Running on Expression Matrices

Expression Matrices

Decomposition of expression signatures using signatureanalyzer. For identifying de novo signatures in expression matrices (ex. single-cell RNA-Seq, bulk RNA-Seq, etc.). The following document is a reference of important considerations when running this method for these data-types.

---

Feature Selection

With most clustering methods on transcriptional data, we recommend a highly variable gene selection step. These are well documented for single-cell and bulk RNA-Seq data and may reduce the input feature space to a few thousand genes of interest.

Objective Function

For mutational signatures, we assume a gaussian distribution of normalized counts and use Fevotte & Tan's derivation of a gaussian objective function for ARD-NMF. Thus, it is important to use the gaussian objective function. We recommend the following normalization methods:

• Bulk RNA-seq: log2(TPM+1); normalize TPMs with DESeq2 size factors
• Single-cell RNA-seq: ln(CP10K+1); normalize CP10K w/ scran size factors

Use:

--objective gaussian

---

Type of Run

This specifies whether or not to do Cosmic mapping for your dataset. For expression matrices, this is not relevent.

Use:

-t matrix

---

Prior on H & W

We generally impose an exponential (L1) or half-normal (L2) prior on the W & H matrices for non-negative matrix factorization.

Use:

--prior_on_H L1 --prior_on_W L1

# or

--prior_on_H L2 --prior_on_W L2

---

Running the method

This method may be run using an input of (n samples x m features).

Use:

signatureanalyzer -n 10 \
                  -t matrix \
                  --objective gaussian \
                  --max_iter 30000 \
                  --prior_on_H L1 \
                  --prior_on_W L1 \
                  input.tsv