We have developed SigNet, a tool to decompose mutational profiles into mutational signatures (COSMIC v3.1).

SigNet leverages artificial neural networks to decompose (or "refit") sample mutational profiles even when the total mutation count in the sample is very low. It outperforms methods based on non-negative least squares for signature decomposition and provides error bars for the weight estimations. In addition to the signature fitting module, SigNet has a "Detector" and a "Generator" module. The former flags samples which deviate from the expected patterns and which may thus harbour biologically interesting information and the latter is a tool to generate quasi-realistic-looking cancer mutation data based on a variational autoencoder (VAE). More details on the method and some of its applications can be found in the preprint Serrano Colome et al. (2023).

SigNet is available for download on github, PyPI and can also easily be run using the singularity image. Last update: 05/2024.

CBaSE

Cancer Bayesian Selection Estimation (CBaSE) is a tool to identify genes under positive or negative selection in tumor genomes.

Based on somatic point mutations, CBaSE first estimates the mutation probability at the gene level without the necessity to know external mutation rate covariates. Unlike other methods, the algorithm does not suppose that the expected value of the number of mutations per gene is gamma-distributed, but instead allows for a wider range of possible model distributions. CBaSE selects the best model and outputs both q-values as well as the magnitude of the selection signal (dN/dS), which assess the strength of negative and positive selection for each gene. Extended sequence context can be taken into account up to pentamers. The method is described in Weghorn & Sunyaev (2017).

CBaSE v1.2 is available as a python script for download, with corresponding documentation here. Last update: 07/2024.

MutPanning

We co-developed MutPanning, a tool to identify genes under positive selection in tumor genomes.

MutPanning accounts for extended context dependence beyond pentamers, which plays an important role for some cancer types. The algorithm is composed of two tests for selection: The first test assesses whether a gene has a signficant excess of nonsynonymous mutations compared to the expectation under neutral evolution. The second test queries whether nonsynonymous mutations on the gene are distributed according to the estimated neutral pattern. This latter feature boosts power to predict cancer driver genes when only a fraction of nonsynonymous sites are under selection. The method is described in Dietlein* & Weghorn* et al. (2020).

MutPanning is available for download as a standalone script and as a module on the GenePattern server. All information can be found at cancer-genes.org.

Weghorn Lab