pyNBS: a Python implementation for network-based stratification of tumor mutations

Abstract

Summary

We present pyNBS: a modularized Python 2.7 implementation of the network-based stratification (NBS) algorithm for stratifying tumor somatic mutation profiles into molecularly and clinically relevant subtypes. In addition to release of the software, we benchmark its key parameters and provide a compact cancer reference network that increases the significance of tumor stratification using the NBS algorithm. The structure of the code exposes key steps of the algorithm to foster further collaborative development.

Availability and implementation

The package, along with examples and data, can be downloaded and installed from the URL https://github.com/idekerlab/pyNBS.

1 Introduction

The biomedical community increasingly relies on genomic information to diagnose and treat many different complex diseases, including cancer (Frampton, 2013; Johnson, 2014). In parallel, developments in molecular interaction mapping technologies and network analysis algorithms have enabled the systematic elucidation of pathways involved in cancer and other complex diseases (Schaefer et al., 2009). These two technologies—genomics and network analysis—have been recently combined to contextualize somatic mutations in tumors against the knowledge contained in molecular interaction networks and disease pathway maps. For example, numerous algorithms now use molecular network information to discover significantly mutated pathways in particular cohorts of patients (Ciriello, 2012; Drake, 2016; Leiserson, 2013, 2014; Paull, 2013; Vandin, 2011a,b; Vaske, 2010).

Recently, we introduced an algorithm that uses molecular network information to guide the stratification of tumor somatic mutation profiles into clinically relevant subtypes (Hofree, 2013). Such mutation profiles have been notoriously difficult to stratify (i.e. cluster) due to their extreme heterogeneity from patient to patient. Our algorithm, called Network-Based Stratification (NBS), relies upon aggregating these mutations in molecular network neighborhoods to gain power in separating patients. The underlying assumption is that cancer arises due to disruptions in specific molecular pathways, not only disruptions in isolated genes (Vanunu et al., 2010). It is commonly observed that similar cancer types arise from mutations that affect different genes that are participants in common pathways. However, traditional gene-wise clustering methods fail to capture similarities that are observed only on the pathway level, since mutations do not necessarily fall on the same genes and therefore do not contribute to any measure of similarity between patients despite affecting the same pathway. The information of each somatic mutation is smoothed across its network neighborhood, spreading the signal to other functionally related genes in network space. It is then possible to obtain robust clusters of patients based on the similarity of these network-smoothed mutation profiles.

In the original publication of NBS, the code used to develop the project was provided in MATLAB, a proprietary programming language, making open access to this software difficult. Additionally, the code lacked modularization, making individual steps of the algorithm difficult to control, analyze and test. In what follows, we implement and organize the NBS algorithm as an installable Python package, which we call pyNBS. This package modularizes and exposes the major steps in the algorithm to better control, analyze and improve the approach in future studies.

2 Materials and methods

The NBS algorithm requires two inputs: a matrix of binary values describing all somatic tumor mutations found within a cohort of cancer patients (patients × genes) and a second file describing the gene-gene interactions defining a reference molecular network. Given these inputs, the NBS algorithm clusters the tumor mutation profiles into molecular subtypes as seen in Figure 1. Additional details of the algorithm are described in the original NBS manuscript (Hofree, 2013).

Fig. 1.

Overview and stepwise factorization of the NBS algorithm

3 Results

3.1 pyNBS usage and validation

The NBS algorithm can be executed using the pyNBS package in two modes: using a wrapper script via the command line, or by running the provided Jupyter Notebooks. Documentation for both code execution modes are provided within a GitHub repository, which can be found at: https://github.com/idekerlab/pyNBS.

It should be noted that each full run of pyNBS does not necessarily produce the exact same cluster assignments on the same cohort. This variation is due to the stochastic nature of the sub-sampling step as well as the non-unique nature of matrix factorization (Cai et al., 2011). However, this variance is largely controlled by the final consensus clustering step.

We tested the pyNBS package by generating patient subtypes in ovarian and uterine cancer using the data and corresponding networks released with the original Hofree et al. manuscript. PyNBS nearly perfectly recovered the original Hofree patient cluster assignments for ovarian and uterine cancer (χ²P-value: 2.3 × 10⁻¹⁰⁷ and 5.3 × 10⁻⁸⁸, respectively). These two test examples are provided, along with the required datasets (re-formatted for usage with pyNBS), as Jupyter Notebooks in the GitHub repository.

3.2 A cancer-specific network for pyNBS

In addition to reconstructing the original NBS algorithm, we also explored alternative reference networks for their ability to separate tumor cohorts into clinically relevant subtypes. The outcome of this exploratory research was a compact cancer reference network that contained only high-confidence interactions specific to cancer. To construct this network, we began with a high-quality network assembled in a previous study, containing 19 781 genes and 2 724 724 interactions supported by multiple lines of evidence (PCNet, Huang et al., 2018). We filtered this network to retain only cancer genes as documented in at least one of four collections (Forbes, 2017; Hanahan and Weinberg, 2011; Iorio, 2016; Vogelstein, 2013). We found that this cancer reference network (CRN) more effectively clusters tumor samples from several different cancer types, as measured by the clusters’ ability to predict patient survival, in comparison to one of the networks used in the original NBS study (Fig. 2A). This cancer reference network, as well as directions on constructing this network and analysis of the effect of different network models on pyNBS, are presented as Jupyter Notebooks located in the GitHub repository.

Fig. 2.

Benchmarking and pyNBS stratification performance. (A) Significance of survival separation between subtypes in bladder (BLCA), colon (COAD), head and neck (HNSC) and uterine (UCEC) cancer as discovered by pyNBS. Cohorts were stratified using the top 10% of edges in HumanNet (HN90, blue), our cancer reference subnetwork (CRN) from PCNet (gold, see text), without network propagation using CRN genes (green), and with propagation over 10 degree-preserving shuffles of CRN (red). Note that the green and red bars provide controls on CRN (gold) and should not be compared to HN90 (blue). HN90 outperforms its analogous controls (Hofree et al., 2013). (B) Consensus clustering convergence rate and runtime performance of pyNBS on TCGA head and neck cancer data with HN90 (blue) and the Cancer Subnetwork (gold). By measuring the agreement of consensus clustering results at each step and the consensus clustering result using 10 fewer sub-sampling iterations, it is clear that the consensus clustering is fairly stable at just 100 sub-sampling iterations

3.3 Practical benchmarking and parameter tuning

The pyNBS algorithm can be expensive in both memory and in run time for large networks, or if many iterations of the sub-sampling and matrix factorization are required. However, we found that 1000 iterations of sub-sampling and consensus clustering, as originally performed by Hofree et al., could be markedly decreased with little reduction in performance, with only 100 iterations being sufficient for the consensus clustering to converge. This reduction can offer 90% run time savings with no appreciable deviation in the results (Fig. 2B). For example, to stratify the TCGA head and neck cancer data using the filtered HumanNet (HN90, as described by Hofree et al.), we reduced the runtime of pyNBS from approximately 21.5 h to 2.2 h.

In addition, using the filtered Cancer Subnetwork (see above), which only has 2291 nodes compared to the 7939 nodes in HN90, we see that pyNBS not only runs much faster, but by reducing the consensus clustering iterations, this also reduces the overall runtime of pyNBS in this scenario from 6.5 h to approximately 40 min (Fig. 2B). Due to the NBS algorithm requiring many matrix multiplications, we recommend running pyNBS on a machine with at least four threads and 4GB of RAM per thread. Such operations also suggest that further optimization can be had by the utilization of GPUs.

While we mainly sought to recreate the original procedure and parameter space for running pyNBS here, we performed an additional exploration on the effect of varying several parameters and algorithmic decisions on the final consensus clustering results in pyNBS. We present some of these results as Jupyter Notebooks located in the GitHub repository.