TL;DR
This paper introduces TumorBoost, a novel normalization method for allele-specific copy number (ASCN) data from tumor samples. By leveraging matched normal samples, TumorBoost significantly improves the signal-to-noise ratio of allelic signals, making it easier to detect subtle genomic alterations like copy-neutral loss of heterozygosity (LOH) that are crucial for cancer research.
What Problem Does This Paper Address?
Cancer genomes are characterized by widespread genomic alterations, including changes in DNA copy number (CN). Genotyping microarrays are powerful tools for detecting these changes, not only by measuring total copy number (TCN) but also by quantifying the contribution of each allele (allelic composition). This allelic information is particularly valuable for identifying events like copy-neutral LOH, which are difficult or impossible to detect from TCN alone. However, raw allelic signal estimates from genotyping arrays are notoriously noisy and suffer from systematic effects, even after standard preprocessing. This noise makes it challenging to reliably identify and interpret these crucial genomic alterations. The paper addresses this gap by proposing a method to effectively normalize these noisy allelic signals, thereby enhancing the power to detect cancer-related genomic events.
Methods
The core of the TumorBoost method lies in its ability to leverage the information from a matched normal sample to correct for systematic biases in the tumor's allelic signals. The method works by:
- Estimating SNP-specific effects: It assumes that systematic biases (referred to as "SNP effects") are largely shared between the tumor and normal samples. These effects are estimated from the normal sample, where the true allele fractions (genotypes) are known or can be reliably inferred (e.g., 0, 1/2, or 1 for homozygous and heterozygous SNPs).
- Correcting tumor signals: These estimated SNP effects from the normal sample are then subtracted from the corresponding tumor allelic signals. This process effectively "normalizes" the tumor data, removing systematic variations that are not due to true biological changes.
- Handling homozygous and heterozygous SNPs differently: The method applies distinct correction strategies for homozygous and heterozygous SNPs, recognizing that their behavior and the nature of their allelic fractions differ. For homozygous SNPs, the correction is straightforward. For heterozygous SNPs, a more nuanced approach is taken, involving a scale factor that accounts for the observed deviation from the expected allele fraction.
TumorBoost is designed to be applied on a per-SNP basis and does not require prior knowledge of copy number change points or regions, making it a flexible preprocessing step. The implementation is available in the R package aroma.cn.
Datasets
The study primarily uses data from The Cancer Genome Atlas (TCGA) project. Specifically, they utilized:
- Ovarian (serous cystadenocarcinoma, OV) and Brain (glioblastoma multiforme, GBM) tumor-normal pairs.
- Affymetrix GenomeWideSNP_6 data: Raw CEL files were used, with allele-specific CRMA v2 preprocessing. Data preprocessed with RMA/median-polish was also used for comparison.
- Illumina Human1M-Duo and HumanHap550 data: XY-normalized and summarized data were used.
- Birdseed genotype calls: Used for evaluating robustness against genotyping errors.
The paper does not introduce new benchmark datasets but rather uses publicly available data to demonstrate the effectiveness of their method.
Key Insights and Ideas
- Shared SNP effects: The central idea that systematic biases in allelic signals are largely shared between tumor and normal samples is a key insight. This allows for effective correction using only a matched normal.
- Improved SNR for allelic data: TumorBoost significantly boosts the signal-to-noise ratio (SNR) of allelic data, making subtle genomic events much more detectable. This is crucial for identifying events like copy-neutral LOH.
- Single-pair analysis capability: The method's ability to work with a single tumor-normal pair is a significant practical advantage, enabling immediate analysis without the need for large reference panels or complex batch effect correction. This is particularly valuable for clinical diagnostics.
- No need for change point detection: TumorBoost normalizes data at the SNP level, independent of any prior knowledge of copy number alterations. This makes it a general preprocessing step that can be integrated into various downstream analysis pipelines.
Key Results
- Enhanced detection of CN events: TumorBoost significantly increases the power to detect somatic copy-number events, including copy-neutral LOH, compared to using raw or standard preprocessed allelic data.
- Improved sensitivity and specificity: The paper demonstrates that CN events are detected with greater sensitivity and specificity after TumorBoost normalization.
- Visual evidence of improvement: Figures 1, 2, and 3 visually show the dramatic reduction in noise and the sharpening of allele B fraction distributions after TumorBoost normalization.
- Quantitative improvements: ROC curve analyses and t-statistics presented in Tables 2, Figures 5, and 6 show substantial improvements in distinguishing between different copy number states after normalization. For example, normalized DHs consistently yield larger t-statistics, indicating better detection power.
- Robustness: The method shows good performance even with naive genotype calls, although using higher confidence genotype calls further improves results.
Limitations and Future Work
- Genotyping errors: While TumorBoost improves SNR, it can still produce outliers if SNPs are incorrectly genotyped (e.g., homozygous SNPs called heterozygous). The authors suggest that downstream segmentation methods are generally robust to these outliers, and methods to improve genotype calling accuracy are discussed.
- Mean level bias: TumorBoost aims to remove systematic effects across SNPs but does not claim to control for the absolute mean levels of allele B fractions, which can still be biased due to platform differences, preprocessing methods, or incomplete offset correction.
- Future Directions:
- Simultaneous segmentation of TCN and DH: The authors advocate for developing methods that jointly segment total copy number (TCN) and decrease in heterozygosity (DH) data, as both carry complementary information.
- Calling PCN states: Developing robust methods for calling copy number states, accounting for normal contamination and tumor ploidy, is identified as a next step.
- Emphasis on matched normals: The paper strongly recommends the routine collection of matched normal samples for cancer studies due to their critical role in quality control, somatic vs. germline event distinction, and improved signal quality.
Summary updated