Quantifying the seed sensitivity of cancer subclonal reconstruction algorithms

Philippa L Steinberg ^1,2,3, Lydia Y Liu ^1,2,3,4,5, Anna Neiman-Golden ^1,2,3

⁰Department of Human Genetics, University of California, Los Angeles, Los Angeles, CA, 90095, USA.

Biorxiv : the Preprint Server for Biology +

|

February 19, 2024

View abstract on PubMed

Summary

The initializing seed significantly impacts subclonal reconstruction (SRC) results, affecting cancer subclone estimates. Reporting and randomizing seeds are crucial for reproducible bioinformatics research.

Area Of Science

Genomics
Bioinformatics
Computational Biology

Background

Intra-tumoural heterogeneity (ITH) is a key challenge in cancer prognosis and treatment.
Subclonal reconstruction (SRC) algorithms estimate ITH by identifying cancer subclones from bulk DNA sequencing data.
Probabilistic SRC algorithms require initialization with a random seed, the impact of which is largely unstudied.

Approach

Benchmarked the seed sensitivity of three probabilistic SRC algorithms (PyClone-VI, DPClust, PhyloWGS).
Utilized fourteen whole-genome sequences from head and neck squamous cell carcinoma.
Evaluated nine SRC pipelines across 1470 subclonal reconstructions (single- and multi-region).

Key Points

All evaluated SRC algorithms demonstrated substantial seed sensitivity, with varying subclone estimates for identical input data.
Subclone estimates differed across SRC pipelines, but seed variability was consistent within each algorithm.
No single seed consistently identified the most frequent number of subclones across all patients for any algorithm.

Conclusions

Seed sensitivity introduces significant variability in quantifying ITH using probabilistic SRC algorithms.
Recommends reporting and randomizing seed choices in publications to enhance reproducibility.
Suggests considering seed sensitivity in future SRC algorithm benchmarking and bioinformatics tool development.