Modeling Intratumor Gene Copy Number Heterogeneity using Fluorescence in Situ Hybridization data

TL;DR

This paper develops a probabilistic model using hierarchical log-linear models to infer tumor progression pathways from FISH data, capturing dependencies among gene copy number changes and producing phylogenetic trees.

Contribution

It introduces a novel Markov chain model for FISH data that accounts for dependencies, along with a theorem for reconstructing oncogenetic trees from such data.

Findings

Successfully modeled tumor progression pathways.

Validated method on breast tumor dataset.

Captured dependencies in gene copy number changes.

Abstract

Tumorigenesis is an evolutionary process which involves a significant number of genomic rearrangements typically coupled with changes in the gene copy number profiles of numerous cells. Fluorescence in situ hybridization (FISH) is a cytogenetic technique which allows counting copy numbers of genes in single cells. The study of cancer progression using FISH data has received considerably less attention compared to other types of cancer datasets. In this work we focus on inferring likely tumor progression pathways using publicly available FISH data. We model the evolutionary process as a Markov chain in the positive integer cone Z_+^g where g is the number of genes examined with FISH. Compared to existing work which oversimplifies reality by assuming independence of copy number changes, our model is able to capture dependencies. We model the probability distribution of a dataset with hierarchical log-linear models, a popular probabilistic model of count data. Our choice provides an attractive trade-off between parsimony and good data fit. We prove a theorem of independent interest which provides necessary and sufficient conditions for reconstructing oncogenetic trees. Using this theorem we are able to capitalize on the wealth of inter-tumor phylogenetic methods. We show how to produce tumor phylogenetic trees which capture the dynamics of cancer progression. We validate our proposed method on a breast tumor dataset.