TL;DR
This paper introduces NIMBus, a novel integrative method that uses negative binomial regression to accurately identify mutation hotspots in cancer genomes. By accounting for both genomic features and individual-level mutation rate heterogeneity, NIMBus offers a more robust approach to discovering cancer drivers and understanding their regulatory mechanisms.
What Problem Does This Paper Address?
Identifying frequently mutated regions in cancer genomes is crucial for understanding cancer progression and discovering potential therapeutic targets. However, this task is complicated by two major challenges:
- Mutation Rate Heterogeneity: Cancer genomes are not uniform. Mutation rates vary significantly across different individuals and even within different regions of the same genome. This heterogeneity, influenced by factors like replication timing and chromatin organization, makes it difficult to assume a constant mutation rate. Traditional methods that assume uniformity can lead to poor data fitting and an overabundance of false positives.
- Integrating Genomic Features: Numerous genomic features (e.g., chromatin status, replication timing) are known to influence mutation processes. Effectively incorporating these features into mutation burden analysis is essential for accurate identification of driver mutations. However, existing methods may be limited by the number and types of features they can integrate, and may struggle with data availability across different tissues or cell types.
The authors aim to address these challenges by developing a method that can simultaneously model mutation rate heterogeneity and integrate a comprehensive set of genomic features.
Methods
NIMBus employs a multi-pronged approach to tackle mutation burden analysis:
- Negative Binomial Regression: At its core, NIMBus uses a negative binomial distribution to model mutation counts. This is a key departure from simpler Poisson models, as the negative binomial distribution inherently accounts for overdispersion, a common characteristic of mutation count data due to biological variability.
- Gamma-Poisson Mixture Model: To specifically address the heterogeneity across individuals, NIMBus models individual mutation rates as random variables following a Gamma distribution. This, combined with a Poisson process for mutation counts within individuals, results in a negative binomial distribution for the pooled mutation counts.
- Integrative Covariate Matrix: NIMBus leverages a vast array of genomic features (381 in total) from resources like REMC and ENCODE. These features, which describe chromatin organization, expression profiles, replication status, and more, are used to build a covariate matrix. This matrix is then used in a negative binomial regression to predict local mutation rates with high precision.
- Principal Component Analysis (PCA): To handle the high dimensionality and potential multicollinearity of the 381 features, PCA is applied to the covariate matrix. This reduces the feature space while retaining the most important sources of variation.
- Network Analysis: For analyzing pathways, NIMBus extends its capabilities by treating genes within a pathway as a network. It calculates mutation burden for individual genes and then combines these p-values using Fisher's method to derive a pathway-level significance score.
Datasets
- Whole Genome Sequencing (WGS) Data: The study utilized WGS variant calls from 649 cancer genomes across seven different cancer types: breast cancer (BRCA), gastric cancer (GACA), liver cancer (LICA), lung adenocarcinoma (LUAD), prostate cancer (PRAD), medulloblastoma (MB), and pilocytic astrocytoma (PA).
- Genomic Feature Data: Comprehensive functional genomic data was sourced from the Roadmap Epigenomics Mapping Consortium (REMC) and the Encyclopedia of DNA Elements (ENCODE) project. This included signal track files for histone modification marks, chromatin status, methylation, mRNA-seq, replication timing, and GC percentage.
- Annotation Data: GENCODE annotation v19 was used for coding regions. For noncoding regions, a comprehensive list of annotations was collected, including promoter regions, transcription start sites (TSS), translated regions (UTR), transcription factor binding sites (TFBS), enhancers, ultra-conserved, and ultra-sensitive sites.
- Cancer Gene Census: The COSMIC database was used to classify burdened genes based on existing literature support.
Key Insights and Ideas
- Overdispersion is Key: The paper strongly emphasizes that simply assuming a constant mutation rate (Poisson distribution) is insufficient for cancer genomics. The inherent overdispersion in mutation counts, driven by biological heterogeneity, necessitates more sophisticated models like the negative binomial distribution.
- Individual Heterogeneity Matters: Treating mutation rates as random variables with a Gamma distribution effectively captures the variability between individuals, leading to more accurate modeling of pooled mutation counts.
- Comprehensive Feature Integration is Powerful: By integrating a large number of diverse genomic features, NIMBus achieves high precision in predicting local mutation rates, highlighting the complex interplay between genomic context and mutation processes.
- Noncoding Regions Hold Significant Drivers: NIMBus successfully identified numerous noncoding regions with significant mutation burden, including well-known drivers like the TERT promoter and novel candidates, underscoring the importance of analyzing these regions for cancer driver discovery.
- Network-Based Analysis Uncovers Pathway Significance: Applying NIMBus to KEGG pathways demonstrates its utility in identifying biologically meaningful networks that are significantly mutated, offering insights into disrupted cellular processes.
Key Results
- Improved Model Fitting: NIMBus, using a negative binomial distribution, showed significantly improved fitting to mutation count data compared to models assuming a constant mutation rate (Figure 2B).
- Accurate Local Mutation Rate Prediction: Regression against PCA-transformed genomic features achieved high correlation between predicted and observed mutation rates, ranging from 0.668 to 0.958 across different cancer types (Figure 3B).
- Controlled P-value Inflation: In coding region analysis, NIMBus effectively controlled P-value inflation, identifying a manageable number of significantly mutated genes (11 in LUAD) compared to methods assuming constant mutation rates (6,023 genes) (Table 1, Figure 4).
- Identification of Known and Novel Drivers: NIMBus identified well-documented cancer drivers like TP53 and TERT promoter mutations, as well as numerous novel burdened genes and noncoding regions.
- Benchmarking Success: NIMBus outperformed other methods like OncodriveFML and the PCAWG Consortium's analysis in identifying burdened noncoding regions, identifying a substantially larger number of unique burdened genes. For example, NIMBus identified 1,380 unique burdened genes in noncoding regions compared to 249 by OncodriveFML.
Limitations and Future Work
- Limited Statistical Power: The authors acknowledge that the limited availability of WGS data in some analyses can lead to appropriately conservative predictions, suggesting that future work with more data will enable the identification of a more complete list of heavily burdened genes.
- Bin Size Optimization: The choice of bin size for analysis is an important parameter, and optimal selection remains a challenge that may require case-by-case investigation. The paper used a 1 Mb bin size for all cancer types.
- Pathway Annotation Constraints: Challenges in interpreting some identified pathways may arise from the constrained availability of pathway annotations, potentially leading to false positives. Future work could focus on building and validating more comprehensive pathway annotations.
- Independence of P-values in Fisher's Method: When combining P-values using Fisher's method, the assumption of independence might not always hold, potentially leading to some false positives. This could be an area for future study.
- Extensibility: The NIMBus model is designed to be flexible and can be retrained with new covariates, sequencing results, and target regions, suggesting future work could involve applying it to different types of genomic data or diseases.
Summary updated