AVADA: toward automated pathogenic variant evidence retrieval directly from the full-text literature

Abstract

The primary literature on human genetic diseases with high penetrance includes descriptions of large numbers of pathogenic variants that can be essential for clinical diagnosis. Variant databases such as ClinVar and HGMD collect pathogenic variants by manual curation of either voluntary submissions or the published literature. AVADA (Automatically curated VAriant DAtabase) represents the first automated tool designed to construct a comprehensive database of highly penetrant genetic variants directly from full-text articles about human genetic disease. 2 AVADA was able to automatically curate almost 60% of the pathogenic variants deposited in 24 HGMD, over 4 times more than approaches parsing only PubMed abstracts. AVADA also 25 contains more than 60,000 pathogenic variants that are in HGMD, but not in ClinVar. Despite 26 being fully automated, 9 of AVADA's top 10 yielding journals are shared with HGMD's top 10, 27 and its mutation type distribution strongly resembles that of both HGMD and ClinVar. We demonstrate the utility of AVADA in clinical practice on a cohort of 245 patients with already diagnosed genetic diseases. Out of 260 causative variants originally reported for these patients, AVADA contained 38 variants described in the literature prior to publication of the patient cohort, compared to 43 using HGMD, 20 using ClinVar and only 13 (wholly subsumed by AVADA's) using an automated abstracts-only based approach. The database of automatically curated variants will be made available upon publication at http://bejerano.stanford.edu/AVADA. 35

The Human Gene Mutation Database (HGMD®): optimizing its use in a clinical diagnostic or research setting

“…Researchers at HGMD have been involved for several years in attempts to automatically extract mutation data from the literature. We recently contributed towards the Automatic Variant Evidence Database (AVADA), a novel machine learning tool that uses natural language processing to automatically identify pathogenic genetic variant evidence in full-text primary literature (Birgmeier et al 2020 ). AVADA automatically retrieved 58% of the likely disease-causing variants deposited in HGMD.…”

The UCSC Genome Browser database: 2021 update

“…Five new tracks were created to support the assessment of sequence variants in a clinical context: gnomAD Constraint Metrics (metrics of pathogenicity per-gene and transcript regions) (Variation Group) ( 8 ); gnomAD Structural Variants (allele frequencies of SVs in the common population) (Variation Group) ( 8 ); dbVar Curated Common Structural Variants (Variation Group) ( 9 ); Automatic Variant evidence Database (AVADA) variants extracted from full-text publications (Phenotype and Literature Group) ( 10 ); the ClinGen track collection, including Gene Dosage Sensitivity (haploinsufficiency and triplosensitivity) (Phenotype and Literature Group) ( 11 ), and Problematic Regions (regions known to cause short-read sequencing analysis artifacts) (Mapping and Sequencing Group).…”

VIP-HL: Semi-automated ACMG/AMP variant interpretation platform for genetic hearing loss

“…For example, Birgmeier and co-authors developed an end-to-end machine learning tool, named AVADA, for the automatic retrieval of variant evidence directly from full-text literature. 39 Suppose we can accumulate enormous datasets of evidence-related sentences or figures, in that case, it is possible to apply machine-learning approaches in the future for evidence retrieval and to automate the remaining ACMG/AMP rules in the next version of VIP-HL. In the meantime, our interface enables curators to manually activate the relevant codes after manual literature curation.…”