TL;DR
This paper introduces C-ToMExO, a novel computational method that infers cancer progression dynamics by analyzing the clonal composition of tumors, rather than just tumor-level mutation data. This approach significantly improves the accuracy and robustness of identifying interactions between cancer driver genes, crucial for understanding tumor evolution.
What Problem Does This Paper Address?
Cancer is driven by the accumulation of mutations in specific "driver genes." Understanding how these mutations interact and occur in sequence (cancer progression dynamics) is key to developing effective treatments. Traditional methods, called Cancer Progression Models (CPMs), analyze mutation patterns across many tumors. However, these methods often treat each tumor as a single, homogeneous entity, ignoring the fact that tumors are frequently composed of multiple distinct cell populations (clones), each with its own unique set of mutations. This "clonal heterogeneity" can lead to misleading interpretations when analyzing tumor-level data, potentially masking important relationships between driver genes or suggesting false ones. The paper aims to overcome this limitation by leveraging finer-grained, clone-level data.
Methods
The core of the paper is the C-ToMExO (Clonal-ToMExO) method. It builds upon a probabilistic model for cancer progression originally developed for tumor-level data (ToMExO). The key innovation is the incorporation of clone-level genotype representations. Instead of a single mutation profile per tumor, C-ToMExO uses a representation that includes the genotypes of individual clones within a tumor and their relative population sizes.
To handle this richer data, C-ToMExO employs a guided Markov Chain Monte Carlo (MCMC) inference algorithm. This algorithm is designed to efficiently explore the vast space of possible cancer progression models. It uses pre-computed mutual exclusivity (ME) and mutual inclusivity (MI) matrices derived from the clone-level data to guide the search towards more probable models. This "informed proposal distribution" significantly speeds up the inference process compared to purely random exploration.
A crucial preprocessing step is also introduced: CNA (Copy Number Alteration) correction. This method accounts for how copy number changes in the genome can affect the observed variant allele frequencies (VAFs) of mutations, allowing for more accurate reconstruction of tumor clonal trees, especially in the presence of complex evolutionary events.
Datasets
The paper evaluates its method on:
- Synthetic datasets: 1200 datasets generated from 100 random generative progression models, varying in the number of genes, tumors, and error rates. This allows for a controlled assessment of the algorithm's performance.
- Biological datasets: Two subsets of the TRACERx Lung Cancer dataset:
- LUAD (Invasive Adenocarcinoma): 61 tumors.
- LUSC (Squamous Cell Carcinoma): 32 tumors.
Key Insights and Ideas
- Clonal heterogeneity is critical: Ignoring the distinct cell populations within a tumor can lead to significant misinterpretations of driver gene interactions, particularly for mutual exclusivity.
- Clone-level data unlocks deeper insights: By analyzing mutations at the clone level, C-ToMExO can more reliably identify true relationships between driver genes, even in the presence of complex evolutionary histories.
- Guided MCMC is essential for scalability: The sheer complexity of cancer progression models necessitates efficient inference. The guided MCMC approach, informed by pairwise gene relationships, dramatically improves the speed and effectiveness of finding the most likely models.
- CNA correction is vital for accurate tumor tree reconstruction: Copy number alterations can obscure mutation signals. The proposed correction method enables more robust tumor tree building, which is a prerequisite for accurate progression modeling.
Key Results
- Improved accuracy: Synthetic data experiments show that C-ToMExO can accurately recover generative progression models, with performance improving with more tumors and lower error rates. Weighted F-scores indicate strong agreement with generative models for identifying relationships among highly mutated genes, even with relatively small datasets (e.g., 30 tumors).
- Computational efficiency: The guided MCMC algorithm significantly speeds up inference compared to previous methods.
- Biological relevance: Analysis of TRACERx lung cancer data identified known and potentially novel relationships between driver genes, such as the mutual exclusivity between EGFR and KRAS in LUAD, which has clinical implications. The method successfully identified these relationships with high statistical significance.
Limitations and Future Work
- Model complexity: While C-ToMExO is more scalable than some previous methods, the space of possible progression models still grows exponentially with the number of genes, posing a challenge for very large numbers of driver genes.
- Data requirements: The method benefits from multi-regional sequencing data to accurately reconstruct clonal trees.
- Future work: The authors suggest further exploration of the inferred relationships, particularly those involving TP53 in LUAD, for deeper biological investigation. They also anticipate applying the method to larger datasets as more data becomes available from cohorts like TRACERx. The paper also mentions the need for robust methods to handle the increasing complexity of tumor evolution.
Summary updated