Using sigLASSO to optimize cancer mutation signatures jointly with sampling likelihood

Shantao Li, Forrest W. Crawford, Mark GersteinNature Communications 2020 · DOI

Abstract

Multiple mutational processes drive carcinogenesis, leaving characteristic signatures in tumor genomes. Determining the active signatures from a full repertoire of potential ones helps elucidate mechanisms of cancer development. This involves optimally decomposing the counts of cancer mutations, tabulated according to their trinucleotide context, into a linear combination of known signatures. Here, we develop sigLASSO (a software tool at github.com/gersteinlab/siglasso) to carry out this optimization efficiently. sigLASSO has four key aspects: (1) It jointly optimizes the likelihood of sampling and signature fitting, by explicitly factoring multinomial sampling into the objective function. This is particularly important when mutation counts are low and sampling variance is high (e.g., in exome sequencing). (2) sigLASSO uses L1 regularization to parsimoniously assign signatures, leading to sparse and interpretable solutions. (3) It fine-tunes model complexity, informed by data scale and biological priors. (4) Consequently, sigLASSO can assess model uncertainty and abstain from making assignments in low-confidence contexts.

Introduction

Cancer is driven by mutational processes—such as UV radiation, tobacco smoke, or DNA repair defects—that leave distinct "fingerprints" or mutational signatures on the genome. These are typically represented as a distribution across 96 trinucleotide contexts. Identifying these signatures in a patient's tumor is critical for understanding the underlying biological drivers of the disease.

The challenge lies in "signature attribution": taking a patient's mutation counts and decomposing them into a linear combination of known signatures. Existing tools often struggle with two major issues. First, in Whole Exome Sequencing (WES), mutation counts are low, leading to high sampling variance. Most tools treat these counts as a fixed distribution, ignoring the inherent uncertainty of small sample sizes. Second, many methods assign too many signatures to a single sample, leading to "dense" solutions that are biologically implausible and difficult to interpret.

Methodology

The authors introduce sigLASSO, a new computational tool that improves the identification of mutational signatures by jointly optimizing for sampling uncertainty and signature fitting. Its architecture rests on four pillars:

  • Joint Likelihood Optimization: Unlike standard regression, sigLASSO explicitly factors a multinomial sampling likelihood into its objective function. It simultaneously estimates the "true" underlying mutation distribution and the signature weights. In low-count scenarios, the model relies more on the signature fit; as counts increase, it trusts the observed data more.
  • L1 Regularization (LASSO): To prevent overfitting, the model applies an L1 penalty to the signature weights. This forces the coefficients of irrelevant signatures to zero, ensuring the solution is "sparse" and biologically realistic.
  • Biological Priors: sigLASSO allows researchers to incorporate prior knowledge. For example, if a sample is from a skin cancer patient, the penalty for the UV signature can be lowered, making it easier for the model to select it.
  • Biconvex Optimization: The algorithm iteratively solves two convex problems: a weight-step and a probability-refinement step. This approach is computationally efficient and converges quickly.
Schematic of sigLASSO's joint optimization framework and performance across mutation counts.

The authors validated sigLASSO across several contexts, including simulated data, Whole Genome Sequencing (WGS) of Papillary Kidney Cancer, WES of Esophageal Carcinoma, and a massive TCGA Pan-Cancer evaluation involving 8,893 samples across 33 cancer types.

Results and Discussion

One of the most novel features is sigLASSO’s ability to "abstain" from making an assignment if uncertainty is too high, signaling that the data is insufficient or an unknown signature might be present. Furthermore, by jointly fitting signatures and sampling parameters, the model allows "knowledge transfer," where the constraints of known signatures help refine the estimate of the noisy multinomial distribution.

Superior Performance in Low Counts

In simulations with low mutation numbers, sigLASSO achieved significantly lower Mean Squared Error (MSE) than point Maximum Likelihood Estimation (MLE). By accounting for the noise in low-count data, sigLASSO provides a more accurate representation of the underlying mutational processes.

Comparison of Mean Squared Error (MSE) between sigLASSO and point Maximum Likelihood Estimation (MLE).

Sparsity and Accuracy

The authors demonstrate that biologically, most tumors are driven by only a few signatures. By enforcing sparsity through LASSO, sigLASSO aligns better with biological reality than "greedy" search methods. On the TCGA dataset, sigLASSO assigned a median of only 1 signature per sample, whereas the popular tool deconstructSigs assigned a median of 4. This suggests sigLASSO is much less prone to "false positive" signature assignments. Additionally, when using biological priors, sigLASSO showed a 5–10% gain in accuracy over deconstructSigs.

Conclusion

sigLASSO offers a robust, efficient, and interpretable solution for mutational signature attribution. While it currently relies on predefined libraries like COSMIC, it sets a new standard for handling low-mutation-count data through joint optimization. Future work may involve automating the generation of biological priors from clinical metadata and exploring the utility of Elastic Net regularization as signature catalogs grow.

Summary updated

Abstract

Multiple mutational processes drive carcinogenesis, leaving characteristic signatures in tumor genomes. Determining the active signatures from a full repertoire of potential ones helps elucidate mechanisms of cancer development. This involves optimally decomposing the counts of cancer mutations, tabulated according to their trinucleotide context, into a linear combination of known signatures. Here, we develop sigLASSO (a software tool at github.com/gersteinlab/siglasso) to carry out this optimization efficiently. sigLASSO has four key aspects: (1) It jointly optimizes the likelihood of sampling and signature fitting, by explicitly factoring multinomial sampling into the objective function. This is particularly important when mutation counts are low and sampling variance is high (e.g., in exome sequencing). (2) sigLASSO uses L1 regularization to parsimoniously assign signatures, leading to sparse and interpretable solutions. (3) It fine-tunes model complexity, informed by data scale and biological priors. (4) Consequently, sigLASSO can assess model uncertainty and abstain from making assignments in low-confidence contexts.
BibTeX
@inproceedings{li2020using,
  author = {Shantao Li and Forrest W. Crawford and Mark Gerstein},
  title = {{Using sigLASSO to optimize cancer mutation signatures jointly with sampling likelihood}},
  booktitle = {Nature Communications},
  year = {2020},
  doi = {10.1038/s41467-020-17388-x},
}
Figure