TL;DR
This paper introduces NetNorM, a novel method that leverages gene interaction networks to transform raw somatic exome mutation data into a more informative representation. NetNorM significantly improves the prediction of patient survival and enables more robust patient stratification compared to existing methods, offering a more nuanced view of cancer's genetic landscape.
What Problem Does This Paper Address?
Cancer development is driven by genetic alterations, particularly somatic mutations. While next-generation sequencing has generated vast amounts of tumor mutation data, analyzing it is challenging. Key issues include:
- Sparsity and Noise: Most mutations are "passenger" events, not directly driving cancer, and they obscure the "driver" mutations.
- Low Overlap: Different patients often have very different sets of mutations, making direct comparisons difficult.
- High Dimensionality: The sheer number of genes and potential mutations creates a complex data space.
To address these challenges, researchers have explored using gene interaction networks as prior knowledge. The idea is that if a gene is mutated, its interacting neighbors might also be affected. This paper aims to improve upon existing network-based approaches by developing a method that better captures cancer-relevant information from mutation data.
Methods
The core of this paper is the NetNorM method. It takes raw binary mutation profiles (where a '1' indicates a mutation in a gene for a patient) and an undirected gene network as input. NetNorM transforms these profiles into a new binary representation with a fixed number of mutations, k, per patient. This normalization is achieved through two main mechanisms:
- Adding "Proxy" Mutations: For patients with fewer than k mutations, NetNorM imputes mutations in non-mutated genes that have a high number of mutated neighbors in the network. These are "proxy" mutations, suggesting that the gene's function might be indirectly affected due to its mutated network context.
- Removing "Non-essential" Mutations: For patients with more than k mutations, NetNorM removes mutations from genes with a low degree (few neighbors) in the network, assuming these mutations have a less significant impact.
The key parameter k can be tuned, often set to the median number of mutations in the cohort or optimized via cross-validation.
The paper also compares NetNorM to:
- Raw Binary Mutation Data: The original, un-normalized mutation profiles.
- Network Smoothing with Quantile Normalization (NSQN): A common approach that diffuses mutation information across the network and then normalizes the resulting values.
- Simplified NSQN (SimpNSQN): A version of NSQN that only considers direct neighbors.
Datasets
The study utilizes data from The Cancer Genome Atlas (TCGA), specifically analyzing whole-exome somatic mutation profiles from 8 different cancer types:
- LUAD (Lung adenocarcinoma)
- SKCM (Skin cutaneous melanoma)
- GBM (Glioblastoma multiform)
- BRCA (Breast invasive carcinoma)
- KIRC (Kidney renal clear cell carcinoma)
- HNSC (Head and Neck squamous cell carcinoma)
- LUSC (Lung squamous cell carcinoma)
- OV (Ovarian serous cystadenocarcinoma)
A total of 3,278 samples were analyzed. For gene networks, the study primarily used Pathway Commons, a comprehensive database of pathway and molecular interactions. Other networks (BioGRID, HPRD, HumanNet, STRING) were also used for sensitivity analysis.
Key Insights and Ideas
- Network Topology Matters: The study strongly suggests that the structure of gene interaction networks provides crucial biological information that can significantly enhance the analysis of mutation data.
- Proxy Mutations Capture Broader Context: NetNorM's "proxy" mutations, imputed in genes with many mutated neighbors, effectively capture information about a patient's overall mutational burden and the local mutational burden within gene networks. This is a key differentiator from simply using raw mutation data.
- Fixed Mutation Count is Beneficial: Normalizing all patients to have the same number of mutations (k) appears to be a beneficial property for downstream statistical models, potentially by providing a more consistent input for methods like survival SVMs and NMF.
- Local Network Information is Sufficient: NetNorM's success, which focuses on direct neighbors, suggests that extensive diffusion across the network (as in NSQN) might not be necessary and could even introduce noise.
Key Results
- Improved Survival Prediction: NetNorM significantly improved the Concordance Index (CI) for survival prediction in LUAD (from 0.56 to 0.62) and SKCM (from 0.48 to 0.61) compared to raw data and NSQN. For SKCM, both NetNorM and NSQN were significantly better than raw data.
- Robust Patient Stratification: NetNorM consistently produced more stable and clinically relevant patient subtypes across multiple cancer types compared to raw data and NSQN. These subtypes often showed significantly distinct survival outcomes.
- Network Randomization Impact: Using randomized gene networks led to significantly worse performance for NetNorM in survival prediction and patient stratification, underscoring the importance of the biological information encoded in real networks.
- Prognostic Power of Proxy Genes: The study identified "proxy genes" that, despite not being directly mutated in a patient, were selected by survival models and contributed to better predictions, highlighting their prognostic value.
- Complementary to Clinical Data: Combining NetNorM-processed mutation data with clinical features further improved survival prediction for LUAD and SKCM, suggesting that mutation information offers complementary insights.
Limitations and Future Work
- Cancer Type Variability: NetNorM showed significant improvements for LUAD and SKCM, but less so for other cancer types. The authors suggest that further research is needed to design representations that increase statistical power for these datasets.
- Hyperparameter Tuning: The performance of NetNorM and NSQN can be sensitive to hyperparameter choices (k for NetNorM, α for NSQN). While cross-validation is used, finding optimal parameters for all tasks and cancer types remains an area for exploration.
- Integration of Functional Impact: NetNorM currently focuses on mutation presence/absence and network topology. The authors suggest that integrating information about the predicted functional impact of mutations could be a valuable extension.
- Empirical Distribution Choices: The choice of the normalized mutation profile distribution (fixed number of mutations for NetNorM, quantile normalization for NSQN) is empirical. Future work could explore optimizing these distributions for specific tasks.
- Network Choice: While Pathway Commons was the primary network, the study notes that no single network clearly stands out as best for all cancers, suggesting that network selection might be task- or cancer-specific.
Summary updated