Single-Sample Tumor Subclonal Reconstruction Algorithm Based on Crowd-Sourced Resources

Background

The evolution of cancer and the genetic heterogeneity of tumors are critical fields in modern oncology research. Tumors evolve from normal cells through progressive acquisition of somatic mutations. These mutations occur probabilistically, influenced by the chromatin structure of the cells and both endogenous and exogenous mutational pressures. If specific mutations provide a selective advantage, the descendant cells can expand in their local environment. Over years of accumulation, this leads to populations of cells with various cancer hallmarks, referred to as clones. Different subpopulations of tumor cells (subclones) can emerge through drift or selective pressures within the whole cell population. This evolutionary characteristic is clinically significant, as genetic heterogeneity is associated with worse prognosis, more mutations, and drug resistance. Therefore, understanding and quantifying the evolutionary process of tumors is crucial for cancer treatment and prognosis evaluation.

Tumor subclonal reconstruction is a common method used to quantify the evolutionary characteristics of tumors by utilizing somatic single-nucleotide variants (SNVs) and copy number aberrations (CNAs) allele frequencies. Many algorithms have been developed for this task, employing various strategies like Bayesian inference. However, there are significant differences in accuracy and applicability among these algorithms, and it is unclear how best to quantify their accuracy. Thus, evaluating existing subclonal reconstruction algorithms and determining the factors affecting their accuracy is of paramount importance.

Research Source and Publication Information

The paper titled “Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction” by Adriana Salcedo et al., was published in the journal Nature Biotechnology. This study was a collaborative effort from multiple institutions, including the University of California, Los Angeles, Free University of Brussels, Ontario Institute for Cancer Research, among others. The study is based on the ICGC-TCGA (International Cancer Genome Consortium–The Cancer Genome Atlas) DREAM Somatic Mutation Calling Challenge (SMC-Het Challenge), conducted over 7 years, using a cloud computing platform to evaluate the performance of 31 subclonal reconstruction algorithms in 51 simulated tumors.

Research Process and Methods

Study Design

To evaluate tumor subclonal reconstruction algorithms, the research team designed 51 tumor lineages based on the ICGC-TCGA DREAM Somatic Mutation Calling Challenge and its tumor simulation framework. These tumors encompass a wide range of biological and technical parameters. Among these, 25 are based on manually curated whole-genome cancer analysis (PCAWG) data, while the remaining 26 are based on non-PCAWG tumors and specific marginal cases from a single breast tumor. The research team utilized Bamsurgeon to simulate normal and tumor BAM files, and used the GATK Mutect tool to identify somatic SNVs, with Battenberg for identifying somatic CNAs and estimating tumor purity.

Algorithm Evaluation

Participating teams submitted 31 containerized workflows, all executed in a reproducible cloud architecture. Researchers added five reference algorithms, including random prediction, PCAWG’s “informed brute-force” clustering algorithm, a single-cluster assignment algorithm, and two state-of-the-art algorithms (DPClust and PhyloWGS). Each method was evaluated in seven sub-challenges: purity (SC1a), the number of subclones (SC1b), SNV cellular prevalence (SC1c), mutation clustering (SC2), and phylogeny (SC3), with SC2 and SC3 further divided into deterministic (SC2a and SC3a) and probabilistic tasks (SC2b and SC3b). Each prediction used an established framework for scoring, with scores normalized within {tumor, sub-challenge} pairs.

Key Research Findings

Algorithm Performance Evaluation

The study found significant differences in the performance of various algorithms across all seven tasks, indicating that algorithm choice has a substantial impact on performance, surpassing the influence of tumor characteristics. Specifically, no single algorithm performed best in all tasks, and existing ensemble strategies did not surpass the best single methods, suggesting a continuing need for research in subclonal reconstruction algorithms.

Best Algorithms

The research team ranked algorithms based on the median score across all tumors, identifying a best performance submission in SC1a and SC2b, while SC1b and SC1c each had two statistically indistinguishable submissions, and SC2a had three statistically undifferentiated submissions. The top algorithm for SC1a inferred purity using only copy number calls, while the next best method combined copy number and SNV clustering purity estimates.

Influencing Factors Analysis

By analyzing tumor and algorithm features, the study identified a few tumor characteristics that strongly impacted reconstruction accuracy. Sensitivity to specific tumor traits explained differences in variation detection and data resolution among algorithms. Notably, tumor purity, copy number status, and mutation burden significantly affected many algorithms’ performance, while algorithms based on Gaussian noise models showed poor performance in SNV co-clustering tasks.

Impact of Intrinsic Data Characteristics and Experimental Design on Accuracy

The study highlighted sequencing coverage as the main controllable technical feature in experimental design. By adjusting sequencing coverage to consider tumor purity and ploidy, the study quantified the impact of NRPCC (Reads Per Chromosome Copy) on subclonal reconstruction. Results indicated that higher NRPCC led to better algorithm performance in SNV co-clustering and tumor purity estimation. However, at high NRPCC levels, differences between algorithms became the primary source of variation.

Error Source Analysis

The study also explored sources of error in SNV cellular prevalence estimation. Most algorithms accurately identified whether an SNV was a clonal mutation but performed poorly in detecting low-frequency subclonal mutations. Furthermore, underlying copy number status significantly affected the accuracy of SNV clonality assignment, particularly for clonal SNVs in subclonal copy number loss regions. Algorithms based on Gaussian noise models performed poorly in handling low-frequency variants, while robustness to copy number changes significantly correlated with overall performance.

Conclusion

This study systematically evaluated the performance of 31 single-sample tumor subclonal reconstruction algorithms in 51 simulated tumors, revealing significant impacts of algorithm choice and experimental design on reconstruction accuracy. The findings will help improve the application of existing methods and the development of new methods to better understand the tumor evolution process. Moreover, the research team provided online tools to help users choose the best algorithm based on data set and research questions.

Research Significance

This study provided a benchmark for evaluating tumor subclonal reconstruction algorithms, promoting comparison and improvement among different algorithms. By identifying key factors affecting algorithm performance, the research offers important guidance for future algorithm development and optimization. Additionally, the study emphasized the importance of high-quality sequencing data and appropriate experimental design in tumor evolution research, providing valuable references for clinical cancer treatment and prognosis evaluation.