sigLASSO: optimizing cancer mutation signatures jointly with sampling likelihood

Shantao Li, Forrest W. Crawford, Mark GersteinbioRxiv 2018 · DOI

TL;DR

This paper introduces sigLASSO, a novel computational framework for identifying cancer mutation signatures. It uniquely accounts for sampling uncertainty, especially crucial in low-data scenarios like whole-exome sequencing, and leverages L1 regularization for biologically interpretable, sparse results.

What Problem Does This Paper Address?

Cancer development is driven by multiple mutational processes, each leaving a distinct "fingerprint" or signature on the tumor genome. Identifying these active signatures is key to understanding cancer initiation and progression. The challenge lies in decomposing observed mutation counts into a combination of known signatures, a task complicated by:

  • Sampling Variance: Especially with low mutation counts (common in whole-exome sequencing), the observed mutations might not perfectly reflect the underlying mutational process.
  • Large Number of Signatures: There are many potential signatures, and determining which ones are truly active in a given sample is difficult.
  • Biological Interpretability: Solutions need to be biologically meaningful and reflect known mutational mechanisms.
  • Model Complexity: Balancing model complexity to avoid overfitting or underfitting is crucial.

Existing methods often struggle with these issues, particularly in low-data regimes, and may not explicitly model sampling uncertainty or provide easily interpretable results.

Methods

sigLASSO tackles these challenges through a novel optimization framework with four key aspects:

  1. Joint Optimization of Sampling and Signature Fitting: Unlike previous methods that might treat observed mutations as perfect representations, sigLASSO explicitly incorporates a multinomial sampling likelihood into its objective function. This allows it to simultaneously learn the underlying mutation distribution and the signature contributions, directly addressing sampling variance.
  2. L1 Regularization for Sparsity: sigLASSO uses L1 regularization (similar to LASSO) on the signature weights. This encourages sparsity, meaning it tends to assign zero weight to signatures that are not active, leading to more parsimonious and biologically interpretable solutions.
  3. Continuous Priors for Model Complexity: Instead of hard-coding which signatures are active or inactive, sigLASSO allows for continuous priors. These priors can be informed by existing biological knowledge (e.g., from previous studies or patient data) and help fine-tune model complexity, making it more robust.
  4. Ability to Abstain: Due to its awareness of uncertainty, sigLASSO can "abstain" from making assignments in low-confidence contexts, flagging them for human review. This is a crucial feature for robust analysis.

The optimization is handled using an iterative approach (alternating between optimizing signature weights and mutation probabilities) that exploits the biconvex nature of the problem.

Datasets

The paper evaluates sigLASSO on:

  • Simulated Datasets: Generated to mimic real cancer mutation data under various conditions (mutation counts, noise levels, number of signatures, prior information).
  • Real Cancer Datasets:
    • 35 Whole Genome Sequencing (WGS) papillary kidney cancer (pRCC) samples.
  • 182 Whole Exome Sequencing (WES) esophageal carcinoma (ESCA) samples.
  • 8,893 TCGA (The Cancer Genome Atlas) samples across 33 cancer types.

The authors also leverage publicly available COSMIC (Catalogue of Somatic Mutations in Cancer) signatures as a source of prior knowledge.

Key Insights and Ideas

  • Sampling Uncertainty is Critical: The paper strongly emphasizes that ignoring sampling variance, especially with low mutation counts, can lead to inaccurate signature assignments. sigLASSO's joint optimization directly addresses this.
  • Priors Enhance Performance: Incorporating biological priors, even if imperfect, significantly boosts sigLASSO's performance, leading to more accurate and robust signature identification.
  • Abstention is a Feature, Not a Bug: The ability of sigLASSO to abstain from making assignments when uncertainty is high is a key advantage, preventing overconfident and potentially incorrect conclusions.
  • Sparsity is Biologically Relevant: The L1 regularization naturally leads to sparse solutions, which align well with the observation that only a subset of mutational processes are typically active in any given cancer sample.

Key Results

  • Improved Performance: sigLASSO generally outperformed deconstructSigs (a popular existing method) in terms of Mean Squared Error (MSE) and support recovery (precision/recall/accuracy), especially in low mutation count scenarios and noisy data.
  • Speed: sigLASSO was found to be significantly faster than deconstructSigs, particularly with high mutation numbers, due to its efficient optimization strategy.
  • Sparsity and Interpretability: sigLASSO consistently produced sparser solutions than deconstructSigs, which are more biologically interpretable.
  • Robustness: The method demonstrated robustness across various scenarios, including low sampling counts and different noise levels.
  • Prior Impact: sigLASSO with priors showed a 5-10% gain in accuracy compared to deconstructSigs and maintained high accuracy (above 0.8) across simulation settings.

Limitations and Future Work

  • Ground Truth: The authors acknowledge that the "ground truth" of operative mutational signatures in real samples is difficult to obtain, making direct evaluation challenging. They rely on simulations and biological plausibility.
  • Unknown Signatures: While sigLASSO can abstain, the presence of truly unknown signatures might still lead to challenges in assignment. The paper suggests that further research into the biological foundations of signatures and quantifying biological priors will be crucial.
  • Elastic Net: While implemented, elastic net (combining L1 and L2 regularization) did not show significant benefits over L1 alone in their simulations with current COSMIC signatures, likely due to low correlation between them. However, they suggest it might become more beneficial as the number of known signatures grows.
  • Future Directions: The authors propose extending their joint optimization framework to other areas with low-count, discrete data, such as single-cell RNA sequencing, and to mutagenesis modeling. They also highlight the potential for sigLASSO to work seamlessly with other decomposition methods like SomaticSignatures.

Summary updated

Abstract

Multiple mutational processes drive carcinogenesis, leaving characteristic signatures on tumor genomes. Determining the active signatures from the full repertoire of potential ones can help elucidate mechanisms underlying cancer initiation and development. This task in-volves decomposing the counts of cancer mutations, tabulated according to their trinucleotide context, into a linear combination of known mutational signatures. We formulate it as an optimization problem and develop sigLASSO, a software tool, to carry it out efficiently. (An R package implementation is available at github.com/gersteinlab/siglasso ). sigLASSO features four key aspects: (1) It jointly optimizes the likelihood of sampling and signature fitting, by explicitly adding multinomial sampling into the overall objective function. This is particularly important when mutation counts are low and sampling variance is high, such as in exome sequencing. (2) sigLASSO uses L1 regularization to parsimoniously assign signatures to mutation profiles, leading to sparse and more biologically interpretable solutions resembling previously well-characterized results. (3) sigLASSO fine-tunes model complexity, informed by the scale of the data and biological-knowledge based priors. In particular, instead of hard thresholding and choosing a priori a discrete subset of active signatures, sigLASSO allows continuous priors, which can be effectively learned from auxiliary information. (4) Because of this, sigLASSO can assess model uncertainty and abstain from making certain assignments in low-confidence contexts. Finally, to evaluate sigLASSO signature assignments in comparison to other approaches, we develop a set of reasonable expectations (e.g. sparsity, the ability to abstain, and robustness to noise) that we apply consistently in a variety of contexts.
BibTeX
@misc{li2018siglasso,
  author = {Shantao Li and Forrest W. Crawford and Mark Gerstein},
  title = {{sigLASSO: optimizing cancer mutation signatures jointly with sampling likelihood}},
  year = {2018},
  doi = {10.1101/366740},
  howpublished = {bioRxiv},
}
Figure