TL;DR
Cancer genomes are rife with structural variations (SVs), but existing detection methods struggle with their complexity and heterogeneity. This paper introduces COMSV, a novel optical DNA mapping-based method specifically designed to accurately identify SVs in cancer samples, even those with low tumor purity or complex sub-clonal architectures. COMSV demonstrates superior performance compared to existing tools, enabling the discovery of hundreds of novel SVs per sample.
What Problem Does This Paper Address?
Structural variations (SVs) are a hallmark of cancer genomes, playing a significant role in gene amplification, deletion, and fusion, which can drive tumorigenesis. Detecting these SVs is crucial for understanding cancer development, diagnosis, and classification. However, traditional short-read sequencing methods face challenges in accurately identifying large SVs due to limitations in read alignment and determining the full genomic span affected. Optical DNA mapping, with its long read lengths (hundreds of kilobases), offers a powerful alternative for detecting large SVs.
Despite the promise of optical mapping, existing SV calling algorithms were primarily developed for non-cancer samples. Cancer genomes present unique challenges:
- Cellular Heterogeneity: Cancer samples often contain a mix of cancer cells, normal cells, immune cells, and stromal cells. This means SVs specific to cancer cells might be supported by only a small fraction of the DNA molecules, making them difficult to detect with methods assuming uniform cell populations.
- Sub-clones: Cancer genomes can harbor multiple sub-clones, each with its own set of SVs, further complicating the detection of rare events.
- Copy Number Aberrations: Cancer genomes frequently exhibit abnormal copy numbers, leading to varying proportions of SV-supporting molecules across the genome, which can mislead standard SV callers.
- Increased SV Complexity and Errors: Cancer genomes tend to have more SVs, and these SVs are often more complex, leading to more errors in optical map alignments and de novo assemblies, which can impact the reliability of SV calls.
This paper addresses the need for an SV detection method that can overcome these specific challenges inherent in cancer genomics.
Methods
The authors introduce COMSV (Cancer Optical Mapping for detecting Structural Variations), a novel pipeline designed to address the complexities of cancer genomes when using optical DNA mapping data. COMSV comprises two main pipelines: an "indel pipeline" for insertions and deletions, and a "complex SV pipeline" for other SV types.
Key methodological innovations include:
- Normalization and Robust Alignment Handling: COMSV normalizes distances between labels on DNA molecules to account for linearization variations. Crucially, it employs a flexible approach to handle abnormal distance deviations, either identifying potential indels, discarding erroneous alignments, or correcting distances based on consensus. This is vital for cancer data, which often has higher alignment error rates.
- Candidate Region-Based Indel Calling: Instead of analyzing label pairs individually, COMSV groups overlapping label pairs into candidate indel regions. This allows for the inclusion of molecules with missing labels near SV breakpoints and reduces redundancy by considering all pairs within a region simultaneously.
- Flexible Clustering for Zygosity and Type Determination: COMSV clusters inter-label distances to identify SV types and zygosity. Its flexible clustering approach is designed to be sensitive to indels regardless of the a priori unknown proportion of supporting molecules, accommodating variations due to mixed cell types, sub-clones, and copy number changes.
- Complex SV Detection: The complex SV pipeline leverages split-aligned molecules, partially unaligned molecules, and de novo assembled contigs. It infers SV types from split-alignment locations and orientations and uses re-alignment of unaligned parts to detect SVs missed by split alignments.
- Integrated Annotation: COMSV provides a final set of non-redundant SVs with detailed annotations, considering redundancy, breakpoint uncertainty, and potential functional consequences.
The method is made available as an open-source tool.
Datasets
The paper utilizes both simulated and real-world datasets for evaluation:
- Simulated Data:
- Simulated indels with varying proportions of SV-supporting molecules (10-90%) to test precision and recall under different support levels.
- Simulated data based on a cell evolution graph, creating four groups of cells (normal, trunk cancer, two sub-clones) and then generating three mixture scenarios of increasing complexity (high tumor content, low tumor content with one sub-clone, low tumor content with two sub-clones). This benchmark is designed to specifically test performance on heterogeneous cancer samples.
- Simulated data based on Direct Labeling Enzyme 1 (DLE-1) recognition sites to evaluate performance with DLS labeling technology.
- Real Data:
- Optical mapping data from cancer cell lines and patient tumor samples.
- Previously published optical mapping data.
- Newly generated optical mapping data.
- Matched blood samples from cancer patients were used to identify cancer-specific SVs and validate COMSV's calls.
Key Insights and Ideas
- Cancer Genomics Demands Specialized Tools: The paper strongly argues that standard SV calling methods are insufficient for cancer samples due to their inherent biological complexity (heterogeneity, sub-clones, copy number variations). This highlights a critical gap in current genomic analysis tools.
- Optical Mapping is Ideal for Cancer SVs: The long read lengths of optical mapping are particularly well-suited for resolving the large and complex structural rearrangements common in cancer, offering a cost-effective alternative to some long-read sequencing approaches.
- Low Support is Not Necessarily Noise: COMSV's ability to detect SVs supported by a small fraction of molecules is a key differentiator. This is crucial because true cancer-specific SVs might be present in only a subset of cells, and COMSV's robust statistical framework can distinguish these from alignment artifacts.
- De Novo Assembly and Re-alignment Enhance Detection: The complex SV pipeline's use of de novo assembly and targeted re-alignment of unaligned segments allows for the detection of SVs that might be missed by methods relying solely on split alignments.
- Novel SVs in Cancer Genes: The identification of novel SVs, particularly those overlapping genes like ZNF429 and CSMD1, which have known roles in cancer, suggests potential new avenues for understanding cancer mechanisms and therapeutic targets.
Key Results
- Superior Performance on Simulated Cancer Data: COMSV significantly outperformed existing methods (Bionano Solve, OMIndel, OMSV) on simulated heterogeneous cancer samples, especially in complex scenarios with sub-clones. Its F1 scores were consistently higher across various SV types (deletions, insertions, inversions, duplications).
- High Precision and Recall: COMSV maintained high precision even with low proportions of SV-supporting molecules (<15%) and showed a much slower deterioration in recall compared to other methods as molecule support decreased.
- Robustness to Alignment Quality: COMSV's performance was only minimally affected when provided with noisy, real-world alignments compared to artificially perfect alignments, indicating its resilience to common issues in optical mapping data.
- Identification of Novel SVs: Application to real cancer samples revealed hundreds of novel SVs per sample, many of which were independently identified across different samples and cancer types.
- Validation Against Previous Studies: Approximately 80% of SVs identified by COMSV had support from previous studies, demonstrating high concordance, while a significant portion (at least 6.7% per sample) represented novel findings.
Limitations and Future Work
- Validation of Novel SVs: While COMSV identifies novel SVs, further experimental validation and functional studies are needed to confirm their significance, especially for those identified independently across multiple samples.
- Comparison with DLS Data: The paper notes that while both Nt.BspQI and DLE-1 (DLS) data were used, the two labeling methods can produce different densities and patterns of labels, potentially leading to inconsistencies. Further work could explore optimizing COMSV for each labeling method or developing a unified approach.
- Comprehensive Catalog: The authors express a desire to apply COMSV to a larger compendium of cancer samples to create a comprehensive catalog of large and complex SVs across various cancer types.
- Application to Other Species: COMSV's potential application to other species with high-quality reference genomes, such as mouse models, is suggested for broader cancer research.
- Functional Significance: Investigating the functional consequences of the identified SVs, particularly the novel ones, is a key area for future research.
Summary updated