BeviMed has a function bevimed for performing Bayesian Evaluation of Variant Involvement in Mendelian Disease. The aim is to estimate the probability of an association between individual configurations of a restricted set of variants and a case-control label, and conditional on the association, the probabilities of pathogenicity for individual variants.
bevimed is an MCMC sampling procedure with many parameters, including the data y (a logical vector), a k by N matrix of variant-level genotypes G (where length(y) == nrow(G)), a mode of inheritance hypothesis encoded as an integer (minimum number of alleles required for affection) min_ac, and others determining the sampling management and prior distributions of the model parameters.
It returns a list of traces for the sampled parameters in an object of class BeviMed. This object can take up a lot of, and hence it may be preferable to store a summarised version by passing it to summary.
Quantities of interest can also be computed directly from the parameters - for example via the function log BF, returning a log Bayes factor between the variant-level model for case-control status and a baseline model whereby there is no association between variants in the region and case-control status - or the function probability_pathogenic a probability of association given some prior probability.
library(BeviMed)
set.seed(0)Firstly, we’ll generate some random data consisting of a disease affection status, y and a variant-wise indicator of pathogenicity, Z, where Z[i] == TRUE if variant i is pathogenic and FALSE otherwise.
y <- c(rep(TRUE, 20), rep(FALSE, 80))
Z <- c(rep(TRUE, 3), rep(FALSE, 20))In the first application, we’ll use variant matrix G1 where there is no association between y and the variants.
G1 <- sapply(y, function(y_i) as.integer(runif(n=length(Z)) < 0.15))
probability_pathogenic(G=G1, y=y)## [1] 0.06988054The results indicate that there is little preference for model \(v\) over model \(n\). In the example with association, we’ll sample the allele counts based on y.
G2 <- sapply(y, function(y_i) as.integer(runif(n=length(Z)) < 0.15 |
(if (y_i) 1:length(Z) == sample(which(Z), size=1) else rep(FALSE, length(Z)))))
probability_pathogenic(G=G2, y=y)## [1] 0.9999853For a more detailed output (containing the variant-level probabilities of pathogenicity), the bevimed function can be called, summarised and stored.
output <- summary(bevimed(G=G2, y=y))
output## ---------------------------------------------------------------------------
## Log Bayes factor between variant level model and null model is 13.94
##
## A confidence interval for the estimate is:
## 2.5% 97.5%
## 13.26 14.49
## ---------------------------------------------------------------------------Note that now the estimated log Bayes factor in favour of the variant-level model is high. Furthermore, we can use print passing the option print_Z=TRUE to print a summary of the results containing the variant-level probabilities of pathogenicity.
print(output, print_Z=TRUE)## ---------------------------------------------------------------------------
## Log Bayes factor between variant level model and null model is 13.94
##
## A confidence interval for the estimate is:
## 2.5% 97.5%
## 13.26 14.49
## ---------------------------------------------------------------------------
## Estimated probabilities of pathogenicity of individual variants
## (conditional on model v)
##
## Variant Controls Cases P(Z_j=1|y,V) Bar Chart
## 1 15 8 0.99 [=================== ]
## 2 10 11 1.00 [====================]
## 3 6 5 0.92 [================== ]
## 4 12 8 0.00 [ ]
## 5 15 3 0.00 [ ]
## 6 11 1 0.00 [ ]
## 7 14 5 0.00 [ ]
## 8 15 3 0.00 [ ]
## 9 15 2 0.00 [ ]
## 10 8 2 0.01 [ ]
## 11 15 2 0.00 [ ]
## 12 8 4 0.01 [ ]
## 13 13 1 0.00 [ ]
## 14 9 3 0.04 [ ]
## 15 10 3 0.00 [ ]
## 16 16 1 0.00 [ ]
## 17 14 2 0.00 [ ]
## 18 13 3 0.00 [ ]
## 19 12 1 0.00 [ ]
## 20 17 2 0.00 [ ]
## 21 13 7 0.00 [ ]
## 22 10 3 0.01 [ ]
## 23 15 6 0.00 [ ]