Abstract

1 Introduction

Cancer develops through the successive expansions of clones, in which certain (epi)genomic alterations, called drivers, confer a fitness advantage and progressively accumulate, in a context of overall scarcity of resources (Nowell, 1976). Specifically, in Nowell’s seminal work, tumor evolution is described in terms of stepwise genetic variation such that growth advantage is the key for the survival and prorogation of the clones. Therefore, one can define cancer progression models, in terms of probabilistic causal graphical models, where the conditional dependencies and the temporal ordering among these alterations are described, revealing the evolutionary trajectories of cancer at the (epi)genome level.

We further distinguish (Caravagna et al., 2015) (i) ensemble-level progression models, describing the statistical trends of accumulation of genomic alterations in a cohort of distinct cancer patients. Such models describe the temporal partial orders of fixation and accumulation of such alterations and represent population-level trends; and (ii) individual-level models, thus accounting for the specific evolutionary history of cancer clones in individual tumors. Such models thus impute the ancestry relations of the observed clones.

Even if the inference of such models is further complicated by a series of theoretical and technical hurdles, such as, e.g. intra- and inter-tumor heterogeneity and the effective detection of drivers, it can benefit from the increasing amount of next-generation sequencing data, currently available through public projects such as The Cancer Genome Atlas (TCGA, https://tcga-data.nci.nih.gov). Usually, such databases provide cross-sectional (epi)genomic profiles retrieved from single biopsies of cancer patients, which can be used to extract ensemble-level models; but higher-resolution data such as multiple-biopsies, or even single-cell sequencing data, are becoming more accessible and reliable, which can be used to infer individual-level models.

Here, we introduce TRanslational ONCOlogy (TRONCO), an R package built to infer cancer progression models from heterogeneous genomic data (in the form of alterations persistently present along tumor evolution.) Currently, TRONCO provides the implementation of two algorithms: (i) CAncer PRogression Extraction with Single Edges (CAPRESE) (Olde Loohuis et al., 2014) and (ii) CAncer PRogression Inference (CAPRI) (Ramazzotti et al., 2015), both based on Suppes’ theory of probabilistic causation (Suppes 1970), but with distinct goals and properties (see Section 2).

TRONCO, in its current form and perspective, should be thought of as a tool that provides the implementation of up-to-date solutions to the progression inference problem. At the time of the writing it can be effectively used as the final stage of a modular pipeline for the extraction of ensemble-level cancer progression models from cross-sectional data (Caravagna et al., 2015). In such a pipeline input data are pre-processed to (i) stratify samples in tumor subtypes, (ii) select driver alterations and (iii) identify groups of fitness-equivalent (i.e. mutually exclusive) alterations, prior to the application of the CAPRI algorithm. The resulting ensemble-level progression models depict the evolutionary dynamics of cancer, with translational impacts on diagnostic and therapeutic processes, especially in regard to precision medicine and personalized drug development.

From the complementary perspective, TRONCO can also exploit the CAPRESE algorithm to infer the clonal evolutionary history in single patients when multiple samples are available, as in the case of multiple biopsies and/or single-cell sequencing data, as long as the set of driver events is selected; see Caravagna et al. (2015).

2 Software implementation

TRONCO implements a set of R functions to aid the user in extracting a cancer progression model from genomic data. At a high level, these functions shall help to import, visualize and manipulate genomic profiles—regardless of their source—eventually allowing the implemented algorithms to run and assess the confidence in a model.

The basics steps of TRONCO’s usage are shown in Figure 1. In Figure 1A, we show multiple input alterations (e.g. somatic mutations or copy number alterations) either from a cohort of patients, or a unique patient (e.g. multi-region or single-cell sequencing); in Figure 1B we show an oncoprint visualization from the tool, i.e. a matrix whose columns represent samples and rows the alterations and their presence per sample; Figure 1C shows an inferred graphical Bayesian progression model obtained with one of the available algorithms; finally, in Figure 1D we show data supported for processing in the tool. For a more detailed explanation of the implementation of the package see Antoniotti et al. (2015).

Open in a separate window

Fig. 1.

(A) TRONCO can process either alterations (e.g. somatic mutations or wider chromosomal lesions) in a cohort of independent samples (top lolliplot diagram), or a set of multiple snapshots from a unique patient (e.g. multi-region or single-cell, bottom panel). (B) Oncoprints allow the user to visualize the data that the tool is processing. Regardless of the source, each row represents a certain alteration—at a custom resolution depending on the cancer under study—and each column a sample. (C) A model inferred with the tool might outline cancer evolution occurring in a population ensemble or in an individual patient. Graphically, alterations are represented as nodes with different colors (e.g. green mutations and blue homozygous deletions). Algorithms such as CAPRI allow describing alterations with logical formulas, in an attempt to find their role as a ‘group’ (see Ramazzotti et al., 2015 for details); we picture such groups with dashed lines. In the panel, we show a hypothetical ensemble-level model predicting a selection pressure on two genes mapped to 17p13, tp53 and hic1, as it may be inferred by analyzing samples harboring either tp53/hic1 mutations or homozygous deletions in the cytoband where any of these two genes map, i.e. here for purely explanatory cases we suppose just tp53, which maps to 17p13.1. The model suggests a trend of selection toward mutations in gene y, which shall be interpreted as a set of preferential clonal expansions characteristic of the population of analyzed samples, involving alterations of the functions mapped to 17p13 and y. (D) TRONCO supports three data types. Custom data, which is supposed to be provided as a binary input matrix storing the presence (1) or absence (0) of a certain alteration in a sample. Or, standard data formats such as the MAF for somatic mutations, as well as the GISTIC format for focal Copy Number Variations. Data can be generated by custom experiments, or collected—along with other ‘omics’—from public databases such as TCGA and cBio portal. For the latter, cBio portal, TRONCO implements a query system to fetch data with minimal effort. The tool engine can then be used to manipulate genomic profiles—regardless of their source—and run progression inference algorithms

3 Discussion

TRONCO provides up-to-date, theoretically well-founded, statistical methods to understand the evolution of a cancer (ensamble-level) or a single tumor (individual-level). The implemented algorithms are demonstrably the state-of-the-art for the progression inference problem, in terms of computational cost, scalability with respect to sample size, accuracy and robustness against noise in the data. The implementation makes straightforward the interaction of TRONCO with other common bioinformatics tools, possibly allowing the creation of a common suite of tools for cancer progression inference.

Finally, we refer to either Caravagna et al. (2015) or the Supplementary materials for a demonstration of the usage of TRONCO on real genomics data both at the ensemble-level and individual-level progression models. In particular, this article outlines the capability of the methods to reproduce much of the current knowledge on the progression for a set of cancer types, as well as to suggest clinically relevant insights. Furthermore, we also provide users with detailed manuals, vignettes and source code to replicate all the analysis presented in the article plus others (case studies: colorectal cancer, clear cell renal cell carcinoma and acute chronic myeloid leukemia) in the Supplementary materials and at the TRONCO official webpage (software section at http://bimib.disco.unimib.it/).

Supplementary Material

Acknowledgement

References