Fig. 3.

Overview of PASTRI algorithm. (a) We observe variant-allele read counts A and the total number of reads D that align to the locus for n mutations across m samples of the tumor. (b) A clustering algorithm that does not require that the data is generated by a phylogenetic tree gives an estimate $Q(\mathbf{F})$ of the posterior distribution over cluster cell fractions F. (c) PASTRI draws samples $\overline{F}$ from $Q(\mathbf{F})$. For each sample $\overline{F}$, PASTRI enumerates the set ${\mathcal{T}}_{\overline{F}}$ of trees T and assignments π of cell fractions to vertices of of T that satisfy the Sum Condition. All trees/vertex-assignment pairs not in ${\mathcal{T}}_{\overline{F}}$ have a probability of 0. Algorithm estimates the posterior probability of each tree using importance sampling