find_weights: Produce annotation-based weightings

In andrewGhazi/bayesianMPRA: A Bayesian framework for modelling MPRA data with an informative, annotation based empirical prior

Description

For the ith variant, adaptively produce a vector of weightings such that at minimum min_num_contributing variants meaningfully contribute (i.e. cumsum(sortedWeights)[30] <= .99) arguments after the 2nd are heuristics that may need tuning

Usage

find_weights(i, dist_mat, min_dist_kernel, min_num_contributing = 30,
  increase_fold = 1.333)

Arguments

i	the index of the variant in the assay
dist_mat	annotation-based distance matrix
min_dist_kernel	the minimum kernel possible to use in the annotation space
min_num_contributing	the minimum number of variants that must contribute to the ith variant's prior
increase_fold	the multiplicative amount by which to increase the kernel in the case the current kernel doesn't allow at least min_num_contributing variants to contribute

Details

The kernel is initialized at some small, pre-computed value then iteratively increased until there are "enough" variants contributing. This keeps the prior for one variant from being too strongly influenced by extremely close neighbors in annotation space.

Value

a similarity-based vector of weights of the other n-1 variants

andrewGhazi/bayesianMPRA documentation built on May 28, 2019, 4:56 p.m.