Limitations

Nuclear SNVs and small insertions/deletions are processed through the ACMG/AMP classifier.

Constitutional deletions and duplications are processed through the dedicated SV/CNV module under the Riggs 2020 scoring framework. The module does not replace upstream structural-variant calling.

Mitochondrial variants use a separate MMDWG 2020 classifier with mtDNA-specific frequency, haplogroup, heteroplasmy, predictor, and NUMT evidence.

Repeat expansions are not detected or interpreted. Disorders caused by tandem-repeat expansion require a dedicated upstream caller and interpretation workflow.

Somatic tumor variants are not supported. Tumor-normal analysis and AMP/ASCO/CAP somatic classification remain outside the product scope.

GRCh38 (hg38) is the primary supported genome build. All reference databases are indexed to GRCh38.

GRCh37 (hg19) VCF files are accepted and automatically lifted over to GRCh38 using CrossMap. Liftover may fail for a small number of variants in regions with structural differences between builds.

Earlier genome builds (hg18 and prior) are not supported.

The nuclear classifier computes evidence that can be derived from variant annotations, reference data, phenotype input, and supported family context. Functional assays, case-control evidence, and other expert-curated observations still require reviewer input.

Gene-specific VCEP specifications are applied where available. Genes without a supported VCEP overlay use the documented general thresholds and guards.

Every automated class remains subject to geneticist review. The software records the applied criteria and evidence path but does not replace clinical judgment.

A full trio can establish parental origin for candidate compound-heterozygous pairs. Without informative parental or read-based phasing, the pair remains a candidate rather than a confirmed in-trans result.

gnomAD v4.1 population frequencies have uneven coverage across ancestry groups. Non-Finnish European samples represent the majority of the cohort. Variants in underrepresented populations may have less precise frequency estimates.

ClinVar assertions vary in quality. Single-submitter assertions with zero review stars carry less weight than expert panel consensus. The platform applies review star filtering but cannot independently validate submitter quality.

dbNSFP functional predictions are limited to missense variants at coding positions. Non-coding variants, UTR variants, and deep intronic variants do not receive BayesDel, SIFT, or AlphaMissense scores.

HPO gene-phenotype associations depend on the completeness of disease gene curation. Recently discovered gene-disease associations may not yet be reflected in the HPO database.

The literature database is a filtered local PubMed snapshot, not a live mirror. Freshness is not guaranteed, so recent, corrected, retracted, or exhaustive literature must be verified against live PubMed and the journal record.

The AI assistant generates clinical interpretations that require independent validation by a qualified clinical geneticist. AI-generated content does not constitute a medical diagnosis.

The assistant only has access to data within the current analysis session. It cannot cross-reference findings across different patients or historical cases.

Conversation context is limited to the last 20 messages. Very long diagnostic discussions may lose early context.

SQL generation from natural language may occasionally produce incorrect queries for ambiguous or highly complex questions. The generated SQL is visible to the user for verification.

The literature search operates on the local PubMed mirror. It cannot access preprint servers, full-text content, or databases outside the indexed PubMed corpus.

Primary variant analysis accepts single-sample VCF files. Family analysis links separately processed proband, parent, and sibling samples into a trio, duo, or extended family configuration.

VCF files must conform to VCF 4.1 or 4.2 specification. Non-standard fields in the INFO or FORMAT columns are ignored.

The platform processes up to approximately 5 million variants per file. Whole genome sequencing files typically contain 4-5 million variants and are fully supported.

Phenotype matching requires HPO terms. Free-text clinical descriptions are supported through automatic HPO term extraction, but manually curated HPO terms produce better matching accuracy.