Identification of key somatic oncogenic mutation based on a confounder-free causal inference model

Yijun Liu, Ji Sun, Huiyan Sun, Yi ChangPLOS Computational Biology 2022 · DOI

TL;DR

This paper introduces CEBP, a novel causal inference framework that uses a deep variational autoencoder to identify key oncogenic mutations by disentangling correlation from causation. It reveals that high mutation frequency doesn't necessarily equate to high oncogenic impact, offering a more nuanced understanding of cancer drivers.

What Problem Does This Paper Address?

Cancer development is a complex process driven by genetic mutations. Identifying "driver" mutations—those that actively promote cancer growth—is crucial for understanding disease mechanisms and developing targeted therapies. Traditional computational methods often rely on identifying associations (correlations) between mutations and cancer-related biological processes like cell proliferation and epithelial-mesenchymal transition (EMT). However, these associations can be misleading due to "confounders"—factors that influence both the mutation and the biological process, creating spurious correlations. Existing methods struggle to account for both observed and unobserved confounders, leading to potentially inaccurate identification of driver mutations. This paper aims to address this by employing causal inference to estimate the true causal effect of mutations, free from confounding biases.

Methods

The core of this work is the Causal Effect of a mutation on cancer Biological Process (CEBP) framework. CEBP integrates genomic and transcriptomic data and employs a Causal Effect Variational Autoencoder (CEVAE).

Here's a breakdown:

  • Structural Causal Model: The paper first defines a causal graph illustrating the relationships between mutations (treatment, M), biological processes (outcome, Y), observed confounders (X, e.g., other gene mutations), and unobserved confounders (Z, e.g., micro-environment stress).
  • Estimating Biological Process Activity: Since direct measurements of biological process activity aren't available, the paper proposes a method to estimate it. This involves identifying "core genes" within a pathway based on correlation, and then using linear regression to relate each sample's gene expression to the average expression of these core genes. This estimated activity serves as the outcome (Y).
  • Causal Effect Variational Autoencoder (CEVAE): This is the novel component. It's a deep generative model that learns a representation of the unobserved confounders (Z) from the observed data (X, M, Y).
  • Encoder (Inference Network): Learns the posterior distribution of Z given X, M, and Y. It uses a two-stage neural network to capture complex relationships and disentangle the effects of mutated vs. non-mutated genes.
  • Decoder (Model Network): Reconstructs the observed data from the learned latent representation of Z.
  • Average Treatment Effect (ATE): By modeling the causal graph and accounting for confounders (both observed and unobserved), CEBP estimates the ATE, which represents the true causal impact of a mutation on a biological process. A positive ATE suggests the mutation promotes the process, while a negative ATE suggests it inhibits it.

Datasets

The study utilizes data from The Cancer Genome Atlas (TCGA). Specifically, they collected genomics (mutation) and transcriptomics (RNA-seq) data for tumor samples across 10 different cancer types. The specific cancer types are listed in Table 3:

  • Breast invasive carcinoma (BRCA)
  • Kidney renal clear cell carcinoma (KIRC)
  • Thyroid carcinoma (THCA)
  • Head and Neck squamous cell carcinoma (HNSC)
  • Lung adenocarcinoma (LUAD)
  • Liver hepatocellular carcinoma (LIHC)
  • Lung squamous cell carcinoma (LUSC)
  • Esophageal carcinoma (ESCA)
  • Kidney renal papillary cell carcinoma (KIRP)
  • Bladder Urothelial Carcinoma (BLCA)

Key Insights and Ideas

  • Causation Over Correlation: The fundamental insight is that identifying driver mutations requires understanding their causal impact, not just their association with cancer phenotypes. CEBP provides a framework to achieve this.
  • Mutation Frequency is Misleading: A key finding is that genes with high mutation rates do not necessarily have the strongest causal effect on cancer progression. This challenges the common heuristic of prioritizing highly mutated genes.
  • Unobserved Confounders Matter: The framework's ability to model and account for unobserved confounders is critical for accurate causal inference in complex biological systems.
  • Dual Roles of Mutations: Some mutations can have opposing effects on different biological processes (e.g., promoting DNA replication while suppressing EMT), which could explain why targeted therapies are not always effective.
  • New Driver Candidates: The method not only validates known driver mutations but also identifies novel candidates that warrant further investigation.

Key Results

  • Identification of Key Mutations: CEBP successfully identified top 10 mutations with the highest causal effect on DNA replication and EMT for each of the 10 cancer types.
  • Consistency with Literature: The identified key mutations largely align with previously reported findings and experimental evidence.
  • No Strong Correlation Between Mutation Rate and ATE: Figures 3 and 4, and Tables 1 and 2, demonstrate that mutations with high ATE values often have low mutation rates, and vice-versa. For example, in LUAD, TTN has a high mutation frequency but a low causal effect on DNA replication.
  • Statistical Significance: Mann-Whitney U tests (Figs. 5-8) confirmed that the biological process activities (DNA replication and EMT) were significantly different between mutated and non-mutated groups for genes identified by CEBP as having high ATE. This validates the biological relevance of the identified mutations.
  • Comparison with Other Methods: CEBP's performance was found to be comparable to existing state-of-the-art driver mutation identification methods (Fig A in S1 Text).

Limitations and Future Work

  • Validation: While the identified mutations are consistent with existing literature, the authors acknowledge that wet experiments are needed for definitive validation of the predicted negative causal effects and novel candidates.
  • Single Gene Mutations: The current framework focuses on the causal effect of single gene mutations. Future work aims to develop models for estimating the bundling effect of candidate mutation sets, which is computationally challenging due to the combinatorial explosion of possibilities.
  • Unconfoundedness and Positivity Assumptions: The paper notes that high-dimensional confounders can make it harder to satisfy the unconfoundedness assumption and increase the risk of violating the positivity assumption (where every individual has a non-zero probability of receiving any treatment). Future research will focus on balancing these assumptions.
  • Data Availability: The authors mention that the availability of comprehensive databases for validating key mutations is limited, which hinders direct comparison with other methods.

Summary updated

BibTeX
@article{liu2022identification,
  author = {Yijun Liu and Ji Sun and Huiyan Sun and Yi Chang},
  title = {{Identification of key somatic oncogenic mutation based on a confounder-free causal inference model}},
  journal = {PLOS Computational Biology},
  year = {2022},
  doi = {10.1371/journal.pcbi.1010529},
}
Figure