--- title: "The lute user's guide" author: - Sean K. Maden - Stephanie C. Hicks date: "`r format(Sys.time(), '%d %B, %Y')`" bibliography: bibliography.bib package: lute output: BiocStyle::html_document: code_folding: show toc: no tocfloat: no BiocStyle::pdf_document: toc: no toc_depth: 0 vignette: > %\VignetteIndexEntry{The lute user's guide} %\usepackage[UTF-8]{inputenc} %\VignetteEncoding{UTF-8} %\VignetteEngine{knitr::rmarkdown} --- ```{r setup, include=FALSE} libv <- c("lute", "SummarizedExperiment", "SingleCellExperiment") sapply(libv, library, character.only=TRUE) knitr::opts_chunk$set(echo=TRUE, eval=TRUE) ``` # How to use this guide This User's Guide describes how to use `lute`, a framework for bulk transcriptomics deconvolution. An introduction to deconvolution experiments is provided, followed by a description of the framework and a small experiment using example data. Begin by passing the following to your active R session: ```{r} library(lute) ``` # Installation Install `lute` by passing the following to the console from an active R session. You must have the `BiocManager` package installed. ```{r, eval = F} BiocManager::install("lute") ``` # Deconvolution overview ## Definition Deconvolution is the task of quantifying signal from a signal mixture, which is sometimes called a signal convolution. Deconvolution experiments try to predict signals from their mixtures as accurately and reliably as possible. ## Transcriptomics deconvolution In the field of transcriptomics, deconvolution is often applied to gene expression datasets in order to predict cell type quantities from cell mixtures. This technique is applied to bulk tissue specimens of multiple cell types, where accurate cell type quantifications can improve bias adjustments or allow testing of new hypotheses. ## Experiment elements Most deconvolution algorithms predict the type-specific proportions from a signature matrix and a convoluted signals matrix in the following manner: $$P \leftarrow [Y, Z]$$ Where we have the following term definitions for bulk transcriptomics deconvolution: * $P$ : $K$-length vector of predicted proportions for the number of cell types $K$. * $Z$ : Signature expression matrix, with dimensions of $G$ total rows of signature genes by $K$ total columns of cell types. * $Y$ : The convoluted signals expression matrix, with dimensions of $G$ total rows of signature genes by $J$ total columns of bulk samples. The above terms are shared by reference-based deconvolution algorithms, which are defined by their requirement of a signature matrix $Z$ to make predictions. Some additional important properties to consider are the preparation steps used to generate the terms $Z$ and $Y$. These steps might include data rescaling and transformation of the expression signals, or certain selection criteria to arrive at the final set of $G$ signature genes. ## Pseudobulking Another important representation of deconvolution is in terms of a pseudobulk equation. A pseudobulk is a simulated bulk sample generated using either cell- or type-level reference data. Pseudobulk analysis is a common task for deconvolution experiments. We may represent the pseudobulked sample $Y_{PB}$ in terms of the $Z$ signature matrix and $P$ cell type proportions, defined as above, as well as some $K$-length vector $S$, which contains cell size estimates used for rescaling $Z$. This looks like the following: $$Y_{PB}=Z * P * S$$ ## Biological importance of cell size scale factors Inclusion of the $S$ vector of cell size scaling factors is an important strategy to correct for potential bias due to systematic differences in sizes between the sizes of predicted cell types. It has been used in deconvolution studies of a variety of tissues, including brain and blood (@monaco_rna-seq_2019, @racle_epic_2020, @sosina_strategies_2021). Proportions predictions not incorporating cell size scale factors assume equal cell sizes between the predicted types, leading to very biased predictions in tissues such as brain that feature principal cell types with very different sizes. ## Deconvolution algorithms There are dozens of different deconvolution algorithms used in the field of trascriptomics alone. These may be organized in several ways, such as whether they incorporate a variance weighting strategy or cell size scale factors. Another useful way to organize algorithms is by whether they incorporate the non-negative least squares (NNLS) algorithm in some way. This is a statistical approach with the added constraint of assumed non-negativity for inputs, which holds when we consider typical gene expression datasets in the form of counts or log-normalized counts. The next section shows a hierarchical class structure for accessing multiple deconvolution algorithms, and we can think of this hierarchy as yet another way of organizing and relating different deconvolution algorithms in a useful and actionable manner. # The `lute` framework for deconvolution `lute` supports deconvolution experiments by coupling convenient management of standard experiment tasks with standard mappings of common inputs to their synonyms in supported algorithms for marker selection and deconvolution. Accessing the framework is as simple as running the `lute()` function using your experiment data (see `?lute` for details). Runnable examples using the framework function are provided below. ## Supported deconvolution algorithms To view a table of supported algorithms and their details, pass the following code to your R session: ```{r} info.table <- luteSupportedDeconvolutionAlgorithms()[,seq(5)] knitr::kable(info.table) ``` ## The `deconvolutionParam` class hierarchy Briefly, `lute` defines a new class hierarchy that organizes deconvolution functions. This was based on the approach uesd by the `bluster` R/Bioconductor package for clustering algorithms. Classes in this hierarchy each have a method defined for the `deconvolution()` generic function (see `?deconvolution` for details). Methods for higher-level parent classes like `referencebasedParam` will perform data summary and marker genes comparisons. For the algorithm-specific classes, such as `nnlsParam` or `musicParam`, the method maps standard inputs like $Y$, $Z$, and $S$ to their algorithm-specific synonyms. The `deconvolutionParam` class hierarchy for supported algorithms is visualized in the following diagram: The `deconvolutionParam` class hierarchy is intended to be extensible to new algorithms. A future vignette will provide a step-by-step guide to supporting a new algorithms using these classes and their methods. # Deconvolution experiment example This section features several runnable examples of deconvolution experiments using the `lute()` framework function. We may begin with an object of type `SingleCellExperiment` (a.k.a. "sce" object) containing cell-level expression data. We may use the main `lute()` framework function to perform a deconvolution experiment with these data. For demonstration, we generate a randomized sce object with `randomSingleCellExperiment()` (see `?randomSingleCellExperiment` for details). Set the $G$ marker genes to be 10, or 5 per cell type, which we will identify from an initial set of 100 genes: ```{r} markersPerType <- 5 totalGenes <- 100 exampleSingleCellExperiment <- randomSingleCellExperiment( numberGenes=totalGenes) ``` We could identify marker genes in `exampleSingleCellExperiment` now by running `lute` without setting the deconvolution algorithm argument: ```{r} markers <- lute(singleCellExperiment=exampleSingleCellExperiment, markersPerType=markersPerType, deconvolutionAlgorithm=NULL) length(markers$typemarkerResults) ``` We identified 10 marker genes from the provided data. To run a more complete experiment, we need to define $Y$. For this demonstration, we define $Y$ as a pseudobulk from the provided `exampleSingleCellExperiment` data using `ypb_from_sce()` (see `?ypb_from_sce` for details). ```{r} examplePseudobulkExpression <- exampleBulkExpression <- ypb_from_sce( singleCellExperiment=exampleSingleCellExperiment) ``` By default, the pseudobulk used all available cell data in `exampleSingleCellExperiment`. Finally, we run marker selection and deconvolution in succession with `lute` as follows: ```{r} experiment.results <- lute(singleCellExperiment=exampleSingleCellExperiment, bulkExpression=as.matrix(exampleBulkExpression), typemarkerAlgorithm=NULL) ``` By default, `lute` used the mean ratios algorithm to get markers, then the NNLS algorithm to get cell type proportion predictions. We can inspect the predicted cell type proportions $P$ from `experiment.results`: ```{r} experiment.results$deconvolution.results ``` # Conclusions and further reading This User's Guide introduced the `lute` framework for deconvolution, including an outline of deconvolution experiment variables, an introduction to the different deconvolution methods for bulk transcriptomics, and a small runnable example showing how to access the NNLS algorithm by calling the `deconvolution` generic on an object of class `nnlsParam`. If you found this guide useful, you may also find the other vignettes included in `lute` to be helpful. The vignette "Pseudobulk cell size rescaling example" shows a small example experiment contrasting predictions with and without using cell scale factors using real human brain single-nucleus RNA-seq data accessed from the `scRNAseq` package. More background about cell scale factors can be found in the @maden_challenges_2023 commentary, and examples of their uses can be read about in @monaco_rna-seq_2019, @racle_epic_2020, and @sosina_strategies_2021. # Session Info ```{r} sessionInfo() ``` # Works cited