Skip to content

Dimensionality reduction and clustering#

Motivation#

Dimensionality reduction and cell clustering are fundamental steps in single-cell RNA-sequencing (scRNA-seq) analysis. In this tutorial, we emphasize the primary parameters associated with dimensionality reduction and clustering. We will also delve into PCA loadings, metrics for evaluating clustering, and UMAP visualization

Step-by-step#

In this hands-on section, we will cover several analytical steps. Initially, the pipeline will combine sample matrices into one Seurat object. Subsequently, it will carry out normalization, dimensionality reduction, and clustering. Collectively, these procedures will generate a preliminary clustering used for distinguishing between malignant and non-malignant cells.

1. Running pipeline#

Currently, the pipeline permits users to modify several parameters, encompassing crucial thresholds like the number of highly variable genes and clustering resolution. Please, see the list below.

1.1. On the HPC#

By default the previous command line considers thresholds.

HPC

  • workflow_level = Basic
  • thr_n_features = 2000
  • input_group_plot = source_name,Sort
  • thr_resolution = 0.25
  • thr_proportion = 0.25
nextflow run main.nf --workflow_level Basic --project_name Training --sample_csv sample_table.csv --meta_data meta_data.csv --cancer_type Ovarian -resume -profile seadragon

1.2. On Cirro#

Alternatively, we execute this task on Cirro.

Cirro

  • Defining the pipeline entrypoint = Basic
  • Number features for FindVariableFeatures = 2000
  • Meta-data columns for UMAP plot = source_name,Sort
  • Resolution threshold = 0.25
  • Cell proportion for ROGUE calculation = 0.25

On Cirro, users should (Do not run):

  • Navigate to the Pipelines tab and enter "BTC scRNA Pipeline" in the search engine.
  • Change the Dataset to BTC Training dataset and the Copy Parameters From option to Run_01.
  • Double-check the aforementioned parameters and click Run.

2. Inspecting report#

For convenience the figures can be located in the Test_merged_report.html and Test_main_cluster_report.html reports. These reports are located within the Run_02 dataset.

2.1. Highly variable genes (HVG)#

The first report produces multiple figures, including the HGV distribution on the dataset. In addition, the user can doublecheck which genes is contributing (loadings) to each principal component.

Image caption

As mentioned earlier, loadings represent the genes/features contribution to each principal component or other dimension reduction. Visualizing loadings provides insights into which genes are driving the separation observed in a particular component. In turn, these genes will strongly affect the clustering process.

Image caption

2.2. UMAP and cluster composition#

The next step on the pipeline will perform clustering over the dimensions (e.g. PCs). Here, we can access the clustering profile, composition, and quality.

Image caption

Image caption

The barplot illustrates the cluster composition per sample. Clusters dominated by a single sample might indicate populations of malignant cells.

2.3. Clustering assessement#

ROGUE is an entropy-based metric designed to evaluate cluster purity in single-cell RNA sequencing (scRNA-seq) data. In essence, it assists in refining clustering by suggesting when clusters should be split or merged. High ROGUE scores correlate with purity, meaning clusters consist of cells displaying similar transcriptional backgrounds. Conversely, low ROGUE scores indicate cluster heterogeneity, suggesting clusters comprise varied cell populations and should be further subdivided.

Image caption

The ROGUE score can also provide insights regarding data quality. For example, samples with a low average ROGUE score may contain a higher proportion of doublets.

3. Exercise: Playing around with multiple paremeters#

Question

What would happen if we changed the features and the resolution threshold? A: Run_Dimensionality and Run_Clustering

Please note: When configuring the pipeline on Cirro, ensure that the Dataset is set to BTC Training dataset and select Run_02 for the Copy Parameters From option. Additionally, configure the Entrypoint parameter to Basic.

Reference#

  1. An entropy-based metric for assessing the purity of single cell populations