Tree of Mutually Exclusive Oncogenes

Mohammadreza Mohaghegh Neyshabouri, Jens LagergrenbioRxiv 2022 · DOI

TL;DR

This paper introduces ToMExO, a novel probabilistic model that uses a tree structure to represent cancer progression, capturing both the temporal order of driver gene mutations and their mutual exclusivity. By combining efficient dynamic programming for likelihood calculations with a sophisticated MCMC inference algorithm, ToMExO can analyze large cancer genomic datasets to uncover complex driver gene interactions.

What Problem Does This Paper Address?

Understanding how cancer progresses is crucial for developing effective treatments. This progression is driven by an accumulation of mutations in specific "driver" genes, which confer selective advantages to cancer cells. Identifying these driver genes and their interrelationships is challenging due to the vast number of mutations in any given tumor, many of which are "passenger" mutations with no significant impact on evolution. Furthermore, driver mutations often exhibit mutual exclusivity, meaning that the occurrence of one mutation can prevent another from occurring, likely because they disrupt the same biological pathways or protein complexes. Existing models often struggle to simultaneously capture both the temporal order of mutations and their mutual exclusivity, especially in large datasets.

Methods

ToMExO tackles this problem by modeling cancer progression as a tree. Each node in the tree represents a set of mutually exclusive driver genes. The edges of the tree represent the temporal order in which these sets of genes are likely to be mutated.

Key methodological contributions include:

  • Probabilistic Progression Model: A tree-based model where nodes contain sets of mutually exclusive driver genes, and edges represent firing probabilities (the likelihood of a mutation occurring given the parent node is mutated).
  • Dynamic Programming for Likelihood Calculation: An efficient algorithm to compute the probability of observing a given set of mutations in a tumor, given a specific progression model. This is crucial for handling large datasets.
  • MCMC Inference Algorithm: A Markov Chain Monte Carlo approach is used to search for the best-fitting progression tree and associated error parameters (false positives and false negatives) from the observed mutation data.
  • Novel MCMC Moves: A suite of specialized "structural moves" are designed to efficiently explore the space of possible tree structures, including merging/splitting nodes, attaching/detaching genes, and swapping genes. This allows for flexible model exploration and avoids getting stuck in local optima.
  • Error Parameter Estimation: The method accounts for false positive and false negative mutation calls, which are common in genomic data. It uses a dynamic programming approach to estimate optimal error parameters for a fixed tree structure and then fine-tunes them with gradient descent.

Datasets

The paper uses two types of datasets:

  • Synthetic Datasets: Generated to systematically evaluate the performance of ToMExO under various controlled scenarios, including different error rates and tumor numbers, and different underlying progression model structures (linear, tree-like, and mixture of trees).
  • Biological Datasets: Real-world cancer mutation data from The Cancer Genome Atlas (TCGA) for three cancer types:
  • Uterine Corpus Endometrial Carcinoma (UCEC)
    • Cholangiocarcinoma (CHOL)
    • Glioblastoma Multiforme (GBM)

These biological datasets were preprocessed to filter silent mutations and select "potentially driver genes" based on inclusion in the IntOGen driver gene list and a minimum mutation frequency.

Key Insights and Ideas

  • Unified Modeling of Temporal Order and Mutual Exclusivity: The tree structure elegantly captures both the sequential nature of cancer evolution and the phenomenon of mutually exclusive mutations within sets of genes.
  • Scalability to Large Datasets: The combination of efficient likelihood computation and sophisticated MCMC moves allows ToMExO to handle datasets with hundreds of genes and thousands of tumors, a significant improvement over some previous methods.
  • Biological Relevance of Inferred Structures: The model successfully identifies known mutually exclusive oncogenes and suggests novel, statistically significant relationships in real cancer data, providing new hypotheses for further research.
  • Robustness to Errors: The explicit modeling and estimation of false positive and false negative mutation rates make the method more robust to the inherent noise in genomic data.

Key Results

  • Synthetic Data Performance: ToMExO accurately reconstructs the underlying generative models from synthetic data, achieving high F-scores (harmonic mean of precision and recall) for identifying mutual exclusivity and progression relationships, especially with error rates up to 0.05 and at least 100 tumors.
  • Biological Data Analysis:
    • UCEC: Identified known mutually exclusive genes like TP53 and CTNNB1, and suggested a significant progression signal from PI3K pathway genes (PIK3CA, PIK3R1) to PTEN.
  • CHOL: Despite a small dataset, the model found interesting perfect mutual exclusivity and progression relations, with zero inferred error parameters.
  • GBM: Recovered PI3K pathway genes in a first-layer node and identified progression signals from TP53 to IDH1 and RB1, and from IDH1 to ATRX.
  • Likelihood Ratios: The inferred models for biological datasets showed significantly higher likelihoods compared to a simple "star tree" model, indicating a better fit to the data. For UCEC, the per-tumor likelihood ratio was 5.49, and for GBM, it was 1.22.

Limitations and Future Work

  • Computational Complexity: While significantly improved, the MCMC approach can still be computationally intensive for extremely large and complex datasets.
  • Model Selection: The paper mentions that flexible MCMC moves help resolve model selection issues, but the inherent complexity of exploring the space of tree structures remains a challenge.
  • Future Work: The authors suggest that their method can be extended to incorporate other types of genomic alterations (e.g., copy number variations) and to analyze larger, more complex datasets. The identified novel progression signals in biological data warrant further experimental validation.

Summary updated

Abstract

Abstract Identifying cancer driver genes and their interrelations is critical in understanding cancer progression mechanisms. In this paper, we introduce ToMExO, a probabilistic method to infer cancer driver genes and how they affect each other, using cross-sectional data from cohorts of tumors. We model cancer progression dynamics using a tree with sets of driver genes in the nodes. This model explains the temporal orders among driver mutations and their mutual exclusivity patterns. We introduce a dynamic programming procedure for the likelihood calculation and build an MCMC inference algorithm. Together with our engineered MCMC moves, our efficient likelihood calculations enable us to work with datasets having hundreds of genes and thousands of tumors in the datasets. We demonstrate our method’s performance on several synthetic datasets covering various scenarios for cancer progression dynamics. We then present the analyses of several biological datasets using the ToMExO method and validate the results using a set of method-independent metrics.
BibTeX
@misc{neyshabouri2022tree,
  author = {Mohammadreza Mohaghegh Neyshabouri and Jens Lagergren},
  title = {{Tree of Mutually Exclusive Oncogenes}},
  year = {2022},
  doi = {10.1101/2022.02.07.479425},
  howpublished = {bioRxiv},
}
Figure