for referral were seizures and neonatal hypotonia. Over 50% of probands were <1 month of age at the time of testing. Primary manual diagnostic analysis yielded a molecular diagnosis in 55% (220/401) of probands, including three with a dual diagnosis1,27.Fig. 2Cohort characteristics.Characteristics of the ACG, RGP and prospective cohorts, including cohort size, prior diagnostic yield, major phenotypes, proband age distribution at the time of the original test, availability of parental data and genomic test modality (exome or genome). NDD, neurodevelopmental disorder; WGS: whole-genome sequencing. Cohort characteristics. Characteristics of the ACG, RGP and prospective cohorts, including cohort size, prior diagnostic yield, major phenotypes, proband age distribution at the time of the original test, availability of parental data and genomic test modality (exome or genome). NDD, neurodevelopmental disorder; WGS: whole-genome sequencing. Among all diagnostic variants previously reported in the cohort, 260 were single-nucleotide variants (SNVs)/insertions and deletions (indels), 13 were CNVs, two were mitochondrial variants, one was an STR and two were retrotransposon insertions. Inheritance patterns included 91 diagnoses due to de novo variants in genes associated with autosomal dominant conditions, 15 due to inherited variants in genes associated with autosomal dominant conditions, 91 due to variants in genes associated with autosomal recessive conditions (54 with compound heterozygous variants, 37 homozygous), 13 X-linked and two mitochondrial. Of the 208 trios with known diagnoses in the ACG cohort, pathogenic variants were detectable within the GATK/GATK-SV/GATK-gCNV pipeline for 196 (94%; Table 1), with the remaining 11 diagnoses attributable to variant types not reliably captured by these tools, including STR expansions or low-level mosaic variants. Talos returned a median of 1.3 candidate variants per trio for manual review (range, 0−13). These candidates included variants responsible for 177 of the known diagnoses (90%). Stricter filtering using high-level phenotype matching (Methods) reduced the number of candidate diagnoses per family to 1.1 and correctly identified 164 diagnoses (83%; Table 1).Table 1Results of Talos analysis in evaluation cohortsACG cohortRGP cohortTrio onlySingletonFull cohortTrio onlySingletonFull cohort Probands377401401390688688 Known diagnosesa196208208549393Talos default filtering Diagnoses177 (90%)170 (82%)186 (89%)47 (87%)81 (87%)80 (86%) Candidate variants4877295235191,4911,095 Candidates per proband1.31.81.31.32.21.6 Candidates per diagnoses2.84.32.811.018.413.7Talos strict filtering (phenotype match required) Diagnoses164 (83%)160 (77%)173 (83%)45 (83%)78 (84%)77 (83%) Candidate variants403491419335777582 Candidates per proband1.11.21.00.91.10.8 Candidates per diagnoses2.53.02.47.49.97.6a Known diagnoses where the pathogenic variant(s) were detectable using the standardized GATK/GATK-SV/GATK-gCNV analysis pipeline described in the Methods. Table 1Results of Talos analysis in evaluation cohortsACG cohortRGP cohortTrio onlySingletonFull cohortTrio onlySingletonFull cohort Probands377401401390688688 Known diagnosesa196208208549393Talos default filtering Diagnoses177 (90%)170 (82%)186 (89%)47 (87%)81 (87%)80 (86%) Candidate variants4877295235191,4911,095 Candidates per proband1.31.81.31.32.21.6 Candidates per diagnoses2.84.32.811.018.413.7Talos strict filtering (phenotype match required) Diagnoses164 (83%)160 (77%)173 (83%)45 (83%)78 (84%)77 (83%) Candidate variants403491419335777582 Candidates per proband1.11.21.00.91.10.8 Candidates per diagnoses2.53.02.47.49.97.6a Known diagnoses where the pathogenic variant(s) were detectable using the standardized GATK/GATK-SV/GATK-gCNV analysis pipeline described in the Methods. Results of Talos analysis in evaluation cohorts a Known diagnoses where the pathogenic variant(s) were detectable using the standardized GATK/GATK-SV/GATK-gCNV analysis pipeline described in the Methods. Of the 20 missed diagnoses, eight (40%) involved moderate or low impact variants that lacked supporting information in ClinVar, all in recessive genes, and another four resulted from incomplete penetrance (Extended Data Table 1). These variants were not returned by Talos due to its intentionally conservative filtering strategy. During the original manual analysis, these variants had been classified as likely pathogenic using a mixture of criteria, including being in trans with another pathogenic/likely pathogenic (P/LP) variant, having a highly specific phenotype match or having supportive functional studies1. Given the frequent use of singleton testing in both research and clinical settings, we next assessed the performance of Talos in the absence of parental data. To simulate singleton analysis, we removed parental data from all 401 probands in the ACG cohort, including those originally tested as duos and not included in the trio analysis above. Talos returned a median of 1.8 variants per singleton and identified 170 (82%) of the 208 achievable diagnoses (Table 1 and Extended Data Table 2). The decline in diagnostic performance compared to trios was primarily due to 16 missense variants in dominant genes, where de novo status was required to support likely pathogenic classification. When phenotype-based filtering was applied, the median number of candidate variants reduced to 1.2 per singleton, with 160 diagnoses identified (77%; Table 1) (Extended Data Table 3). When all ACG families were considered (trios and non-trios), Talos identified 186 of 208 diagnoses (89%), returning a median of 1.3 candidate variants per family (Table 1). Phenotype-based filtering reduced the median number of candidate variants per family to 1.0 and correctly identified 173 diagnoses (83%). Rare Genomes Project cohort The Rare Genomes Project (RGP) cohort includes families with rare disease recruited through a remote study model launched in 2018, enabling direct partnership with families and advocacy groups to enroll participants from all 50 US states and Puerto Rico. Most had already undergone clinical genetic testing with negative or inconclusive results. For this study, we analyzed data from 688 RGP-enrolled families (298 singletons and 390 trios) with genome sequencing performed by the Broad Institute prior to September 2022. At enrollment, probands ranged in age from 7 months to 82 years, with a median age of 26 years (Fig. 2). Prior analysis identified 106 molecular diagnoses, the diagnostic yield of 15.4% reflective of a cohort with previous uninformative clinical testing. Among the 106 prior diagnoses in the RGP cohort, 98 were due to SNVs or indels, 15 were CNVs or SVs, two were mitochondrial variants, eight involved STRs and three were detected using gene-specific callers. Inheritance patterns included 58 diagnoses attributed to autosomal dominant conditions (29 de novo; four inherited, including one with incomplete penetrance; and 25 with unknown inheritance due to absence of parental data), 34 attributed to autosomal recessive conditions (23 compound heterozygous and 11 homozygous), 13 attributed to X-linked conditions (10 de novo and three inherited or unconfirmed) and two attributed to conditions with mitochondrial inheritance. Of the 57 trios with known diagnoses in the RGP cohort, pathogenic variants were detectable within the GATK/GATK-SV/GATK-gCNV pipeline for 54 (95%), with the remaining two diagnoses attributable to one STR and one CNV in SMN1 (Table 1). Talos returned a median of 1.3 candidate variants per trio for manual review (range, 0−12). These candidates included 47 of the 54 diagnoses possible (87%). Stricter filtering using high-level phenotype matching (Methods) reduced the number of candidate diagnoses per family to 0.9 and correctly identified 45 diagnoses (83%; Table 1). In the absence of parental data (simulating singleton analysis), Talos returned a median of 2.2 candidate variants per singleton and identified 81 (87%) of the 93 possible diagnoses (Table 1). When phenotype-based filtering was applied, the median number of candidate variants reduced to 1.1 per singleton, with 78 diagnoses identified (84%) (Extended Data Table 4). Across all 688 RGP families, Talos identified 80 of 93 possible diagnoses (86%), returning a median of 1.6 candidate variants per family (Table 1). When phenotype-based filtering was applied, the median number of candidate variants dropped to 0.8 per family, correctly identifying 77 diagnoses (83%). Comparison with other tools To contextualize these findings and evaluate how Talos compares to existing automated variant prioritization pipelines, we benchmarked this aspect of its performance against Exomiser, a widely used tool28. SNVs and indel call sets from the ACG cohort were analyzed using Exomiser v.14 with default configuration. In the ACG cohort trios, 194 families had known diagnoses attributable to SNVs or indels. When all variants prioritized by Exomiser were considered, it successfully identified the causative variant in 172 families (89%), with the lowest-ranked diagnosis at 32. Performance decreased when limiting the number of variants reviewed per family. The causative variant was identified in 164 families (85%) when only the top 10 ranked variants were considered, in 157 families (81%) when limited to the top five and in 127 families (65%) when only the highest-ranked variants were reviewed. Because outcomes are paired within the same families, we used McNemarʼs exact test on discordant pairs, defined as families where Talos identified the causative variant but Exomiser did not versus families where Exomiser identified the causative variant but Talos did not. This showed no significant difference at the all-variants and top 10 settings (Talos-only 12 versus Exomiser-only 12, P = 1; 18 versus 10, P = 0.18, respectively) but favored Talos at top five and top one (25 versus 10, P = 0.017; 53 versus 8, P < 0.0001, respectively). These results suggest that Talos and Exomiser have similar sensitivity for SNVs and indels, although their design strategies differ. Although Exomiser relies on ranking and returns a broader list, Talos is optimized to return a smaller number of highly specific unranked candidate variants per family. The specific variants identified by each tool differed (Extended Data Fig. 1). Iterative reanalysis in research and clinical datasets We next evaluated real-world utility through deployment of Talos at scale in a large, unselected cohort of 4,735 individuals who had undergone prior genomic testing but remained undiagnosed (Fig. 2). This included 962 pediatric and adult research participants recruited as part of Australian Genomics rare disease flagship studies investigating the diagnostic utility of genomic testing across a range of clinical indications29. Genomic sequencing was performed by clinically accredited laboratories between 2017 and 2022 and was largely undertaken as singletons (n = 739, 76%), with average diagnostic yield of 33%, varying between 17% and 54% based on clinical indication for testing29. We also included 3,773 pediatric and adult patients referred for genomic testing to a single diagnostic laboratory (Victorian Clinical Genetics Services) between 2019 and 2022, who remained without a full diagnosis after initial testing. The full cohort of 4,735 individuals included 3,739 who had exome sequencing and 996 with genome sequencing. Most were analyzed as singletons (n = 3,737), with 998 probands sequenced as part of a trio. One-third were children under 5 years of age at the time of the original test (n = 1,440; 30%). The most common reasons for referral were neurodevelopmental disorders (n = 1,662; 35%), followed by cardiac (n = 997; 21%), renal (n = 695; 15%) and neurological (n = 416; 9%) conditions. Comprehensive genomic analysis had previously been undertaken in 2,086 individuals (44%), whereas the remaining 2,649 (56%) had analysis restricted to virtual panels based on clinical indication (for example, cardiac or renal panels; Fig. 2). Review of Talos output identified 268 likely causative variants consistent with 241 new diagnoses in 238 individuals (5.1% additional yield; Fig. 3 and Supplementary Table 1). All likely causative variants were subsequently classified as P/LP by accredited clinical laboratories. Of the variants identified, 168 (62%) were coding SNVs and indels, 32 (12%) were noncoding SNVs and indels and 69 (26%) were CNVs and SVs. Three individuals received a dual diagnosis.Fig. 3Results from iterative reanalysis using Talos in prospective cohorts of individuals remaining without a diagnosis after genomic testing.a, Sources of new diagnoses in the full cohort and in the three main subcohorts: neurodevelopmental, cardiac and renal. Each tile represents a diagnosis. b, Genomic data aggregation timeline showing the entry of exome and genome data into the program between 2023 and 2025. c, Diagnostic yield based on year of original analysis, including reason for reanalysis diagnosis. Each tile represents a diagnosis. d, Timelines of three illustrative cases, demonstrating the short gap between new information becoming available and reanalysis diagnosis through Talos (highlighted in orange). NDD, neurodevelopmental disorder; QC, quality control; WGS: whole-genome sequencing. Results from iterative reanalysis using Talos in prospective cohorts of individuals remaining without a diagnosis after genomic testing. a, Sources of new diagnoses in the full cohort and in the three main subcohorts: neurodevelopmental, cardiac and renal. Each tile represents a diagnosis. b, Genomic data aggregation timeline showing the entry of exome and genome data into the program between 2023 and 2025. c, Diagnostic yield based on year of original analysis, including reason for reanalysis diagnosis. Each tile represents a diagnosis. d, Timelines of three illustrative cases, demonstrating the short gap between new information becoming available and reanalysis diagnosis through Talos (highlighted in orange). NDD, neurodevelopmental disorder; QC, quality control; WGS: whole-genome sequencing. More than half of the additional diagnoses resulted from changes in knowledge about either gene−disease relationships (n = 78; 32%) or variant−disease relationships (n = 54; 22%). The remaining 109 diagnoses (45%) comprised variants that could have potentially been identified as part of the original analysis. Analysis for variant types not included in the original test, such as CNVs and SVs, identified 69 additional diagnoses, including CNVs ranging in size from 0.6 kilobases (kb) to 1.8 megabases (Mb) and one translocation transecting the PKD2 gene in an individual with cystic renal disease. Of the 68 CNVs, 63 (93%) were small intragenic deletions considered to be below the resolution of standard chromosomal microarray analysis (CMA) (200 kb). Other variants were not reported during the original diagnostic analysis either due to overly restricting analysis based on phenotype (n = 27) or based on filtering criteria (n = 10), or they were missed due to human error (n = 10). Of the 78 diagnoses attributed to new gene−disease relationships, 46 (59%) were not yet curated in Online Mendelian Inheritance in Man (OMIM) at the time of reanalysis, and 11 (14%) were due to new modes of inheritance reported for well-established disease genes, such as ALPK3 and IFT140. Diagnostic yield from reanalysis was similar across clinical indications: 5.2% additional yield (86/1665) for neurodevelopmental disorders, 5.3% (53/997) for cardiac disorders and 5.9% (41/695) for renal disorders. Sources of new diagnoses varied by clinical indication for testing. New gene−disease associations and newly detected CNVs accounted for 78% of new diagnoses in individuals with neurodevelopmental disorders, whereas variant reclassification accounted for 58% of new diagnoses in individuals with cardiac disorders (Fig. 3b). The diagnostic yield from reanalysis of exome cases was 4.8% (180/3,741) compared to 6.1% (61/996) in genome cases. Of the diagnoses obtained in the genome cohort, six (10.5%) were only tractable due to availability of genome data (four noncoding variants in RNU4-2, one deep intronic variant in MRPL39 and one chromosomal translocation). Diagnostic yield varied based on time elapsed since the original analysis, with 8.3% additional yield in individuals undergoing diagnostic testing in 2019 (73/879), 6.4% in 2020 (68/1,068), 3.2% in 2021 (42/1,319) and 4.0% in 2022 (49/1,234) (Fig. 3c). Twenty-nine cycles of iterative monthly reanalysis were performed. Cohorts entered the reanalysis program at different timepoints (Fig. 3a), with all cohorts undergoing at least three cycles of iterative reanalysis. Whereas 221 diagnoses (92%) were identified as part of the initial cycle of reanalysis upon entry into the program, 20 diagnoses (8%) were obtained as part of iterative monthly cycles. All were due to new information about gene−disease relationships. During iterative reanalysis, Talos only returns candidate variants where the supporting evidence has changed since the previous round, such as updates to ClinVar or PanelApp Australia annotations. As a result, later cycles returned an average of five new candidate variants per 1,000 cases. The average time between new knowledge becoming publicly available and a new diagnosis identified was 32 days, with the shortest being 1 day. Examples illustrating timelines to diagnosis from reanalysis are presented in Fig. 3d, and examples illustrating sources of diagnoses and clinical impact are presented in Table 2. Over 50 additional family members have had cascade testing based on reanalysis results to date, clarifying risk and the need for surveillance and treatment, particularly in cardiac disorders.Table 2Illustrative cases: sources of new diagnoses and clinical impact (gene names in italics)Clinical features and details of original testAutomated reanalysis result and reason for new diagnosisClinical impactAdult with bone pain since childhood; extensive family history. Assessed by multiple specialists; discordant opinions about possible diagnosis of hypophosphataemic rickets.Singleton exome with analysis limited to Hypophosphataemic Rickets panel.PHEXNM_000444.6:c.*231 A > Guntranslated region variant, linked marker for intragenic pathogenic duplication.Noncoding variant filtered out in original analysis. Detected due to new ClinVar entries. Duplication confirmed and reported as pathogenic.Diagnosis of hypophosphataemic rickets confirmed, ending diagnostic odyssey lasting decades.Additional seven family members tested for this treatable condition.Adult with hypertrophic cardiomyopathy.Singleton genome with analysis limited to Hypertrophic Cardiomyopathy panel.FHOD3NM_001281740.3(FHOD3):c.1287-243_1721+917delReported as variant of uncertain significance in original analysis.Improved understanding of mechanism of disease resulted in upgrade to pathogenic.Enabled cascade testing within family to identify relatives at risk.Adult, sudden unexpected death. Hypertrophic cardiomyopathy on postmortem.Singleton exome with analysis limited to Hypertrophic Cardiomyopathy panel.Dual diagnosis:PTPN11NM_002834.5:c.188 A > G p.Tyr63CysClinical features of Noonan syndrome not recognized. Gene not included in panel.MYBPC3NM_000256.3:c.1224-52 G > ADeep intronic variant filtered out in original analysis. Detected due to new ClinVar entries.Provided closure to family and enabled cascade testing to identify relatives at risk.Two additional individuals with this intronic MYBPC3 variant detected in cohort.Child with severe neurodevelopmental disorder.Trio genome sequencing with comprehensive analysis.RNU4-2NR_003137.2:n.64_65insTNew gene−disease association.Gene added to PanelApp Australia, and this recurrent variant was added to ClinVar based on preprint.Three additional individuals diagnosed in cohort with same variant, ending diagnostic odysseys and enabling families to connect.Reassurance provided regarding low recurrence risk in future pregnancies.Young adult with polycystic kidney disease.Singleton genome sequencing with analysis limited to Cystic Kidney disease panel.46,XX,t(4;7)(q22;p14)Transects PKD2Detected with improved SV calling.Confirmation of diagnosis enabled counseling regarding prognosis; cascade testing in family members; and reproductive counseling regarding translocation.Adult with polycystic kidney disease.Singleton genome sequencing with analysis limited to Cystic Kidney disease panel.IFT140NM_014714.4:c.1147 C > T p.Gln383TerNew gene−disease relationship.Bi-allelic variants in this gene previously established as a cause of multisystem ciliopathy.Mono-allelic variants newly established as a cause of slowly progressive, isolated polycystic kidney disease.Confirmation of diagnosis enabled counseling regarding prognosis: end-stage renal disease unlikely.Seven additional individuals diagnosed in cohort with LOF variants in IFT140.Adult with aortic dissection.Singleton exome with analysis limited to Aortopathy panel.COL3A1, COL5A2NC_000002.12:g.188971952_189077786delDetected by CNV analysis of exome data.Provided closure regarding cause of death.Enabled screening in five first-degree relatives. Table 2Illustrative cases: sources of new diagnoses and clinical impact (gene names in italics)Clinical features and details of original testAutomated reanalysis result and reason for new diagnosisClinical impactAdult with bone pain since childhood; extensive family history. Assessed by multiple specialists; discordant opinions about possible diagnosis of hypophosphataemic rickets.Singleton exome with analysis limited to Hypophosphataemic Rickets panel.PHEXNM_000444.6:c.*231 A > Guntranslated region variant, linked marker for intragenic pathogenic duplication.Noncoding variant filtered out in original analysis. Detected due to new ClinVar entries. Duplication confirmed and reported as pathogenic.Diagnosis of hypophosphataemic rickets confirmed, ending diagnostic odyssey lasting decades.Additional seven family members tested for this treatable condition.Adult with hypertrophic cardiomyopathy.Singleton genome with analysis limited to Hypertrophic Cardiomyopathy panel.FHOD3NM_001281740.3(FHOD3):c.1287-243_1721+917delReported as variant of uncertain significance in original analysis.Improved understanding of mechanism of disease resulted in upgrade to pathogenic.Enabled cascade testing within family to identify relatives at risk.Adult, sudden unexpected death. Hypertrophic cardiomyopathy on postmortem.Singleton exome with analysis limited to Hypertrophic Cardiomyopathy panel.Dual diagnosis:PTPN11NM_002834.5:c.188 A > G p.Tyr63CysClinical features of Noonan syndrome not recognized. Gene not included in panel.MYBPC3NM_000256.3:c.1224-52 G > ADeep intronic variant filtered out in original analysis. Detected due to new ClinVar entries.Provided closure to family and enabled cascade testing to identify relatives at risk.Two additional individuals with this intronic MYBPC3 variant detected in cohort.Child with severe neurodevelopmental disorder.Trio genome sequencing with comprehensive analysis.RNU4-2NR_003137.2:n.64_65insTNew gene−disease association.Gene added to PanelApp Australia, and this recurrent variant was added to ClinVar based on preprint.Three additional individuals diagnosed in cohort with same variant, ending diagnostic odysseys and enabling families to connect.Reassurance provided regarding low recurrence risk in future pregnancies.Young adult with polycystic kidney disease.Singleton genome sequencing with analysis limited to Cystic Kidney disease panel.46,XX,t(4;7)(q22;p14)Transects PKD2Detected with improved SV calling.Confirmation of diagnosis enabled counseling regarding prognosis; cascade testing in family members; and reproductive counseling regarding translocation.Adult with polycystic kidney disease.Singleton genome sequencing with analysis limited to Cystic Kidney disease panel.IFT140NM_014714.4:c.1147 C > T p.Gln383TerNew gene−disease relationship.Bi-allelic variants in this gene previously established as a cause of multisystem ciliopathy.Mono-allelic variants newly established as a cause of slowly progressive, isolated polycystic kidney disease.Confirmation of diagnosis enabled counseling regarding prognosis: end-stage renal disease unlikely.Seven additional individuals diagnosed in cohort with LOF variants in IFT140.Adult with aortic dissection.Singleton exome with analysis limited to Aortopathy panel.COL3A1, COL5A2NC_000002.12:g.188971952_189077786delDetected by CNV analysis of exome data.Provided closure regarding cause of death.Enabled screening in five first-degree relatives. Illustrative cases: sources of new diagnoses and clinical impact (gene names in italics) Adult with bone pain since childhood; extensive family history. Assessed by multiple specialists; discordant opinions about possible diagnosis of hypophosphataemic rickets. Singleton exome with analysis limited to Hypophosphataemic Rickets panel. PHEX NM_000444.6:c.*231 A > G untranslated region variant, linked marker for intragenic pathogenic duplication. Noncoding variant filtered out in original analysis. Detected due to new ClinVar entries. Duplication confirmed and reported as pathogenic. Diagnosis of hypophosphataemic rickets confirmed, ending diagnostic odyssey lasting decades. Additional seven family members tested for this treatable condition. Adult with hypertrophic cardiomyopathy. Singleton genome with analysis limited to Hypertrophic Cardiomyopathy panel. FHOD3 NM_001281740.3(FHOD3):c.1287-243_1721+917del Reported as variant of uncertain significance in original analysis. Improved understanding of mechanism of disease resulted in upgrade to pathogenic. Adult, sudden unexpected death. Hypertrophic cardiomyopathy on postmortem. Singleton exome with analysis limited to Hypertrophic Cardiomyopathy panel. Dual diagnosis: PTPN11 NM_002834.5:c.188 A > G p.Tyr63Cys Clinical features of Noonan syndrome not recognized. Gene not included in panel. MYBPC3 NM_000256.3:c.1224-52 G > A Deep intronic variant filtered out in original analysis. Detected due to new ClinVar entries. Provided closure to family and enabled cascade testing to identify relatives at risk. Two additional individuals with this intronic MYBPC3 variant detected in cohort. Child with severe neurodevelopmental disorder. Trio genome sequencing with comprehensive analysis. RNU4-2 NR_003137.2:n.64_65insT New gene−disease association. Gene added to PanelApp Australia, and this recurrent variant was added to ClinVar based on preprint. Three additional individuals diagnosed in cohort with same variant, ending diagnostic odysseys and enabling families to connect. Reassurance provided regarding low recurrence risk in future pregnancies. Young adult with polycystic kidney disease. Singleton genome sequencing with analysis limited to Cystic Kidney disease panel. 46,XX,t(4;7)(q22;p14) Transects PKD2 Detected with improved SV calling. Adult with polycystic kidney disease. Singleton genome sequencing with analysis limited to Cystic Kidney disease panel. IFT140 NM_014714.4:c.1147 C > T p.Gln383Ter New gene−disease relationship. Bi-allelic variants in this gene previously established as a cause of multisystem ciliopathy. Mono-allelic variants newly established as a cause of slowly progressive, isolated polycystic kidney disease. Confirmation of diagnosis enabled counseling regarding prognosis: end-stage renal disease unlikely. Seven additional individuals diagnosed in cohort with LOF variants in IFT140. Adult with aortic dissection. Singleton exome with analysis limited to Aortopathy panel. COL3A1, COL5A2 NC_000002.12:g.188971952_189077786del Detected by CNV analysis of exome data. Provided closure regarding cause of death. Enabled screening in five first-degree relatives. Computational and resource requirements Talos is designed to run on a single machine or virtual machine and is readily deployable on both local high-performance computing (HPC) clusters and cloud environments. All analyses described in this study were performed on a single virtual machine with 16 cores and 64 GB of memory. On Google Cloud Platform, the one-time cost of running the variant annotation workflow for 1,000 genomes from single-sample variant call formats (VCFs) was $11.25 USD (Extended Data Table 5). Running the variant filtering and prioritization workflow on the same cohort cost $0.21 AUD per run, equivalent to $1.65 USD annually if reanalysis were performed monthly. Extended Data Table 5 provides detailed cost estimates for cohorts ranging from five to 1,000 genomes. Discussion We introduce Talos, a publicly available tool for automated reanalysis of genomic data in research and clinical settings. Given that clinical and laboratory resources are the principal barrier to systematic reanalysis at scale16, we optimized Talos for high specificity to increase scalability, acceptability and ease of adoption and conducted performance evaluation in two independent validation cohorts, totaling 1,089 probands. When deployed in a large, unselected cohort of previously undiagnosed individuals with rare disease, Talos identified new diagnoses in 5.1% while returning a median of 1.2 candidate variants per family on initial analysis. This 5.1% yield is cohort specific and will vary across settings depending on case mix, prior workup and time since initial analysis. Variant burden decreased markedly on iterative cycles, addressing sustainability concerns while shortening the time between new information becoming publicly available and diagnosis. These findings demonstrate the feasibility of delivering frequent, systematic reanalysis at scale. Consistent with previous manual reanalysis studies5,9, just over half of new diagnoses were due to new information about gene−disease and variant−disease relationships, highlighting the critical importance of rapidly and openly sharing knowledge in improving diagnostic outcomes30. Integrating new knowledge into automated pipelines requires careful balancing between quality of evidence against sensitivity and timeliness. We leveraged the PanelApp Australia platform, as the monthly frequency of updates11 supports rapid incorporation of emerging gene−disease relationships, with the median time from publication of new information to diagnosis in this cohort being just 32 days. The structured curation of evidence for gene−disease associations, including the use of standardized terms for mode of inheritance (MOI), further refined variant filtering, while automating virtual panel assignment to support phenotype matching further decreased variant burden. The use of ClinVar similarly poses challenges relating to quality of evidence and the dynamic nature of the database, with thousands of new entries added each year. Although efforts are underway to curate content and resolve discrepant ratings between submitters and ClinGen expert panels31,32, we elected to use the star rating system as a threshold that denotes entries with sufficient evidence to trigger manual review. Our ClinvArbitration module further downweights entries from outlier submitters and prioritizes variants where the balance of evidence had shifted toward pathogenicity over time. The module can be deployed independently—for example, to support the periodic reevaluation of variants identified through targeted sequencing, including in cancer predisposition testing, where the clinical implications of variant reclassification can be profound33,34. Notably, just under half of the diagnoses identified by Talos were potentially identifiable during the original analysis, highlighting the importance of cohort-level reanalysis in quality assurance and continuous improvement. Genomic analysis approaches continue to evolve. Overly restrictive clinical and laboratory practices based on virtual panels, quality metrics, definition of target regions and inheritance filters have repeatedly been identified as sources of missed diagnoses35–39. Furthermore, the integration of analysis for additional variant types, in particular CNVs and STRs, as part of clinically accredited pipelines is increasing40,41. Automated reanalysis of stored data from historical cohorts thus allows patients and research participants to benefit from current best-practice analysis methods, exemplified here by the large number of new diagnoses identified by CNV analysis on exome data. With scalable, systematic reanalysis now technically feasible, questions of policy, professional practice guidance and funding models must be addressed. Although reanalysis is supported at a policy level13–15, clinical and laboratory practice based on manual models is highly variable, occurring at low volumes and resulting in inequity of access16,17. Automated models offer a consistent and efficient process to identify a large proportion of additional diagnoses over time17. The process can be communicated clearly to clinicians and, in turn, to patients and families, setting expectations, including as part of consent and result return discussions. Although many consent forms and supporting documents refer to result reinterpretation over time16,42, clinicians, laboratories and families would benefit from the clarity on timing, initiation and process of reanalysis that an automated, cohort-based approach brings43. Although the automated model described here scales, simplifies and standardizes reanalysis17, skilled manual review and iterative communication between laboratories and referring clinicians44 remain central to responsible implementation16. The performance evaluation against manual analysis and benchmarking against other automated analysis tools presented here will aid development of evidence-based professional guidance on reanalysis. Automated and manual approaches should be considered complementary: automated pipelines such as Talos enable systematic, frequent, low-burden reanalysis across large cohorts, whereas manual reanalysis remains essential in specific contexts such as evolving clinical features and needs. Beyond clinical considerations, automated reanalysis models raise implementation questions about infrastructure and bioinformatics capability. Cohort-based models rely on economies of scale to optimize benefit. However, legal and logistical barriers may prevent centralized data aggregation, limiting feasibility in some jurisdictions. This may disadvantage underresourced laboratories and worsen inequitable access. Health economic evaluations will be critical to inform funding models and policy decisions. We recognize several important limitations to our approach. Iterative reanalysis was performed only on individuals who remained undiagnosed or partially diagnosed after their initial test. A small but significant number of individuals who initially receive a diagnosis have the diagnosis retracted over time as gene−disease and variant−disease associations are refuted. Automated reanalysis pipelines can potentially have an important role to play in ensuring accuracy of diagnoses over time, including identification of new information that results in variant downgrades. At present, Talos reappraises genomic data using only the clinical information provided at the time of the original test request. However, clinical features can evolve, and automating updates to clinical information—for example, by directly mining electronic medical records—may improve performance. Other areas that may benefit from further automation include direct literature mining for new information associating variants and genes to disease45. Although Talos is currently optimized for diagnostic use, it can be modified to support discovery research—for example, by using lists of candidate genes to identify additional cases to support novel and emerging gene−disease associations. Deployment in additional settings that require scalability, such as primary diagnostic analysis and population screening, will be important to assess wider user needs and operational performance. Talos is also compatible with long-read sequencing data provided in VCF format, offering a further avenue for future application. In summary, we demonstrate the feasibility of iterative automated reanalysis in rare disease, scaled to thousands of datasets, and provide an open-source tool to enable broad adoption. Our work forms a solid foundation for the integration of machine learning and artificial intelligence approaches and has potential for broad applicability to support timely and equitable genomic care. Methods Ethics This study is governed and administered by the Murdoch Children’s Research Institute in Melbourne, Australia. Reanalysis of existing research cohorts using Talos was performed with Human Research Ethics Committee (HREC) approval (HREC/77735/RCHM-2021). The Australian Genomics rare disease cohorts, including the ACG study, have HREC approval (HREC/16/MH/251). The RGP cohort has Human Institutional Review Board approval by Mass General Brigham (protocol number 2016P001422). Participants and/or their guardians provided consent for participation in research, after genetic counseling. Reanalysis of clinical cohorts was performed as part of ongoing diagnostic care and laboratory quality assurance processes, and consent was obtained as part of clinical care. Participants did not receive compensation. Design and implementation of Talos Talos is an open-source, cohort-based variant filtering and prioritization tool designed for scalable, automated and iterative reanalysis of genomic data in rare disease (https://github.com/populationgenomics/talos). It comprises two core workflows:Variant annotation workflow: This initial workflow performs one-time processing of variant data, annotating SNVs, indels and SVs using a comprehensive set of static data sources. These include transcript consequences, population allele frequencies, in silico pathogenicity predictions and gene models. The output is stored in a Hail MatrixTable format for efficient downstream processing.Variant filtering and prioritization workflow: This workflow is designed to be run iteratively and overlays the static annotations, with dynamic annotations from ClinVar and PanelApp Australia updated at runtime. It performs variant filtering, applies a configurable set of variant logic modules, evaluates variant segregation based on expected MOI and prioritizes variants based on phenotype matching. This workflow generates structured reports suitable for clinical review or research interpretation. Variant annotation workflow: This initial workflow performs one-time processing of variant data, annotating SNVs, indels and SVs using a comprehensive set of static data sources. These include transcript consequences, population allele frequencies, in silico pathogenicity predictions and gene models. The output is stored in a Hail MatrixTable format for efficient downstream processing. Variant annotation workflow: This initial workflow performs one-time processing of variant data, annotating SNVs, indels and SVs using a comprehensive set of static data sources. These include transcript consequences, population allele frequencies, in silico pathogenicity predictions and gene models. The output is stored in a Hail MatrixTable format for efficient downstream processing. Variant filtering and prioritization workflow: This workflow is designed to be run iteratively and overlays the static annotations, with dynamic annotations from ClinVar and PanelApp Australia updated at runtime. It performs variant filtering, applies a configurable set of variant logic modules, evaluates variant segregation based on expected MOI and prioritizes variants based on phenotype matching. This workflow generates structured reports suitable for clinical review or research interpretation. Variant filtering and prioritization workflow: This workflow is designed to be run iteratively and overlays the static annotations, with dynamic annotations from ClinVar and PanelApp Australia updated at runtime. It performs variant filtering, applies a configurable set of variant logic modules, evaluates variant segregation based on expected MOI and prioritizes variants based on phenotype matching. This workflow generates structured reports suitable for clinical review or research interpretation. Both workflows are implemented using Nextflow and packaged in a Docker environment for reproducibility. Analyses are executed at the family level but benefit from optimizations when run at cohort scale. Talos input requirements Talos accepts the following inputs:SNV/indel variant data (mandatory): SNV and indel calls can be provided either as a set of single-sample VCF files or, preferably, as a multisample VCF (msVCF) containing variant calls for all individuals in the cohort to be analyzed.SV/CNV variant data (optional): SV/CNV calls can be provided either as a set of single-sample VCF files or, preferably, as an msVCF containing variant calls for all individuals in the cohort to be analyzed.Pedigree information (mandatory): provided in PED formatPhenotype data (optional): encoded as Human Phenotype Ontology (HPO) terms per individual; used to refine gene panels and prioritize variantsConfiguration file (mandatory): a TOML configuration file specifying parameters for panel selection, phenotype matching and variant filtering SNV/indel variant data (mandatory): SNV and indel calls can be provided either as a set of single-sample VCF files or, preferably, as a multisample VCF (msVCF) containing variant calls for all individuals in the cohort to be analyzed. SNV/indel variant data (mandatory): SNV and indel calls can be provided either as a set of single-sample VCF files or, preferably, as a multisample VCF (msVCF) containing variant calls for all individuals in the cohort to be analyzed. SV/CNV variant data (optional): SV/CNV calls can be provided either as a set of single-sample VCF files or, preferably, as an msVCF containing variant calls for all individuals in the cohort to be analyzed. SV/CNV variant data (optional): SV/CNV calls can be provided either as a set of single-sample VCF files or, preferably, as an msVCF containing variant calls for all individuals in the cohort to be analyzed. Pedigree information (mandatory): provided in PED format Pedigree information (mandatory): provided in PED format Phenotype data (optional): encoded as Human Phenotype Ontology (HPO) terms per individual; used to refine gene panels and prioritize variants Phenotype data (optional): encoded as Human Phenotype Ontology (HPO) terms per individual; used to refine gene panels and prioritize variants Configuration file (mandatory): a TOML configuration file specifying parameters for panel selection, phenotype matching and variant filtering Configuration file (mandatory): a TOML configuration file specifying parameters for panel selection, phenotype matching and variant filtering As Talos implements a full suite of annotation and filtering logic, input variant data should not be prefiltered beyond standard quality control. Providing broadly called, quality-control-passed variants ensures that Talos’s own filtering criteria are applied consistently and without inadvertent loss of clinically relevant calls. The variant annotation workflow needs to be run only once per dataset. For subsequent reanalyses, only the variant prioritization workflow is rerun, using updated annotation sources. Additional data resources required by Talos Talos relies on multiple publicly available data resources for variant annotation. The first time the Talos workflow is executed, any resources not detected will be automatically downloaded and made available for subsequent Talos runs.PanelApp Australia: gene−disease associations and modes of inheritance. PanelApp is a publicly available knowledge base that allows virtual gene panels related to human disorders to be created, stored and queried. PanelApp Australia (https://panelapp-aus.org/) is an instance of PanelApp used by Australian diagnostic laboratories. Talos accesses PanelApp Australia at runtime via API to gene−disease associations and MOI descriptions. PanelApp Australia was selected as the source of gene−disease association information in Talos as it is updated each month to incorporate novel gene−disease associations published in the literature11.ClinVar via ClinvArbitration: The ClinVar database46 provides a catalog of variants previously reported as P/LP. Talos uses the ClinvArbitration tool to aggregate ClinVar submissions for each variant, prioritizing recent, high-confidence submissions. ClinvArbitration can be run locally, or monthly data releases can be downloaded from the ClinvArbitration website (https://github.com/populationgenomics/ClinvArbitration).Gene-to-phenotype associations: Gene−phenotype annotation database and ontology structure for phenotype matching are accessed from the HPO project47 and the Monarch Initiative48.Genome Aggregation Database (gnomAD) population allele frequencies: Talos uses population allele frequencies from gnomAD v.4 preprocessed into an echtvar index file for rapid annotation49.In silico missense variant effect prediction: Talos uses in silico effect prediction scores for missense variants generated by AlphaMissense50. PanelApp Australia: gene−disease associations and modes of inheritance. PanelApp is a publicly available knowledge base that allows virtual gene panels related to human disorders to be created, stored and queried. PanelApp Australia (https://panelapp-aus.org/) is an instance of PanelApp used by Australian diagnostic laboratories. Talos accesses PanelApp Australia at runtime via API to gene−disease associations and MOI descriptions. PanelApp Australia was selected as the source of gene−disease association information in Talos as it is updated each month to incorporate novel gene−disease associations published in the literature11. PanelApp Australia: gene−disease associations and modes of inheritance. PanelApp is a publicly available knowledge base that allows virtual gene panels related to human disorders to be created, stored and queried. PanelApp Australia (https://panelapp-aus.org/) is an instance of PanelApp used by Australian diagnostic laboratories. Talos accesses PanelApp Australia at runtime via API to gene−disease associations and MOI descriptions. PanelApp Australia was selected as the source of gene−disease association information in Talos as it is updated each month to incorporate novel gene−disease associations published in the literature11. ClinVar via ClinvArbitration: The ClinVar database46 provides a catalog of variants previously reported as P/LP. Talos uses the ClinvArbitration tool to aggregate ClinVar submissions for each variant, prioritizing recent, high-confidence submissions. ClinvArbitration can be run locally, or monthly data releases can be downloaded from the ClinvArbitration website (https://github.com/populationgenomics/ClinvArbitration). ClinVar via ClinvArbitration: The ClinVar database46 provides a catalog of variants previously reported as P/LP. Talos uses the ClinvArbitration tool to aggregate ClinVar submissions for each variant, prioritizing recent, high-confidence submissions. ClinvArbitration can be run locally, or monthly data releases can be downloaded from the ClinvArbitration website (https://github.com/populationgenomics/ClinvArbitration). Gene-to-phenotype associations: Gene−phenotype annotation database and ontology structure for phenotype matching are accessed from the HPO project47 and the Monarch Initiative48. Gene-to-phenotype associations: Gene−phenotype annotation database and ontology structure for phenotype matching are accessed from the HPO project47 and the Monarch Initiative48. Genome Aggregation Database (gnomAD) population allele frequencies: Talos uses population allele frequencies from gnomAD v.4 preprocessed into an echtvar index file for rapid annotation49. Genome Aggregation Database (gnomAD) population allele frequencies: Talos uses population allele frequencies from gnomAD v.4 preprocessed into an echtvar index file for rapid annotation49. In silico missense variant effect prediction: Talos uses in silico effect prediction scores for missense variants generated by AlphaMissense50. In silico missense variant effect prediction: Talos uses in silico effect prediction scores for missense variants generated by AlphaMissense50. Talos variant annotation workflow This is a lightweight workflow designed for rapid and efficient annotation of variants from large cohorts. It is intended to be run once per cohort and label each variant with predicted transcript consequence and annotations from sources that do not evolve rapidly. Transcript consequence annotations are provided by BCFtools/csq for SNV/indels and SVAnnotate for CNV/SVs using Ensembl v.133 gene models. Each variant is labeled with observed population allele frequencies from gnomAD v.4 using echtvar. The workflow outputs all variants in Hail MatrixTable format for efficient processing in the variant prioritization workflow. Talos variant filtering and prioritization workflow The Talos variant prioritization workflow consists of three main stages:For each family, determine the set of genes eligible for consideration.For each variant within an eligible gene, apply each Talos variant logic module to identify those that could plausibly be pathogenic.For each gene, consider if a variant (or variant pair) identified in step 2 segregates with the expected MOI known for that disorder. For each family, determine the set of genes eligible for consideration. For each family, determine the set of genes eligible for consideration. For each variant within an eligible gene, apply each Talos variant logic module to identify those that could plausibly be pathogenic. For each variant within an eligible gene, apply each Talos variant logic module to identify those that could plausibly be pathogenic. For each gene, consider if a variant (or variant pair) identified in step 2 segregates with the expected MOI known for that disorder. For each gene, consider if a variant (or variant pair) identified in step 2 segregates with the expected MOI known for that disorder. The set of genes eligible for consideration is drawn from PanelApp Australia and, by default, will consist of the (currently) 4,427 genes on the ‘Mendeliome’ gene panel—that is, all genes associated with Mendelian disorders with exception of a panel of 120 genes commonly associated with incidental findings (the ‘Incidentalome’ panel). Talos will only consider genes annotated with a ‘Green’ level of supporting evidence and approved for diagnostic use. The default panel used by Talos is configurable, allowing a user to select either a more permissive base panel (for example, to also include all genes commonly associated with incidental findings) or a more selective base panel (for example, to restrict analysis to a virtual panel of genes associated with a specific disorder). The default gene panel source (PanelApp Australia) is configurable and can be substituted with local resources where preferred. If a patient has an HPO term consistent with one or more disease-group-specific panels, the union of genes on matched panel(s) and the Mendeliome are presented for consideration for a given family. In the second step of the workflow, Talos identifies potentially pathogenic variants through a modular system of independently executed variant logic modules (more detail below). Each module implements a specific rule-based decision tree representing a distinct set of evidence that indicates that a variant should be considered as potentially causative (for example, predicted loss-of-function (LOF) variants, de novo variants or variants previously reported as (likely) pathogenic in ClinVar). In the third stage of the variant prioritization workflow, all variants in each eligible gene are evaluated against the PanelApp Australia-assigned MOIs of the gene. Talos implements decision trees to assess MOI constraints using the following logic:Autosomal dominant: present in all affected and no unaffected individualsAutosomal recessive (homozygous): present in all affected, absent in all unaffectedAutosomal recessive (compound heterozygous): two variants in a gene, only one or neither in unaffectedX-linked dominant or recessive: assessed in context of participant sex and zygosity Autosomal dominant: present in all affected and no unaffected individuals Autosomal dominant: present in all affected and no unaffected individuals Autosomal recessive (homozygous): present in all affected, absent in all unaffected Autosomal recessive (homozygous): present in all affected, absent in all unaffected Autosomal recessive (compound heterozygous): two variants in a gene, only one or neither in unaffected Autosomal recessive (compound heterozygous): two variants in a gene, only one or neither in unaffected X-linked dominant or recessive: assessed in context of participant sex and zygosity X-linked dominant or recessive: assessed in context of participant sex and zygosity Partial penetrance mode can be optionally enabled to relax strict family-level filters. Talos variant logic modules Talos variant logic modules each implement a specific rule-based decision tree representing a distinct set of evidence that indicates that a variant should be considered as potentially causative. Modules are applied to each variant independently. Variants that meet a moduleʼs criteria are annotated with the corresponding label. Module labels are not mutually exclusive; variants may carry multiple labels. Variants that do not match any module are excluded from downstream analyses. The set of variant logic modules applied in a given analysis is configurable. Each module can be assigned either ‘primary’ or ‘supporting’ significance within the run configuration file. Variants annotated by a primary module are eligible to be reported as candidates in both dominant and recessive disease genes. By contrast, variants annotated only by supporting modules can contribute to diagnoses in recessive genes but only as a second hit alongside a variant with primary-level evidence. Homozygous variants annotated only with supporting labels are excluded from analysis. Standard variant logic modules used by Talos are as follows:ClinVar P/LP: labels variants reported as P/LP by ClinvArbitration aggregate classification score based on in ClinVar submissionsHigh impact: labels variants with a high-impact protein consequence on at least one transcript. By default, high-impact protein consequences are defined as frameshift, splice_acceptor, splice_donor, start_lost, stop_gained, stop_lost or transcript_ablation as determined by BCFtools/csq51.De novo: labels variants that are de novo in an affected individual and have a high-impact or missense protein consequence or are present in a transcript with biotype ‘snRNA’PM5: labels missense variants that lead to a different amino acid substitution in the same codon as previously reported P/LP variants. This module uses the same ClinvArbitration aggregate classifications used in the ClinVar P/LP module.SV LOF: labels SV and CNV variants with a predicted consequence of LOF by the SVAnnotate tool52In silico: labels variants that are predicted to be ‘likely_pathogenic’ based on analysis by AlphaMissense50. By default, this module is configured as providing ‘supporting’ significance only. When analyzing trios, use of ‘in silico’ as a primary category can result in increased Talos sensitivity to detect causative missense variants at a moderate cost to specificity. However, if applied to singleton analysis, this configuration results in substantial increase in noise. ClinVar P/LP: labels variants reported as P/LP by ClinvArbitration aggregate classification score based on in ClinVar submissions ClinVar P/LP: labels variants reported as P/LP by ClinvArbitration aggregate classification score based on in ClinVar submissions High impact: labels variants with a high-impact protein consequence on at least one transcript. By default, high-impact protein consequences are defined as frameshift, splice_acceptor, splice_donor, start_lost, stop_gained, stop_lost or transcript_ablation as determined by BCFtools/csq51. High impact: labels variants with a high-impact protein consequence on at least one transcript. By default, high-impact protein consequences are defined as frameshift, splice_acceptor, splice_donor, start_lost, stop_gained, stop_lost or transcript_ablation as determined by BCFtools/csq51. De novo: labels variants that are de novo in an affected individual and have a high-impact or missense protein consequence or are present in a transcript with biotype ‘snRNA’ De novo: labels variants that are de novo in an affected individual and have a high-impact or missense protein consequence or are present in a transcript with biotype ‘snRNA’ PM5: labels missense variants that lead to a different amino acid substitution in the same codon as previously reported P/LP variants. This module uses the same ClinvArbitration aggregate classifications used in the ClinVar P/LP module. PM5: labels missense variants that lead to a different amino acid substitution in the same codon as previously reported P/LP variants. This module uses the same ClinvArbitration aggregate classifications used in the ClinVar P/LP module. SV LOF: labels SV and CNV variants with a predicted consequence of LOF by the SVAnnotate tool52 SV LOF: labels SV and CNV variants with a predicted consequence of LOF by the SVAnnotate tool52 In silico: labels variants that are predicted to be ‘likely_pathogenic’ based on analysis by AlphaMissense50. By default, this module is configured as providing ‘supporting’ significance only. When analyzing trios, use of ‘in silico’ as a primary category can result in increased Talos sensitivity to detect causative missense variants at a moderate cost to specificity. However, if applied to singleton analysis, this configuration results in substantial increase in noise. In silico: labels variants that are predicted to be ‘likely_pathogenic’ based on analysis by AlphaMissense50. By default, this module is configured as providing ‘supporting’ significance only. When analyzing trios, use of ‘in silico’ as a primary category can result in increased Talos sensitivity to detect causative missense variants at a moderate cost to specificity. However, if applied to singleton analysis, this configuration results in substantial increase in noise. Talos phenotype matching Talos uses patient-supplied HPO terms to highlight variants whose genes are plausibly related to the observed phenotype. By default, Talos tests whether any patient HPO term is similar to the HPO terms associated with a gene or with a disease-specific gene panel, and it flags such genes as phenotype matched in the report. This signal is advisory by default. Talos does not exclude variants on the basis of phenotype but surfaces potential matches to draw attention for review. Users may optionally enable a strict phenotype filter that retains only variants in phenotype-matched genes or panels, which increases specificity and reduces reviewer burden. The sensitivity tradeoff for this mode will depend on the completeness and accuracy of patient phenotyping and on how comprehensively HPO terms capture the phenotypic scope of each gene. Talos uses three strategies to identify genes associated with phenotypes observed in a specific patient. Each strategy is intentionally inclusive, such that if one or more of a patient’s phenotypes is matched to a given gene, then that gene is considered a ‘phenotype match’ for the family. Patient phenotype to gene matching Direct fuzzy matching of patient HPO terms with the HPO associated with a specific gene is performed using the Semsimian48 tool to calculate pairwise semantic similarity scores. If any pair of HPO terms has an Ancestor Information Content score above the configurable threshold (default = 14), the gene is considered a phenotype match for the patient. Patient phenotype to gene panel matching PanelApp Australia contains 261 gene panels, each related to a specific disease or group of disorders. These panels have been annotated with one or more high-level HPO terms representative of the group of disorders. For each gene panel, if one or more panel HPO terms either match or are ancestors of a patient HPO term, then the patient is considered a phenotype match for that panel and all of the genes in that panel. Cohort to gene panel matching If the cohort being analyzed consists of patients of a specific disease group (for example, mitochondrial disorders), one or more gene panels can be applied as a ‘phenotype match’ to all families. Allele frequency filtering The Talos variant prioritization workflow filters variants based on their observed allele counts or frequencies. Thresholds are configurable and tailored to the expected MOI. Two sources of allele frequency are used in filtering: population allele frequencies from gnomAD v.4 and cohort allele frequencies calculated from the dataset under analysis. When the cohort is sufficiently large, cohort-specific frequencies are useful for removing common pipeline-specific artifacts that may not be present in gnomAD. To support more permissive filtering, separate thresholds can be specified for variants previously reported in ClinVar as P/LP. ClinVar summarization using ClinvArbitration The ClinVar database maintains an aggregate classification score for each variant based on the classifications provided by each submitter. However, due to the consensus-based logic used to determine the aggregate classification, many variants with a clear majority of P/LP submissions are labeled as ‘Conflicting classifications of pathogenicity’ due to a minority of ‘Uncertain’ submissions. To generate ClinVar variant aggregate classifications better suited to the needs of automated variant prioritization, we developed an alternative aggregation approach implemented in a tool called ClinvArbitration. The ClinvArbitration algorithm uses a modified ‘majority wins’ strategy to resolve conflicts. Where more recent submissions exist, ClinvArbitration will also disregard ClinVar submissions made before the widespread use of modern ACMG variant curation guidelines53. ClinvArbitration allows for variant submissions from specific submitters to be excluded during processing. To prevent circular logic during the evaluation of Talos, this feature was used to exclude any ClinVar submissions made by the laboratories that analyzed the evaluation cohorts in this study. Specifically, the submitters ‘Broad Center for Mendelian Genomics (Broad Institute of MIT and Harvard), Broad CMG’ and ‘Laboratory for Molecular Medicine; Mass General Brigham Personalized Medicine’ were excluded from ClinvArbitration summaries used to analyze the RGP cohort, and the submitter ‘Victorian Clinical Genetics Services, Murdoch Children’s Research Institute’ was excluded from ClinvArbitration summaries used to analyze all other cohorts. ClinvArbitration is run for each monthly release of ClinVar. The open-source software and monthly data releases are available from https://github.com/populationgenomics/ClinvArbitration. Variant calling SNV calling was performed using GATK variant calling best-practices workflows. For the RGP cohort, genome sequencing data were processed through a pipeline based on Picard, using base quality score recalibration and local realignment at known indels. The BWA aligner54 was used for mapping reads to the human genome build 38. SNVs and indels are jointly called across all samples using the GATK HaplotypeCaller package55 v.4.0. Default filters were applied to SNV and indel calls using the GATK Variant Quality Score Recalibration (VQSR) approach. Genome sequencing data were additionally processed to detect SVs using GATK-SV52. All remaining cohorts were reprocessed from FASTQs by the Center for Population Genomics using DRAGMAP v.1.3.0 and GATK v.4.2.6.1. SV calling of genomes was performed using the GATK-SV structural variation discovery pipeline v.0.28.4. CNV calling in exomes was performed using GATK-gCNV v.4.5.0.0 (ref. 56). Exomiser comparison We benchmarked Talos against Exomiser (v.14, default configuration) on the same ACG trios with known SNV or indel diagnoses. Success for Talos was defined as the causative variant being returned under Talos default settings. Success for Exomiser was defined as the causative variant ranking within the top 1, top 5 or top 10 variants per family, plus an unconstrained ‘all’ condition. Because outcomes are paired within the same families, we compared tools using McNemarʼs exact test applied to the discordant pairs (Talos-only versus Exomiser-only) at each threshold, reporting two-sided P values. This test is appropriate for paired binary data and conditions on the total number of discordant outcomes, which controls for case-level heterogeneity and removes ties where both tools either succeed or fail. We used the exact version because the number of discordant pairs was modest. Tests that assume independent samples, such as two-sample proportion tests or Fisherʼs exact test, are not suitable for these paired comparisons. Talos outputs Due to the diverse workflows used for manual review and curation of variants across laboratories, the primary output of Talos is a single machine-readable file in JSON format. The output file contains a list of all variants meeting the criteria to be considered potentially causative, a summary of the evidence supporting each variant and metadata describing the configuration used to perform the analysis. The JSON output file can be transformed into a format suitable for integration into the curation workflow of a specific laboratory. The default Talos workflow contains a JSON to HTML report generator script as a reference implementation. Variant confirmation and reporting Variants found by the reanalysis but not identifiable by the original clinical laboratory using currently clinically accredited tools on existing short-read data were orthogonally confirmed by a range of methods. CNVs were generally confirmed by high-resolution chromosome single-nucleotide polymorphism (SNP) microarray (hrCMA, GDA-Cyto v.1; Illumina), if at least three informative probes were available. CNVs not identifiable by hrCMA and other complex variants (for example, inversions) were confirmed using long-read adaptive sampling on a Promethion 24 sequencer using R10.4.1 flow cells and V14 ligation and sequencing chemistry (Oxford Nanopore Technologies). In one case, insufficient remaining DNA necessitated long-range polymerase chain reaction (PCR). One reciprocal translocation was confirmed using clinical G-banded chromosome metaphase karyotyping. All variants identified as being likely diagnostic through manual review of Talos output, including those undergoing orthogonal validation, were forwarded to the relevant clinical laboratory for assessment. Variants were assessed using clinically accredited processes for variant classification and reporting, based on ACMG guidelines, and reported to the relevant clinician where a new diagnosis was identified. Statistics and reproducibility This study evaluated the performance of an automated reanalysis tool in pre-existing rare disease cohorts. No statistical method was used to predetermine sample size. Cohort sizes reflect the full set of eligible individuals available from each contributing study: the two validation cohorts (ACG, n = 401 probands; RGP, n = 688 families) comprised all individuals from the respective studies with genome or exome sequencing data of sufficient quality for reprocessing through the standardized pipeline; the prospective reanalysis cohort (n = 4,735) comprised all undiagnosed or partially diagnosed individuals meeting the inclusion criteria above. No data were excluded from the analyses; individuals harboring diagnostic variants in classes not reliably detected by the standardized pipeline (for example, STR expansions, low-level mosaic variants and retrotransposon insertions) were retained and are reported as missed diagnoses were applicable. The experiments were not randomized; the study used a benchmarking design in observational cohorts rather than a randomized comparison. The investigators were not blinded to allocation during experiments and outcome assessment; for the validation cohorts, prior diagnoses were known at the time of evaluation to enable benchmarking against manual analysis, although Talos itself is a deterministic, automated pipeline whose output is generated without investigator input. To support reproducibility, Talos is open-source and packaged within a fixed software environment (Docker, Nextflow), with all configuration files used in this study provided with the code (https://github.com/populationgenomics/talos). External annotation sources (PanelApp Australia, ClinVar and gnomAD) are versioned at runtime, enabling exact reproduction of any individual reanalysis cycle. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Online content Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41591-026-04477-5. Supplementary information Supplementary InformationTable 1 and Supplementary Text. Reporting Summary Table 1 and Supplementary Text. Reporting Summary