The baseline methods we propose here are wrappers of already published unsupervised deconvolution algorithms (ICA-based or NMF-based). We assume here that \(A\), \(T_{MET}\) and \(T_{RNA}\) are unknown and need to be estimated, either independently (single-omic pipelines) or integratively (double-omic pipelines).
Before deconvolution, we systematically apply a pre-treatment step of dimensionality reduction based on feature selection. All baseline methods source code is downloadable on the DECONbench platform.
All baselines relies on unsupervised deconvolution algorithms, which consists in solving D=TA, either by ICA-based (i) or NMF-based (ii) approaches.
This following table gives an overview of all the methods implemented as baseline is Deconbench, in particular their characteristics in terms of:
- omics data they work on
- the feature selection they use (FS)
- the deconvolution algorithm they are based on
- their computational time.
A more detailed description of each method is given below the table.
| RNA_wICA | RNA_wNMF | DNAm_Edec | DNAm_MeDeCom | DNAm_wICA | both_wICA | both_wNMFMeDeCom | both_meanwNMFMeDeCom | |
|---|---|---|---|---|---|---|---|---|
| Data type | RNA | RNA | DNAm | DNAm | DNAm | both | both | both |
| FS DNAm | / | / | 5K most variable probes | 5K most variable probes | / | / | 5K most variable probes | 5K most variable probes |
| FS RNA | ICA - selection of top-contributing genes and filtering of duplicated genes | ICA - selection of top-contributing genes | / | / | / | ICA - selection of top-contributing genes | ICA - selection of top-contributing genes | |
| Deconvolution DNAm | / | / | Edec | MeDeCom | ICA weighted on 30 most important genes | ICA weighted on 30 most important genes | MeDeCom with the A matrix computed on RNA as startA parameter | MeDeCom |
| Deconvolution RNA | ICA weighted on 30 most important genes | NMF with snmf/r method | / | / | ICA weighted on 30 most important genes | NMF with snmf/r method | NMF with snmf/r method | |
| Time 10 A | ~10min | ~20min | ~3h | ~17h | ~10min | ~10min | ~17h | ~17h30 |
| Time 1 A | ~1min | ~2min | ~20min | ~1h40 | ~1min | ~1min | ~1h40 | ~1h45 |
| Reference of the tool | ref [1-3] | ref [4] | ref [5] | ref [6] | ref [1-3] | ref [1-3] | ref [4][6] | ref [4][6] |
Time 1 A is the time to run the method one time, Time 10 A is the time to run the method on all the benchmark datasets of DeconBench.
Eight methods have been included in the benchmark as baseline to be compared with the accuracy of new methods.
The method RNA_wICA (r_WIC) uses transcriptomic data as input and is based on the ICA algorithm for both feature selection and deconvolution. It relies on the use of the functions “runICA” and “getGenesICA” developed by P. Nazarov (sablab.net/scripts/LibICA.r) and the deconica R package [3].
STEP1: feature selection. For the ICA-based feature selection, the function “runICA” is run at first with the parameters $ncomp = 10$ and ntry = 50. Then, the function “getGenesICA” selects top-contributing genes with a FDR of 0.2, the feature selection is done on these contributing genes belonging to a component having an average stability greater than 0.8. Finally, duplicated genes are removed.
STEP2: deconvolution. First, we perform FastICA unsupervised deconvolution (deconica::run_fastica is run with the parameters overdecompose = FALSE and n.comp = 5; remaining parameters are set to default). Second, we compute the abundance of the identified cell type following deconica pipeline, assuming that the weighted-mean of the \(30_{top}\) genes of a component is a unique proxy of each component signal.The 30 most important genes of each ICA component are extracted by the function deconica::generate_markers with the parameter return = "gene.ranked“. These genes are used to weight the component scores with the function deconica::get_scores, with the log counts of the ICA as “df” parameter, the list of 30 genes as “markers.list” parameter, and the parameter summary = “weighted.mean”. Finally, the proportions are extracted with the function deconica::stacked_proportions_plot on the transpose of the deconica::get_scores output.
The method DNAm_wICA (m_WIC) uses DNA methylation data as input.
STEP1: feature selection. It has no feature selection step.
STEP2: deconvolution. The deconvolution step is based on ICA, similarly to what was described for the second step of RNA_wICA, but applied on the DNA methylation matrix.
The method both_wICA (b_WIC) combines transcriptomics and DNA methylation information.
STEP1: feature selection. It has no feature selection step.
STEP2: deconvolution. The deconvolution is in two steps, one on each data type. The transcriptomics and DNA methylation are separately deconvoluted with the same deconvolution step as in r_WIC and m_WIC respectively to estimate AMET and ARNA
STEP3: integration. Finally, the mean of both AMET and ARNA estimated proportion matrices is computed as the final method output. To compute the average, the cell types of the both deconvolution matrices are matched by iteration. The cell types of the methylation result matrix are reordered 1,000 times, and the one that best correlates with the transcriptomic result matrix is kept.
The method RNA_wNMF (r_WNM), is a two step-approach that uses transcriptomic data as input.
STEP1: feature selection. The first step uses ICA to perform a feature selection as described for RNA_wICA, although duplicated genes are kept. This step therefore allows genes that contribute to several components to be present several times in the data.
STEP2: deconvolution. The deconvolution is based on sparse NMF and least-squares optimization to minimise \(||D-TA||_{2}\) [4]. It is called by the NMF::nmf function, with the parameter method = "snmf/r".
STEP1: feature selection. The method DNAm_EDec (m_EDC), uses DNA methylation data as input and follows the pipeline implemented in the R package medepir7. The feature selection is performed by medepir::feature_selection [7] for keeping highly variable probes (5000 most variable probes).
STEP2: deconvolution. The NMF-based algorithm of the method EDec9 is used for the deconvolution part, with the function medepir::Edec [5] and all the selected probes as “infloci” parameter. The algorithm consist in minimizing the error term \(||D-TA||_{2}\) with constraints on methylation values.
STEP1: feature selection. The method DNAm_MeDeCom (m_MDC), uses DNA methylation data as input and is based on the pipeline of the R package medepir [7]. The feature selection is performed as for DNAm_EDec above to select the 5000 most variable probes.
STEP2: deconvolution. The deconvolution step, however, uses the MeDeCom R package8. It is run with the function MeDeCom::runMeDeCom [6], with the lambda parameter set to 0.01. As EDec implementation of NMF algorithm, MeDeCom algorithm consists in minimizing the error term \(||D-TA||_{2}\) with constraints on methylation values.
The method both_wNMFMeDeCom (b_COM) combines transcriptomics and DNA methylation information. It is the combination of the two methods RNA_wNMF and DNAm_MeDeCom. The method r_WNM is first applied on the RNAseq matrix.
STEP1: feature selection. The DNA methylation matrix is pre-treated as described in the m_MDC method, with the selection of 5000 most variable.
STEP2-3: deconvolution-integration. Finally, the MeDeCom algorithm is run on the DNAm data, with the result of r_WNM as the initialization parameter startA.
The method both_meanwNMFMeDeCom (b_MEA), which integrates transcriptomics and DNA methylation, applies r_WNM to the transcriptomics matrix, m_MDC to the DNA methylation matrix.
STEP1: feature selection. Feature selection is performed on DMET and DRNA matrices as described in r_WNM and m_MDC sections.
STEP2: deconvolution. Deconvolution is performed on DMET and DRNA matrices as described in r_WNM and m_MDC sections to estimate AMET and ARNA matrices.
STEP3: integration. We computed the mean of the two estimated AMET and ARNA matrices, similarly to b_WIC.
[1] fastICA: FastICA Algorithms to Perform ICA and Projection Pursuit. https://CRAN.R-project.org/package=fastICA.
[2] Hyvarinen, A. Fast and robust fixed-point algorithms for independent component analysis. IEEE Trans. Neural Netw. 10, 626–634 (1999).
[3] Czerwinska, U. UrszulaCzerwinska/DeconICA: DeconICA first release. (Zenodo, 2018). doi:10.5281/zenodo.1250070.
[4] Frichot, E., Mathieu, F., Trouillon, T., Bouchard, G. & François, O. Fast and Efficient Estimation of Individual Ancestry Coefficients. Genetics 196, 973–983 (2014).
[5] Onuchic, V. et al. Epigenomic Deconvolution of Breast Tumors Reveals Metabolic Coupling between Constituent Cell Types. Cell Rep. 17, 2075–2086 (2016).
[6] Lutsik, P. et al. MeDeCom: discovery and quantification of latent components of heterogeneous methylomes. Genome Biol. 18, 1–20 (2017).
[7] Decamps, C., Privé, F., Bacher, R. et al. Guidelines for cell-type heterogeneity quantification based on a comparative analysis of reference-free DNA methylation deconvolution software. BMC Bioinformatics. 2020;21, 16.