Dimensionality Reduction for Visualization

Reduction of high-dimensional datasets to 2D for visualization & interpretation.

Repository: openproblems-bio/task_dimensionality_reduction

Description

Data visualisation is an important part of all stages of single-cell analysis, from initial quality control to interpretation and presentation of final results. For bulk RNA-seq studies, linear dimensionality reduction techniques such as PCA and MDS are commonly used to visualise the variation between samples. While these methods are highly effective they can only be used to show the first few components of variation which cannot fully represent the increased complexity and number of observations in single-cell datasets. For this reason non-linear techniques (most notably t-SNE and UMAP) have become the standard for visualising single-cell studies. These methods attempt to compress a dataset into a two-dimensional space while attempting to capture as much of the variance between observations as possible. Many methods for solving this problem now exist. In general these methods try to preserve distances, while some additionally consider aspects such as density within the embedded space or conservation of continuous trajectories. Despite almost every single-cell study using one of these visualisations there has been debate as to whether they can effectively capture the variation in single-cell datasets [@chari2023speciousart].

The dimensionality reduction task attempts to quantify the ability of methods to embed the information present in complex single-cell studies into a two-dimensional space. Thus, this task is specifically designed for dimensionality reduction for visualisation and does not consider other uses of dimensionality reduction in standard single-cell workflows such as improving the signal-to-noise ratio (and in fact several of the methods use PCA as a pre-processing step for this reason). Unlike most tasks, methods for the dimensionality reduction task must accept a matrix containing expression values normalised to 10,000 counts per cell and log transformed (log-10k) and produce a two-dimensional coordinate for each cell. Pre-normalised matrices are required to enforce consistency between the metric evaluation (which generally requires normalised data) and the method runs. When these are not consistent, methods that use the same normalisation as used in the metric tend to score more highly. For some methods we also evaluate the pre-processing recommended by the method.

Authors & contributors

name	roles
Luke Zappia	maintainer, author
Michal Klein	author
Scott Gigante	author
Ben DeMeo	author
Robrecht Cannoodt	author
Kai Waldrant	contributor
Sai Nirmayi Yasa	contributor
Juan A. Cordero Varela	contributor

API

flowchart LR
  file_common_dataset("Dataset")
  comp_process_dataset[/"Data processor"/]
  file_dataset("Dataset")
  file_solution("Solution data")
  comp_control_method[/"Control method"/]
  comp_method[/"Method"/]
  comp_metric[/"Metric"/]
  file_embedding("Embedding")
  file_score("Score")
  file_common_dataset---comp_process_dataset
  comp_process_dataset-->file_dataset
  comp_process_dataset-->file_solution
  file_dataset---comp_control_method
  file_dataset---comp_method
  file_solution---comp_control_method
  file_solution---comp_metric
  comp_control_method-->file_embedding
  comp_method-->file_embedding
  comp_metric-->file_score
  file_embedding---comp_metric

Loading

File format: Dataset

The dataset to pass to a method.

Example file: resources_test/common/cxg_mouse_pancreas_atlas/dataset.h5ad

Format:

AnnData object
 obs: 'cell_type'
 var: 'hvg_score'
 layers: 'counts', 'normalized'
 uns: 'dataset_id', 'dataset_name', 'dataset_url', 'dataset_reference', 'dataset_summary', 'dataset_description', 'dataset_organism', 'normalization_id'

Data structure:

Slot	Type	Description
obs["cell_type"]	string	Classification of the cell type based on its characteristics and function within the tissue or organism.
var["hvg_score"]	double	High variability gene score (normalized dispersion). The greater, the more variable.
layers["counts"]	integer	Raw counts.
layers["normalized"]	double	Normalized expression values.
uns["dataset_id"]	string	A unique identifier for the dataset.
uns["dataset_name"]	string	Nicely formatted name.
uns["dataset_url"]	string	(Optional) Link to the original source of the dataset.
uns["dataset_reference"]	string	(Optional) Bibtex reference of the paper in which the dataset was published.
uns["dataset_summary"]	string	Short description of the dataset.
uns["dataset_description"]	string	Long description of the dataset.
uns["dataset_organism"]	string	(Optional) The organism of the sample in the dataset.
uns["normalization_id"]	string	Which normalization was used.

Component type: Data processor

A dimensionality reduction dataset processor.

Arguments:

Name	Type	Description
--input	file	The dataset to pass to a method.
--output_dataset	file	(Output) The dataset to pass to a method.
--output_solution	file	(Output) The data for evaluating a dimensionality reduction.

File format: Dataset

The dataset to pass to a method.

Example file: resources_test/task_dimensionality_reduction/cxg_mouse_pancreas_atlas/dataset.h5ad

Format:

AnnData object
 var: 'hvg_score'
 layers: 'counts', 'normalized'
 uns: 'dataset_id', 'normalization_id'

Data structure:

Slot	Type	Description
var["hvg_score"]	double	High variability gene score (normalized dispersion). The greater, the more variable.
layers["counts"]	integer	Raw counts.
layers["normalized"]	double	Normalized expression values.
uns["dataset_id"]	string	A unique identifier for the dataset.
uns["normalization_id"]	string	Which normalization was used.

File format: Solution data

The data for evaluating a dimensionality reduction.

Example file: resources_test/task_dimensionality_reduction/cxg_mouse_pancreas_atlas/solution.h5ad

Format:

AnnData object
 obs: 'cell_type'
 var: 'hvg_score'
 layers: 'counts', 'normalized'
 uns: 'dataset_id', 'dataset_name', 'dataset_url', 'dataset_reference', 'dataset_summary', 'dataset_description', 'dataset_organism', 'normalization_id'

Data structure:

Slot	Type	Description
obs["cell_type"]	string	Classification of the cell type based on its characteristics and function within the tissue or organism.
var["hvg_score"]	double	High variability gene score (normalized dispersion). The greater, the more variable.
layers["counts"]	integer	Raw counts.
layers["normalized"]	double	Normalized expression values.
uns["dataset_id"]	string	A unique identifier for the dataset.
uns["dataset_name"]	string	Nicely formatted name.
uns["dataset_url"]	string	(Optional) Link to the original source of the dataset.
uns["dataset_reference"]	string	(Optional) Bibtex reference of the paper in which the dataset was published.
uns["dataset_summary"]	string	Short description of the dataset.
uns["dataset_description"]	string	Long description of the dataset.
uns["dataset_organism"]	string	(Optional) The organism of the sample in the dataset.
uns["normalization_id"]	string	Which normalization was used.

Component type: Control method

Quality control methods for verifying the pipeline.

Arguments:

Name	Type	Description
--input	file	The dataset to pass to a method.
--input_solution	file	The data for evaluating a dimensionality reduction.
--output	file	(Output) A dataset with dimensionality reduction embedding.

Component type: Method

A dimensionality reduction method.

Arguments:

Name	Type	Description
--input	file	The dataset to pass to a method.
--output	file	(Output) A dataset with dimensionality reduction embedding.

Component type: Metric

A dimensionality reduction metric.

Arguments:

Name	Type	Description
--input_embedding	file	A dataset with dimensionality reduction embedding.
--input_solution	file	The data for evaluating a dimensionality reduction.
--output	file	(Output) Metric score file.

File format: Embedding

A dataset with dimensionality reduction embedding.

Example file: resources_test/task_dimensionality_reduction/cxg_mouse_pancreas_atlas/embedding.h5ad

Format:

AnnData object
 obsm: 'X_emb'
 uns: 'dataset_id', 'method_id', 'normalization_id'

Data structure:

Slot	Type	Description
obsm["X_emb"]	double	The dimensionally reduced embedding.
uns["dataset_id"]	string	A unique identifier for the dataset.
uns["method_id"]	string	A unique identifier for the method.
uns["normalization_id"]	string	Which normalization was used.

File format: Score

Metric score file

Example file: resources_test/task_dimensionality_reduction/cxg_mouse_pancreas_atlas/score.h5ad

Format:

AnnData object
 uns: 'dataset_id', 'normalization_id', 'method_id', 'metric_ids', 'metric_values'

Data structure:

Slot	Type	Description
uns["dataset_id"]	string	A unique identifier for the dataset.
uns["normalization_id"]	string	Which normalization was used.
uns["method_id"]	string	A unique identifier for the method.
uns["metric_ids"]	string	One or more unique metric identifiers.
uns["metric_values"]	double	The metric values obtained for the given prediction. Must be of same length as ‘metric_ids’.