A probabilistic disease-gene finder for personal genomes

Abstract

VAAST (the Variant Annotation, Analysis & Search Tool) is a probabilistic search tool for identifying damaged genes and their disease-causing variants in personal genome sequences. VAAST builds on existing amino acid substitution (AAS) and aggregative approaches to variant prioritization, combining elements of both into a single unified likelihood framework that allows users to identify damaged genes and deleterious variants with greater accuracy, and in an easy-to-use fashion. VAAST can score both coding and noncoding variants, evaluating the cumulative impact of both types of variants simultaneously. VAAST can identify rare variants causing rare genetic diseases, and it can also use both rare and common variants to identify genes responsible for common diseases. VAAST thus has a much greater scope of use than any existing methodology. Here we demonstrate its ability to identify damaged genes using small cohorts (n = 3) of unrelated individuals, wherein no two share the same deleterious variants, and for common, multigenic diseases using as few as 150 cases.

Figure 1.

VAAST uses a feature-based approach to prioritization. Variants along with frequency information, e.g., 0.5:A 0.5:T, are grouped into user-defined features (red boxes). These features can be genes, sliding windows, conserved sequence regions, etc. Variants within the bounds of a given feature (shown in red) are then scored to give a composite likelihood for the observed genotypes at that feature under a healthy and disease model by comparing variant frequencies in the cases (target) compared to control (background) genomes. Variants producing nonsynonymous amino acid changes are simultaneously scored under a healthy and disease model.

Figure 2.

Observed amino acid substitution frequencies compared to BLOSUM62. Amino acid substitution frequencies observed in healthy and reported for OMIM disease alleles were converted to LOD-based scores for purposes of comparison to BLOSUM62. The BLOSUM62 scores are plotted on the y-axis throughout. (Red circles) stops; (blue circles) all other amino acid changes. The diameter of the circles is proportional to the number of changes with that score in BLOSUM62. (A) BLOSUM62 scoring compared to itself. Perfect correspondence would produce the diagonally arranged circles shown. (B) Frequencies of amino acid substitutions in 10 healthy genomes compared to BLOSUM62. (C) OMIM nonsynonymous variant frequencies compared to BLOSUM62.

Figure 3.

Impact of population stratification and platform bias. Numbers of false positives with and without masking. (A) Effect of population stratification. (B) Effect of heterogeneous platform and variant calling procedures. (Red line) Number of false positives without masking; (blue line) after masking. Note that although masking has little effect on population stratification, it has a much larger impact on platform bias. This is an important behavior: Population stratification introduces real, but confounding signals into disease gene searches; these signals are unaffected by masking (A); in contrast, VAAST's masking option removes false positives due to noise introduced by systematic errors in platform and variant calling procedures (B).

Figure 4.

Genome-wide VAAST analysis of Utah Miller Syndrome Quartet. VAAST was run in its quartet mode, using the genomes of the two parents to improve specificity when scoring the two affected siblings. Gray bars along the center of each chromosome show the proportion of unique sequence along the chromosome arms, with white denoting completely unique sequence; black regions thus outline centromeric regions. Colored bars above and below the chromosomes (mostly green) represent each annotated gene; plus strand genes are shown above and minus strand genes below; their width is proportional to their length; height of bar is proportional to their VAAST score. Genes colored red are candidates identified by VAAST. Only two genes are identified in this case: DNAH5 and DHODH. Causative allele incidence was set to 0.00035, and amino acid substitution frequency was used along with variant-masking. This view was generated using the VAAST report viewer. This software tool allows the visualization of a genome-wide search in easily interpretable form, graphically displaying chromosomes, genes, and their VAAST scores. For comparison, the corresponding figure, without pedigree information, is provided as Supplemental Figure 1.

Figure 5.

Benchmark analyses using 100 different known disease genes. In each panel the y-axis denotes the average rank of the disease gene among 100 searches for 100 different disease genes. Heights of boxes are proportional to the mean rank, with the number above each box denoting the mean rank of the disease gene among all RefSeq annotated human genes. Error bars encompass the maximum and minimum observed ranks for 95% of the trials. (A) Average ranks for 100 different VAAST searches. (Left half of panel) The results for genome-wide searches for 100 different disease genes assuming dominance using a case cohort of two (blue box), four (red box), and six (green box) unrelated individuals. (Right half of panel) The results for genome-wide searches for 100 different recessive disease genes using a case cohort of 1 (blue box), 2 (red box), and 3 (green box). (B) Impact of missing data on VAAST performance. (Left and right half of panel) Results for dominant and recessive gene searches as in panel A, except in this panel the case cohorts contain differing percentages of individuals with no disease-causing variants in the disease gene. (Blue box) Two-thirds of the individuals lack a disease-causing allele; (red box) one-third lack a disease-causing allele; (green box) all members of the case cohort contain disease-casing alleles. (C) Comparison of VAAST performance to that of ANNOVAR and SIFT. (Left half of panel) The results for genome-wide searches using VAAST, ANNOVAR, and SIFT to search for 100 different dominant disease genes using a case cohort of six unrelated individuals. (Right half of panel) The results for genome-wide searches using VAAST, ANNOVAR, and SIFT to search for 100 different recessive disease genes using a case cohort of three unrelated individuals.

Figure 6.

Statistical power as a function of number of target genomes for two common disease genes. (A) NOD2, using a data set containing rare and common nonsynonymous variants. (B) LPL, using a data set containing only rare nonsynonymous variants. For each data point, power is estimated from 500 bootstrapped resamples of the original data sets, with α = 2.4 × 10⁻⁶ except where specified. y-axis: probability of identifying gene as implicated in disease in a genome-wide search; x-axis: number of cases. The number of controls is equal to the number of cases up to a maximum of 327 for LPL (original data set) and 163 for NOD2 (original data set + 60 Europeans from 1000 Genomes). (VAAST + OMIM) VAAST using AAS data from OMIM as its disease model; (VAAST + BLOSUM) VAAST using BLOSUM62 as its disease model; (VAAST no AAS) VAAST running on allele frequencies alone; (WSS) weighted sum score of Madsen and Browning (2009); (GWAS) single variant GWAS analysis. NOD2 and LPL data sets were taken from Lesage et al. (2002) and Johansen et al. (2010), respectively.

Figure 7.

VAAST search procedure. One or more variant files (in VCF or GVF format) are first annotated using the VAAST annotation tool and a GFF3 file of genome annotations. Multiple target and background variant files are then combined by the VAAST annotation tool into a single condenser file; these two files, one for the background and one for the target genomes, together with a GFF3 file containing the genomic features to be searched are then passed to VAAST. VAAST outputs a simple text file, which can also be viewed in the VAAST viewer.