TL;DR
This paper introduces a novel machine learning framework, Multi-Modal Correlated Topic Models (MMCTM), that integrates different types of genomic mutations (single nucleotide variants and structural variations) to more accurately identify cancer mutation signatures. This approach significantly improves signature discovery, especially in datasets with sparse mutation data, and enables better patient stratification for personalized treatment strategies.
What Problem Does This Paper Address?
Cancer genomes accumulate mutations from various sources, both external (like UV radiation) and internal (like DNA repair errors). These sources leave distinct "fingerprints" or "mutation signatures" that can reveal a tumor's origin, predict its behavior, and suggest therapeutic vulnerabilities. Traditionally, researchers analyze different types of mutations (e.g., single nucleotide variants or SNVs, and structural variations or SVs) separately. However, these mutation types can be interconnected and arise from shared underlying processes, particularly in cancers with DNA repair deficiencies. The challenge is to develop computational methods that can jointly analyze these diverse mutation types to uncover more comprehensive and accurate mutation signatures, leading to better understanding and treatment of cancer.
Methods
The core innovation of this paper is the development and application of Multi-Modal Correlated Topic Models (MMCTM). This is an extension of Latent Dirichlet Allocation (LDA), a technique commonly used in natural language processing to discover underlying "topics" in documents.
Here's how it works:
- Topic Models for Signatures: In this context, "topics" are analogous to mutation signatures, and "documents" are individual cancer samples. Each sample is represented as a mixture of these signatures.
- Correlated Topic Models (CTM): Unlike standard LDA, CTM accounts for correlations between these "topics" (signatures). This means if certain signatures tend to appear together in cancer genomes, CTM can capture that relationship.
- Multi-Modal: MMCTM extends CTM to handle multiple types of data ("modalities") simultaneously. In this paper, the two modalities are SNV counts and SV counts. MMCTM infers signatures from both SNVs and SVs independently but models their co-occurrence probabilities using a multivariate Gaussian distribution, allowing for cross-modality correlations.
- Novelty: The key novelty lies in the integration of different mutation types (SNVs and SVs) and the modeling of correlations between these signatures. This allows the model to leverage information from both mutation types, leading to more robust and accurate signature inference, especially when data is sparse.
The authors also implemented "independent-feature" versions of these models (ILDA, ICTM, IMMCTM) for comparison.
Datasets
The study utilized two primary datasets:
- Breast Cancer Data: 560 whole-genome sequenced breast cancer samples from a previously published study.
- Ovarian Cancer Data: 195 whole-genome sequenced ovarian cancer samples, a combination of data from a previous study and the International Cancer Genome Consortium (ICGC) ovarian cancer dataset.
The authors also generated synthetic datasets based on known signatures from the breast cancer data to rigorously benchmark their methods under controlled conditions, including varying levels of mutation counts. This synthetic dataset approach is a valuable contribution for future method development and validation in this field.
Key Insights and Ideas
- Integration is Key: Analyzing SNVs and SVs together, rather than in isolation, provides a more complete picture of the mutational processes affecting a cancer genome.
- Correlations Matter: The relationships (correlations) between different mutation signatures are not random; they often reflect shared underlying biological mechanisms, especially in DNA repair-deficient cancers. Capturing these correlations significantly boosts the accuracy of signature discovery.
- Robustness to Sparsity: MMCTM is particularly effective in datasets with low mutation counts (e.g., for SVs), where traditional methods struggle. This is crucial as whole-genome sequencing data can be noisy and incomplete.
- Uncovering Subtypes: By integrating SNV and SV signatures, the MMCTM can identify novel patient subgroups with distinct molecular profiles and potentially different clinical outcomes, offering a more refined stratification than using either mutation type alone.
- Biological Relevance: The identified signatures and their correlations often align with known biological processes (e.g., APOBEC activity, DNA repair deficiencies like HRD), validating the biological interpretability of the model's output.
Key Results
- Improved Signature Inference: MMCTM consistently outperformed standard methods like Non-negative Matrix Factorization (NMF) and basic LDA in inferring both SNV and SV signatures, especially when dealing with sparse data.
- Enhanced Patient Stratification: In breast and ovarian cancer datasets, MMCTM successfully segregated patients into distinct clusters based on their combined SNV and SV signature profiles. These clusters often corresponded to known subtypes (e.g., BRCA1/2-like HRD) and, in some cases, revealed novel prognostic stratifications.
- Robustness to Data Sparsity: Performance degradation was significantly less pronounced for MMCTM when mutation counts were reduced, highlighting its advantage in real-world, often sparse, genomic datasets. For example, MMCTM showed better performance than other methods when only 1% of SNV counts were retained.
- Reproducible Signatures: The inferred signatures were reproducible across different datasets and aligned well with previously established signatures from public databases like COSMIC.
Limitations and Future Work
- Interpretation of Probabilities: The topic models output "signature probabilities" rather than direct "activity estimates." While related, the precise interpretation and optimal thresholding for downstream analysis of these probabilities require further investigation.
- Model Assumptions: The effectiveness of MMCTM relies on the existence of correlations between signatures. If no such correlations exist, the benefits of this specific model might be limited, and simpler methods like NMF or LDA could suffice.
- Extensibility: While the current MMCTM integrates SNVs and SVs, the authors envision its extension to incorporate other genomic features like copy number variations, double-strand breaks, and telomeric insertions, creating a more comprehensive multi-modal framework.
- Cancer Type Generalizability: While applied to breast and ovarian cancers, further validation across a broader range of cancer types is needed to fully assess the generalizability of the approach.
- Computational Cost: While outperforming other methods, fitting complex multi-modal models can be computationally intensive, though the authors provide code for reproducibility.
Summary updated