VarSage is a bioinformatics decision-support platform for interpretation of germline sequence variants. It is intended for exome or genome VCF/table outputs that have already passed through an upstream alignment, variant calling, and primary quality-control workflow. VarSage does not call variants from reads; it integrates existing variant calls with phenotype, inheritance, population, transcript, clinical-assertion, and computational evidence to produce a prioritization report.
The application is organized around three analytical contexts: Rare Disease - Singleton, carrier screening analysis, and Rare Disease - Trio. Across all contexts, VarSage treats the variant as a structured allele defined by chromosome, reference allele, alternate allele, selected sample genotype, and annotation fields. This deterministic coordinate model is deliberately kept separate from AI-generated summaries.
Single-proband analysis that ranks candidate alleles by phenotype fit, molecular consequence, rarity, inheritance compatibility, quality, and provisional evidence.
Independent partner review followed by same-gene comparison for recessive reproductive-risk signals and scope-specific carrier findings.
Proband-first prioritization followed by parental allele-state comparison for possible de novo, inherited, and compound-heterozygous models.
VarSage follows a layered interpretation architecture. First, submitted phenotypes and variants are normalized into controlled representations. Next, transcript and coordinate annotations are attached while preserving original source fields. The resulting candidate set is filtered for assay scope, quality, benign assertions, and population frequency. Finally, retained variants are ranked using phenotype similarity, inheritance compatibility, molecular consequence, rarity, clinical assertions, computational predictors, and provisional ACMG/AMP-oriented evidence.
The final report is not just a score table. It is an audit trail: which sample was analyzed, which file type was detected, which variants were removed, which databases matched, which HPO terms were used, which inheritance model was plausible, and which evidence remains incomplete. This is essential in clinical bioinformatics because the same numeric rank can arise from very different biological and evidentiary situations.
Use when an affected proband has WES/WGS variants and the primary question is phenotypic plausibility of candidate alleles.
Use when two partners are assessed for heterozygous variants in the same recessive disease gene or in configured carrier scopes.
Use when maternal and/or paternal variant files can inform de novo, inherited, or biallelic hypotheses in the proband.
The Rare Disease - Singleton workflow is a single-proband prioritization pipeline for suspected Mendelian disease. Its purpose is to reduce a large variant set to a biologically interpretable shortlist while preserving the evidentiary chain behind each retained candidate. The workflow explicitly distinguishes review priority from pathogenicity classification and separates variant-level evidence from case-level causality.
After ranking and optional synthesis, VarSage adds a cautious case-level decision spectrum. This summarizes whether the retained candidate set contains no positive candidate, an inconclusive candidate, a positive candidate requiring confirmation, or an externally confirmed cause. The decision is derived from structured evidence and report-review signals; it is not a molecular diagnosis.
Assigned when no retained candidate is available, or when the retained candidates do not contain pathogenic/likely-pathogenic or primary report-review evidence. This is not a statement that a genetic diagnosis is excluded.
Assigned when candidates remain important for review, but the current evidence is not strong enough for a cautious positive decision. The result should remain inconclusive or candidate-only until the uncertainty is resolved.
Assigned when a prioritized candidate has a pathogenic/likely-pathogenic provisional classification, a matching ClinVar-like assertion, or a primary report-review signal. Expert review, orthogonal confirmation, and relevant segregation or phase checks are still required.
This is a reporting endpoint for human interpretation, not an automatic VarSage decision. The tool never assigns it; it requires qualified expert review and laboratory confirmation where appropriate.
The main table orders variants by a composite score that combines phenotype fit, inheritance support, consequence, rarity, and quality.
Each row can expand to show annotation, filtering, phenotype overlap, literature links, and provisional ACMG evidence.
The workflow exports HTML, TSV, and Microsoft word results so the shortlist can be reviewed and shared outside the app.
The carrier-screening workflow is built for reproductive-risk review in a couple. It analyzes each partner independently, applies carrier-scope filtering, and then performs a gene-level comparison between retained heterozygous findings. This design prevents partner-specific evidence from being conflated while still making shared-gene signals immediately visible.
Each partner gets its own retained-variant summary, including panel genes, secondary findings, and X-chromosome review where relevant.
The couple-level table highlights genes where both partners still have retained heterozygous variants after filtering.
The workflow writes downloadable TSV, HTML, and Word outputs with separate findings for each partner and the couple summary.
The Rare Disease - Trio workflow extends proband prioritization with parental allele-state evidence. It is designed for family-aware interpretation: the proband is ranked first, and parental calls are then used to evaluate inheritance hypotheses. Missing parent files, low-quality parental calls, or absent coverage are treated as missing evidence rather than definitive evidence of absence.
The proband still gets a full Rare Disease - Singleton ranking table with phenotype, inheritance, and evidence context.
The trio summary separates possible de novo, compound-heterozygous, and causative candidate views from the parental comparisons.
Mother and father each receive their own structured summary, and the final exports include the trio-specific interpretation sections.
Single Variant Analysis looks up one small variant at a time without uploading a VCF or variant table. The user types a genomic locus, chooses the reference genome, and VarSage returns transcript consequences, coordinate-level evidence from the super database, protein 3D structure and 2D domain context, and cross-species conservation for that exact allele. It is intended for quick look-ups, for checking a variant reported elsewhere, and for reviewing a single candidate in more depth.
Open the Single Variant Analysis tab on the analysis page (sign-in required). Enter the variant as VCF-style genomic coordinates (1-based position, reference allele, alternate allele) and select hg38 / GRCh38 (default) or hg19 / GRCh37. The reference allele must match the selected build.
| Accepted form | Example |
|---|---|
| chromosome:position REF>ALT | chr17:7675088 C>T |
| Dash-separated | 17-7675088-C-T |
| Colon-separated | 17:7675088:C:T |
| Space-separated | 17 7675088 C T |
| Indel with anchor base | chr7:117559590 ATCT>A |
chr prefix is optional. Chromosomes 1-22, X, Y and MT (or M) are accepted.7,675,088.Gene, variant type (SNV, MNV, insertion, deletion or deletion-insertion), consequence and impact, transcript and exon/intron number, cDNA change, protein change, and dbSNP ID. Any warnings, such as an unavailable annotation resource, are listed here.
One row per transcript with effect, impact, exon/intron, HGVS c. and p., CDS and amino-acid position/length, and biotype. The default row is highlighted. Choosing another row in the Use column updates the summary and protein views for that transcript.
Max population allele frequency, ClinVar significance, REVEL, the automated InterVar classification, dbSNP ID, and the InterVar ACMG/AMP criteria met for this allele (pathogenic codes in red, benign codes in green). All super database fields expands the full record.
For a protein-coding change, the AlphaFold 3D structure with the residue projected on it and the 2D InterPro domain map are shown, using the same viewers described in Protein 3D Structure. Intronic, UTR, splice-region and intergenic variants have no residue to show.
A UCSC view of the region around the variant with phyloP and phastCons (100-vertebrate) tracks and a species-by-species alignment. The flanking window can be set to ±25, ±100, ±250 or ±500 bp.
Every gene-bearing result view in the web application uses the same on-demand GTEx panel. Rare Disease - Singleton ranked candidates, Carrier screening same-gene warnings and partner tables, and Rare Disease - Trio proband, inheritance-finding, compound-heterozygous, causative, parent-review, and literature-context tables all expose the same Expression control.
Opening the control retrieves the gene's GTEx Portal profile and shows median TPM by tissue, tissue ranking, expression breadth, tissue specificity, a distribution summary, phenotype-linked tissue context when a broad local mapping is available, a locally saved reviewer note, and a link to the GTEx gene page. Expression data are presented as biological plausibility context rather than as variant-classification evidence.
VarSage uses GTEx Portal API v2, with gtex_v8 and GENCODE v26 as the defaults. Expression is loaded only when a signed-in user opens a gene panel, and unavailable remote data does not interrupt the analysis.
The authenticated VarSage endpoint receives the selected gene and phenotype labels. VarSage sends the gene lookup to GTEx and performs phenotype-to-tissue matching locally; patient files, patient IDs, and clinical narrative are not sent to GTEx by this feature.
Expression is biological context, not an ACMG/AMP criterion and not a pathogenicity measure. It does not change filtering, ranking, classification, or reportability. Low bulk-tissue expression does not exclude a gene because cell type, developmental stage, transcript, and disease state matter.
VarSage uses the Human Phenotype Ontology (HPO) as the controlled vocabulary for clinical features. HPO terms are used for Rare Disease - Singleton ranking, phenotype-concern scope in carrier screening, Rare Disease - Trio proband interpretation, and gene-disease plausibility review. The system preserves the difference between observed findings, explicitly absent findings, and inferred ontology matching context.
The matcher recognizes HPO IDs, canonical names, synonyms, selected clinically equivalent aliases, punctuation-normalized terms, and common plural forms. HPO hierarchy is used when mapping phenotypes to genes: for example, Migraine (HP:0002076) is a child of Headache (HP:0002315), so a gene annotated with Headache can remain in scope for a patient with migraine. The extractor may add direct, clinically useful parent terms such as Headache to the editable suggestion list; users can remove any suggestion before submitting the case. Co-symptoms such as photophobia, nausea, visual impairment, or aura are not inferred unless the source text states them or the user confirms them. AI suggestions still require a valid ontology ID and supporting text, and explicit negation takes precedence.
Type HPO terms directly into the Known HPO Terms field. The autocomplete searches by term name and synonym. Add terms one at a time; they appear as removable chips below the input.
Use Extract HPO terms with AI below each pipeline's clinical description. The model suggests present and explicitly absent terms, including strongly supported ontology parents, while deterministic validation requires valid identifiers and evidence from the submitted text. Suggestions appear in removable chips and remain editable before the analysis starts.
Terms entered in the Absent HPO Terms field are recorded as excluded findings. They help refine the phenotype profile but do not contribute positively to the ranking score.
The deterministic matcher normalizes text, handles common plurals (seizures matches Seizure), and searches HPO term names and synonyms. It prioritizes more specific (longer) terms when the 25-term cap is reached. Gene-scope matching additionally recognizes a patient term's HPO ancestors, but it never treats unrelated associated symptoms as observed findings.
VarSage loads a gene-phenotype mapping file configured through GENE_PHENOTYPES_FILE. HGNC aliases and HPOA disease annotations can enrich this layer by connecting previous gene symbols, OMIM/ORPHA disease identifiers, disease-level phenotypes, and known inheritance terms. Phenotype overlap is therefore a structured plausibility measure, not a proof of causality.
For selected rare-disease candidates, VarSage can show predicted 3D protein structure context beside the structured variant evidence. This view helps reviewers localize an annotated amino-acid position on the AlphaFold reference model and consider whether the residue sits in a visible domain, interface, helix, sheet, or exposed region.
The structure panel is shown when the candidate has enough gene, protein, and residue information to connect the variant to a UniProt/AlphaFold reference model. If the mapping is incomplete, the report simply omits the panel.
The panel shows the reference model, a variant-context residue projection, confidence information when available, and links to external protein resources for deeper manual review.
Reviewers can use the image to discuss structural plausibility, domain location, and residue context alongside phenotype fit, inheritance, population rarity, clinical assertions, and functional evidence.
It does not prove pathogenicity, simulate a validated altered fold, replace molecular modeling, or change the variant rank without supporting evidence from the rest of the analysis.
Standard Variant Call Format. VarSage reads genotype, depth, genotype quality, filter status, INFO annotations, FORMAT fields, and transcript annotations. Multi-sample VCFs are supported when the analyzed sample ID is selected.
Tabular files with separate coordinate columns or a single parseable variant column. Common aliases for chromosome, position, reference allele, alternate allele, gene, consequence, and frequency are recognized.
Plain text files are treated as delimited tables when a consistent separator is present. They are most appropriate for small, already-curated variant lists.
VarSage can infer the parser from filename and content. For non-standard exports, optional AI schema detection can identify columns, after which coordinate parsing remains deterministic.
BED files define genomic intervals of interest, usually assay targets or capture regions. When a BED target set is selected, variants outside those intervals are removed before Jannovar annotation and external database merging. This improves runtime and keeps the report aligned with the technical scope of the assay.
BED coordinates use the standard 0-based half-open convention, while VCF and most variant tables use 1-based variant positions. VarSage handles the comparison internally, but the BED file must correspond to the same genome build as the uploaded variants.
The VarSage SuperDB layer is a local, build-specific coordinate annotation resource. It is intended to provide high-value fields needed for rapid interpretation: population rarity, clinical assertion context, prediction scores, splice scores, and stable variant identifiers. Large SuperDB files can be BGZIP-compressed and Tabix-indexed so only positions present in the current analysis are queried.
maxFreq summarizes population rarity by retaining the maximum observed allele frequency across configured population resources for the same allele. It provides a conservative rarity screen, but disease prevalence, penetrance, ancestry, and inheritance mode still determine the appropriate clinical threshold.
REVEL, CADD, SpliceAI, InterVar-like fields, ClinVar-like assertions, and custom laboratory columns can be attached when present in local annotation tables. These fields support review but do not replace primary evidence assessment.
VarSage offers optional AI-assisted features, but the platform remains functional without any model call. AI can help translate clinical text into reviewable HPO candidates, identify unusual table schemas, summarize evidence for top-ranked variants, and draft cautious report narrative from structured data. It is not the authority for coordinate extraction, allele identity, ACMG evidence, or clinical classification.
AI reads the free-text clinical description and suggests present or explicitly absent HPO terms. Suggestions must resolve to valid ontology identifiers and should be interpreted as phenotype curation assistance.
When enabled for a case, AI reviews a bounded top-ranked candidate set plus selected rescue candidates and provides a narrative summary of phenotype fit, inheritance support, evidence quality, and missing confirmation steps.
For AI-reviewed candidates, the model may suggest ACMG/AMP criteria that appear compatible with the supplied structured evidence. Deterministic validation checks the code, strength, and prerequisites before accepting any provisional refinement. Expert verification remains mandatory.
For unusual TSV/CSV headers, AI can identify which columns appear to encode chromosome, position, reference, alternate, gene, consequence, quality, and annotation fields. VarSage then parses coordinates deterministically.
When enabled, AI drafts a case-level synthesis from structured candidate data. The synthesis separates review priority, provisional classification, case-level causality, reportability, and missing evidence.
VarSage provides conservative ACMG/AMP-oriented evidence summaries for germline Mendelian sequence variants. These summaries are provisional decision-support aids, not clinical laboratory classifications. Evidence is deliberately labeled by state so reviewers can distinguish explicit auditable data from hypotheses and contextual information.
Derived from HPO overlap between the patient profile and gene-disease associations. High overlap supports phenotype-gene plausibility, but it is not sufficient by itself for variant pathogenicity.
Based on zygosity, sex, expected inheritance mode, known gene inheritance, and family clues. Trio and phase evidence remain provisional until sample identity, parentage, callability, and segregation are reviewed.
Derived from maxFreq and other frequency fields. Rarity supports review, but maximum credible allele frequency depends on prevalence, penetrance, inheritance mode, and disease-specific guidance.
REVEL, CADD, SpliceAI, and similar predictors support computational-impact review only when their thresholds and applicability are appropriate. Correlated predictors should not be counted independently.
Depth, genotype quality, and filter status are used to reduce technical noise. Quality metrics support confidence in a call, not pathogenicity.
| State | Meaning | Example |
|---|---|---|
| Applied | The prerequisite fact is explicitly present in the source data and can be audited. | A supplied validated functional assay strength, or de novo evidence with parentage confirmation. |
| Candidate | The data suggest a criterion, but disease-specific or laboratory prerequisites are incomplete. | Rarity compatible with PM2_Supporting, or possible compound heterozygosity without confirmed trans phase. |
| Informational | The field is useful context but is not counted as ACMG/AMP evidence. | ClinVar label, GTEx expression, LitVar context, or AI narrative without primary supporting evidence. |
PVS1 is not assigned from consequence alone; loss-of-function disease mechanism, transcript relevance, NMD expectation, exon rescue, and the ClinGen PVS1 decision tree require review. PM2 is treated conservatively at Supporting strength. PP3 and BP4 require calibrated predictors. PP5 and BP6 are not generated from ClinVar labels. De novo, trans-phase, co-segregation, and functional evidence require the corresponding laboratory and disease-specific prerequisites.
Back to topThe pharmacogenomics section annotates patient variants against build-specific PGx databases. When a variant matches a PGx entry, its drug, gene, phenotype category, predicted effect, evidence level, clinical score, and relevant genotype-response text are attached to the result. PGx findings are kept strictly separate from Rare Disease - Singleton ranking and ACMG/AMP classification because genotype-based drug-response guidance describes pharmacological context, not Mendelian disease causality.
Results that include PGx annotations display a hub-and-spoke network diagram that groups findings by pharmacogenomic category. Each hub node (large sphere) represents one category such as Toxicity, Efficacy, or Metabolism/PK. Satellite nodes around each hub show the genes, drugs or regimens, and variant identifiers belonging to that category. Hub size reflects the number of matched variants; color encodes category type. The diagram appears in the live results view, in saved HTML reports, and - as a compact pivot table - in Word report exports.
A PGx guideline recommends a prescribing action based on the detected genotype. Confirmatory testing and prescriber review are required before acting on any automated finding.
The genotype is relevant to medications the patient may receive in the future. No current prescribing change is implied, but the finding should be available in the patient record.
Evidence supports a possible prescribing consideration but does not yet reach guideline-level recommendation. Expert clinical pharmacology review is indicated.
The variant is associated with drug response in the literature but lacks prescribing-level guidance at this time. Treat as background context, not a clinical trigger.
Each PGx finding carries an evidence level, a clinical score derived from guideline strength and clinical relevance, a medication relevance indicator based on patient history where available, and source annotation links. The PGx priority score combines guideline actionability strength, medication relevance, genotype-phenotype match quality, evidence strength, clinical consequence relevance, and evidence consistency. This score orders the PGx summary table but does not represent a standard pharmacogenomic classification.
Next: GWAS Association ContextThe GWAS section attaches population-level trait associations from the GWAS Catalog to patient variants. GWAS context highlights the strongest phenotype signals overlapping the patient variant set and is placed after all diagnostic review sections because GWAS associations describe susceptibility at the population level, not Mendelian disease causality for this individual patient.
When GWAS context is available, a hub-and-spoke network diagram groups loci by confidence tier. Each hub represents one tier (Tier 1 - High confidence through Tier 5 - Low confidence, plus Excluded). Satellite nodes show the prioritized effector genes, matched patient phenotypes or GWAS traits, and genomic loci belonging to that tier. Hub size reflects the number of loci in the group; color encodes tier level from high-confidence red through low-confidence slate. The diagram appears in the live results view, saved HTML reports, and as a pivot table in Word exports.
Patient-relevant trait, genome-wide significant p-value (≤ 5×10⁻⁸), and strong variant-level evidence including fine-mapping or functional support. Highest patient-locus score.
Patient-relevant trait with suggestive significance (p ≤ 1×10⁻⁵) or moderate variant evidence. Warrants review in the context of the diagnostic question.
Association with a phenotype related to but not identical to the entered patient features. Provides secondary plausibility context.
Broad trait association with limited phenotype overlap. May indicate shared genetic architecture with the patient's condition but should be interpreted cautiously.
Each GWAS locus receives two scores. The patient-locus relevance score (0-100) combines phenotype relevance (40%), matched-trait GWAS evidence strength (25%), replication across studies (20%), and ancestry-study compatibility (15%). The variant evidence score (0-100) captures fine-mapping causal probability, functional annotation, QTL shared-variant support, patient effect-allele dosage, and gene-tissue plausibility. Both scores are shown in the GWAS summary table and referenced in the network hub size.
Where available, eQTL and sQTL records are attached to GWAS loci. These indicate that the variant or a correlated variant modulates expression or splicing of a nearby gene in a relevant tissue. QTL support is biological plausibility context; it does not substitute for direct functional evidence of pathogenicity.
Next: Genetic Ancestry EstimationVarSage can estimate a patient's genetic ancestry from a VCF file by comparing observed genotypes against a compact reference panel of ancestry-informative SNPs (AIMs). The output is a set of population-affinity proportions: each value reflects how well a reference population's allele-frequency pattern statistically explains the patient's observed genotype set. This approach follows the constrained genotype-likelihood mixture framework established in the population-genetics literature[1,2,11] and used by tools such as STRUCTURE[1] and ADMIXTURE,[2] adapted here for individual-level inference against a pre-computed reference panel.
Ancestry estimation has become a standard step in clinical and research genomics pipelines.[3,16] It supports population-stratified variant filtering (allele frequencies are population-specific), flags potential reference-panel mismatches, and provides provenance context for pharmacogenomic and disease-risk interpretation. Understanding the statistical method behind the result is essential for interpreting confidence levels and for recognizing when the estimate is reliable versus when it should be treated with caution.
A confidence label is assigned to every result based on the number of usable SNPs,[7] whether the optimizer converged, and the PCA distance to the nearest reference centroid.[9,10]
| Usable SNPs | Confidence tier | Interpretation guidance |
|---|---|---|
| < 1,000 | Estimation aborted - INSUFFICIENT_MARKERS | Too few markers to produce a reliable estimate. Common with WES or gene panels.[7] |
| 1,000 – 4,999 | LOW | Broad population assignment may be possible; fine-grained estimates are unreliable.[8,9] |
| 5,000 – 9,999 | MEDIUM | Suitable for superpopulation-level components.[5,8] Interpret with caution for closely related groups. |
| ≥ 10,000 | HIGH | Robust estimate for the populations represented in the installed reference database.[2,3] |
Total SNPs in the installed reference panel for this genome build. A larger panel improves resolution, especially for closely related populations.[8,9]
Panel positions found in the patient VCF with a genotype call at or above quality thresholds. Low values suggest sparse coverage, low-depth sequencing, or a gene-panel assay.[7]
Observed SNPs that were successfully reconciled between the patient VCF and the reference panel - including strand-flipped sites.[12] This is the count used for inference and confidence assignment.
Direct matches required no allele reorientation; reversed matches had their dosage flipped to correct strand orientation.[12] A high fraction of reversed matches is not a problem - it simply reflects opposite-strand representation in the two datasets.
A REFERENCE_MISMATCH flag appears when the patient's PCA projection[3,4] lies beyond a configured distance threshold from all reference population centroids.[9,10] This most commonly occurs when the patient has ancestry from a region underrepresented in the installed database - for example, a patient with Middle Eastern ancestry analyzed against a database built solely from the five 1000 Genomes[5] superpopulations. In this case, the mixture proportions will be spread across the nearest available components and may not accurately reflect the patient's true ancestral background.[9] The appropriate resolution is to install a database that includes the relevant reference population, not to treat the result as a sequencing quality indicator.
A result such as EUR 68% · SAS 24% · Other 8% means that the combination of 68% European-like and 24% South Asian-like reference allele frequencies provides the statistically best explanation for the genotypes in this VCF, given the installed reference panel.[1,2] It does not mean the patient identifies with these groups, was born in these regions, or has a specific genealogical history.[9,16] The percentages are population-affinity measures derived from allele-frequency patterns at thousands of SNPs simultaneously.
The ancestry feature requires a locally installed reference database. VarSage does not bundle reference data or connect to external ancestry services - the database is built by a maintainer from a legally obtained reference cohort (such as the 1000 Genomes Project[5] or the NHLBI TOPMed Program[17]) using the supplied build_ancestry_database.py script. After building, the compact panel is installed at the path configured in ANCESTRY_DATABASE_HG38 or ANCESTRY_DATABASE_HG19 and enabled with ANCESTRY_ENABLED=true. WES or gene-panel data typically yields fewer than 5,000 usable markers;[7] for reliable superpopulation-level estimates, a WGS VCF against a panel of 20,000 or more AIMs is strongly recommended.
| # | Citation (Vancouver style) |
|---|---|
| 1 | Pritchard JK, Stephens M, Donnelly P. Inference of population structure using multilocus genotype data. Genetics. 2000;155(2):945-59. |
| 2 | Alexander DH, Novembre J, Lange K. Fast model-based estimation of ancestry in unrelated individuals. Genome Res. 2009;19(9):1655-64. |
| 3 | Price AL, Patterson NJ, Plenge RM, Weinblatt ME, Shadick NA, Reich D. Principal components analysis corrects for stratification in genome-wide association studies. Nat Genet. 2006;38(8):904-9. |
| 4 | Patterson N, Price AL, Reich D. Population structure and eigenanalysis. PLoS Genet. 2006;2(12):e190. |
| 5 | 1000 Genomes Project Consortium, Auton A, Brooks LD, Durbin RM, Gargiulo EP, Kang HM, et al. A global reference for human genetic variation. Nature. 2015;526(7571):68-74. |
| 6 | Hardy GH. Mendelian proportions in a mixed population. Science. 1908;28(706):49-50. |
| 7 | Anderson CA, Pettersson FH, Clarke GM, Cardon LR, Morris AP, Zondervan KT. Data quality control in genetic case-control association studies. Nat Protoc. 2010;5(9):1564-73. |
| 8 | Rosenberg NA, Pritchard JK, Weber JL, Cann HM, Kidd KK, Zhivotovsky LA, et al. Genetic structure of human populations. Science. 2002;298(5602):2381-5. |
| 9 | Lawson DJ, van Dorp L, Falush D. A tutorial on how not to over-interpret STRUCTURE and ADMIXTURE bar charts. Nat Commun. 2018;9(1):3258. |
| 10 | McVean G. A genealogical interpretation of principal components analysis. PLoS Genet. 2009;5(10):e1000686. |
| 11 | Tang H, Peng J, Wang P, Risch NJ. Estimation of individual admixture: analytical and study design considerations. Genet Epidemiol. 2005;28(4):289-301. |
| 12 | Marchini J, Howie B. Genotype imputation for genome-wide association studies. Nat Rev Genet. 2010;11(7):499-511. |
| 13 | Purcell S, Neale B, Todd-Brown K, Thomas L, Ferreira MA, Bender D, et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. Am J Hum Genet. 2007;81(3):559-75. |
| 14 | Hellenthal G, Busby GB, Band G, Wilson JF, Capelli C, Falush D, et al. A genetic atlas of human admixture history. Science. 2014;343(6172):747-51. |
| 15 | Kivinen J, Warmuth MK. Exponentiated gradient versus gradient descent for linear predictors. Inf Comput. 1997;132(1):1-63. |
| 16 | Martin AR, Kanai M, Kamatani Y, Okada Y, Neale BM, Daly MJ. Clinical use of current polygenic risk scores may exacerbate health disparities. Nat Genet. 2019;51(4):584-91. |
| 17 | Taliun D, Harris DN, Kessler MD, Carlson J, Szpiech ZA, Torres R, et al. Sequencing of 53,831 diverse genomes from the NHLBI TOPMed Program. Nature. 2021;590(7845):290-9. |
| 18 | Privé F, Luu K, Blum MGB, McGrath JJ, Vilhjálmsson BJ. Efficient toolkit implementing best practices for principal component analysis of population genetic data. Bioinformatics. 2020;36(16):4449-57. |
VarSage is an independent project. With thanks to adbioinformatics.com for supporting VarSage and for writing about the platform in their blog article on AI exome variant prioritization.
We also thank Ahmad Golestani Far and Amir Shahbazi for the time they devoted to testing the tool.
Back to top