Fork me on GitHub!


Build Status Coverage Status

A Julia package implementing several topic models used for mutation signature estimation.

The Multi-modal correlated topic model (MMCTM) takes an array of arrays of matrices, where the first column of each matrix is a mutation type index, and the second column is the mutation count for a particular sample.

The following example shows how to perform inference using the MMCTM on SNV and SV counts:

using MultiModalMuSig
using CSV
using DataFrames
using VegaLite
using Random


snv_counts ="data/brca-eu_snv_counts.tsv", delim='\t')
sv_counts ="data/brca-eu_sv_counts.tsv", delim='\t')

X = format_counts_mmctm(snv_counts, sv_counts)
model = MMCTM([7, 7], [0.1, 0.1], X)
fit!(model, tol=1e-5)

snv_signatures = DataFrame(hcat(model.ϕ[1]...))
snv_signatures[:term] = snv_counts[:term]
snv_signatures = melt(
    snv_signatures, :term, variable_name=:signature, value_name=:probability
snv_signatures |> @vlplot(
    :bar, x={:term, sort=:null}, y=:probability, row=:signature,


Since these types of models can get stuck in poor local optima, it’s a good idea to fit many models and pick the best one.