Abstract
We present a novel hierarchical Bayes statistical model, xseq, to systematically quantify the impact of somatic mutations on expression profiles. We establish the theoretical framework and robust inference characteristics of the method using computational benchmarking. We then use xseq to analyse thousands of tumour data sets available through The Cancer Genome Atlas, to systematically quantify somatic mutations impacting expression profiles. We identify 30 novel cis-effect tumour suppressor gene candidates, enriched in loss-of-function mutations and biallelic inactivation. Analysis of trans-effects of mutations and copy number alterations with xseq identifies mutations in 150 genes impacting expression networks, with 89 novel predictions. We reveal two important novel characteristics of mutation impact on expression: (1) patients harbouring known driver mutations exhibit different downstream gene expression consequences; (2) expression patterns for some mutations are stable across tumour types. These results have critical implications for identification and interpretation of mutations with consequent impact on transcription in cancer.
Introduction
The genomic revolution has enabled the cataloging of millions of mutations across various cancer types. However, a significant challenge remains: distinguishing functional "driver" mutations from neutral "passenger" mutations. Traditional methods often rely on sequence conservation or mutation frequency, frequently overlooking the functional consequences within the cellular environment. This study introduces the concept of a "transcriptional shadow"—the measurable changes in gene expression resulting from functional mutations. By quantifying these effects, we can better understand the molecular context of cancer-driving events.
Methods: The xseq Framework
To map this transcriptional shadow, we developed xseq, a generative probabilistic model that links somatic mutations to gene expression profiles. The model integrates three primary data types: a mutation matrix, a gene expression matrix (mRNA levels), and a gene influence graph representing known functional relationships.
The xseq model distinguishes between two types of functional impacts:
- Cis-effects: Mutations affecting the expression of the mutated gene itself, often through mechanisms like nonsense-mediated decay.
- Trans-effects: Mutations in one gene (e.g., a transcription factor) that disrupt an entire downstream network of connected genes.
The statistical backbone of xseq is a hierarchical Bayes approach utilizing three-component mixture models to represent downregulation, neutral, and upregulation states. Inference is performed using the Belief Propagation algorithm, while parameters are learned via Expectation Maximization (EM).
Results
We applied xseq to data from 2,786 patients across 12 tumor types from The Cancer Genome Atlas (TCGA). The model demonstrated high precision, achieving an area under the curve (AUC) of 0.99 in computational benchmarks.
Novel Driver and Suppressor Identification
xseq identified 150 genes impacting expression networks in trans, including 89 novel predictions. Furthermore, the model identified 30 novel cis-effect tumor suppressor candidates. These candidates were significantly enriched in loss-of-function mutations and biallelic inactivation, validating the model's ability to find functionally relevant genes that frequency-based methods might miss.
Heterogeneity and Stability
A key discovery was the existence of "inert" driver mutations. In some patients, known drivers like CTNNB1 showed no downstream expression impact, particularly in ultramutated tumors. This suggests that even established drivers can act as passengers depending on the genomic context. Conversely, some mutations (e.g., in RB1 or NFE2L2) exhibited stable expression signatures across different tumor types, suggesting "hard-wired" oncogenic pathways that could be targeted consistently across organs.
Discussion and Future Directions
While xseq provides a powerful tool for functional genomics, it currently faces limitations in handling multiple mutations within the same gene and relies on the quality of the input influence graph. Future work will involve applying this logic to single-cell data to account for clonal heterogeneity and refining influence graphs as our understanding of protein-protein interactions grows. Ultimately, xseq bridges the gap between DNA sequence and functional outcome, offering a more nuanced view of the cancer genome.
Summary updated