DGAT-onco: A powerful method to detect oncogenes by integrating differential mutational analysis and functional impacts of somatic mutations

Haoyang Zhang, Junkang Wei, Zifeng Liu, Xun Liu, Yutian Chong, Yutong Lu, Huiying Zhao, Yuedong YangbioRxiv 2020 · DOI

TL;DR

This paper introduces DGAT-onco, a novel computational method for identifying oncogenes by integrating mutation frequency and functional impact analysis. By comparing somatic mutations in cancer patients with germline mutations in healthy individuals, DGAT-onco offers a more robust approach to discovering cancer-driving genes, outperforming existing methods.

What Problem Does This Paper Address?

Discovering oncogenes – genes that drive cancer development when mutated – is crucial for understanding cancer mechanisms and developing new therapies. Traditional methods often rely solely on the frequency of mutations within cancer populations. However, this approach has limitations: it can miss genes with fewer but functionally significant mutations, and it's susceptible to biases from background mutation rates that vary across the genome and between individuals. Many more oncogenes likely exist that are not frequently mutated but still play critical roles. The challenge is to develop methods that can identify these less obvious oncogenes by considering not just how often a gene is mutated, but also how impactful those mutations are, and by using a more accurate baseline for comparison.

Methods

DGAT-onco tackles this by employing a two-pronged approach:

  1. Differential Mutational Profile Analysis: It compares the mutational patterns of genes between cancer patients and a healthy population. This helps identify genes that show a significantly different mutation profile in cancer.
  2. Functional Impact Integration: Instead of treating all mutations equally, DGAT-onco weighs mutations based on their predicted functional impact. It utilizes a suite of 23 different pathogenic scoring systems (from the dbNSFP database) to quantify how likely a mutation is to be damaging. The method then sums these scores for all mutations within a gene for each individual.

The core of DGAT-onco involves creating a mutation matrix where rows represent genes and columns represent individuals. For each gene-individual pair, a "summed functional impact score" is calculated. This matrix is then normalized, and the distribution of these normalized scores is compared between the cancer (TCGA) and natural (1000 Genomes) populations using the unidirectional Earth Mover's Distance (uEMD). Statistical significance is determined through permutation testing and Q-value calculation. The authors also explored various pathogenic scoring functions, finding that M-CAP (Mendelian Clinically Applicable Pathogenicity) yielded the best performance.

Datasets

  • TCGA Dataset: Somatic mutation data from 33 cancer types, collected through standardized variant calling pipelines. This served as the primary dataset for training and evaluation.
  • 1000 Genomes Project (1000G) Dataset: Germline mutation data from 2504 individuals, used as the "natural population" baseline for comparison.
  • TS19 Dataset: An independent test set comprising somatic mutations from 19 cancers sourced from cBioPortal, used to validate the method's reliability on data outside of TCGA.

Key Insights and Ideas

  • Beyond Frequency: The study strongly emphasizes that simply counting mutations isn't enough. Integrating functional impact scores is crucial for a more accurate and comprehensive oncogene discovery.
  • Germline as a Better Baseline: Using germline mutations from a healthy population (1000 Genomes) provides a more accurate and less biased background against which to compare cancer mutations, compared to methods that only use cancer data.
  • The Power of Pathogenic Scores: Different pathogenic scoring systems have varying strengths. The authors systematically evaluated many of them, highlighting M-CAP as particularly effective for this task.
  • Complementarity: DGAT-onco identifies a distinct set of potential oncogenes compared to other methods, suggesting it can uncover candidates missed by frequency-based or other functional impact approaches.

Key Results

  • Superior Performance: DGAT-onco achieved the highest average Area Under the Precision-Recall Curve (AUPRC) in both the TCGA (0.208) and TS19 (0.215) datasets, outperforming 6 alternative methods.
  • Significant Validation: 22.8% of the significant genes identified by DGAT-onco were verified as known oncogenes by the Cancer Gene Census (CGC).
  • M-CAP Advantage: The M-CAP scoring function, when integrated into DGAT-onco, consistently yielded better performance across various cancer types compared to other scoring functions.
  • Broad Applicability: The method demonstrated strong performance across 33 different cancer types in TCGA and was validated on an independent set of 19 cancers.
  • Functional Enrichment: Genes predicted by DGAT-onco were significantly enriched in cancer-related pathways, such as "Pathways in cancer" and "Signaling by Receptor Tyrosine Kinases."

Limitations and Future Work

  • Non-coding Mutations: DGAT-onco currently focuses on non-synonymous coding mutations. The authors acknowledge that mutations in non-coding regions (like 3'-UTR and 5'-UTR) can also contribute to cancer and suggest developing new metrics for these regions or exploring frequency-based methods for their inclusion.
  • Single-Gene Focus: The method analyzes genes individually. The authors note that cancer is often driven by a network of genes and suggest future work could involve integrating gene interaction networks or clustering methods for a more holistic interpretation.
  • Pathogenicity Score Development: The performance of DGAT-onco is dependent on the quality of available pathogenicity classifiers. As these classifiers improve, the method's predictive power is expected to increase.

Summary updated

Abstract

Abstract Motivation Oncogenes are genes whose malfunctions play critical roles in cancer development, and their discovery is a major aim of cancer mechanisms study. By counting the mutation frequency, oncogenes have been identified with frequent mutations, while it is believed that many more oncogenes could be discovered by differential mutational profile analysis. However, it is common that current methods only utilize mutations in the cancer population, which have an obvious bias in background mutation modelling. Methods To predict oncogenes efficiently, we developed a method, DGAT-onco that analyzed the frequency distribution and functional impacts of mutations in both cancer and natural population. Our method can capture the mutational difference of two population, and provide a comprehensive view of genomics basis underlying cancer development. DGAT-onco was constructed by germline mutations from the 1000 Genomes project and somatic mutations of 33 cancer types from the Cancer Genome Atlas (TCGA) dataset. Its reliability was verified on an independent test set including 19 cancers from other sources. Results We demonstrated that our method is more effective than alternative methods in oncogenes discovering. Using this approach achieves higher classification performance in oncogene discovery than 6 alternative methods, and 22.8% significant genes identified by our method were verified as oncogenes by the Cancer Gene Census (CGC). Availability DGAT-onco is available at https://github.com/zhanghaoyang0/DGAT-onco . Contact [email protected] or [email protected]
BibTeX
@misc{zhang2020dgat,
  author = {Haoyang Zhang and Junkang Wei and Zifeng Liu and Xun Liu and Yutian Chong and Yutong Lu and Huiying Zhao and Yuedong Yang},
  title = {{DGAT-onco: A powerful method to detect oncogenes by integrating differential mutational analysis and functional impacts of somatic mutations}},
  year = {2020},
  doi = {10.1101/2020.02.15.947085},
  howpublished = {bioRxiv},
}

Loading related content...

Figure