Journal: BMC Bioinformatics

Article Title: Statistically principled feature selection for single cell transcriptomics

doi: 10.1186/s12859-025-06240-y

Figure Lengend Snippet: Performance of randomly selected genes as features. A UMAPs of PBMCs calculated using either HVG-selected genes (left) or the same number of randomly selected genes (right). B Adjusted Rand index (ARI) and normalized mutual information (NMI) of prediction of cell type as a function of number of genes selected. The predictions were generated by a linear Support Vector Machine (SVM), trained on the top principal components (PCs) calculated on randomly selected genes. Ten sets of random genes were selected at each step. C UMAP of CD+ T cells, including the T regulatory cell (Tregs) subset, calculated using either HVG-selected genes (left) or same number of randomly selected genes (right). D ARI and NMI of Treg identification, as a function of number of randomly selected genes. The predictions were generated by a linear Support Vector Machine (SVM) as in panel B. Twenty sets of random genes were selected at each step.

Article Snippet: To compare the speed and memory usage of HVGs, SCT, FvF and BigSur when varying numbers of cells, we used the 1M PBMC dataset from Parse Biosciences.

Techniques: Generated, Plasmid Preparation

Journal: BMC Bioinformatics

Article Title: Statistically principled feature selection for single cell transcriptomics

doi: 10.1186/s12859-025-06240-y

Figure Lengend Snippet: Clustering performance using values and significance levels of modified corrected Fano factors. The Fano factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi$$\end{document} ( A ) and the modified corrected Fano factor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} ( B ) were calculated for all genes in the CD4+ T cell dataset and are plotted as a function of gene expression level and colored according to the p -value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} . C Values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} and their p -values for genes in the CD4 + T cell are plotted against one another. D–E Enrichment scores (see main text) for FOXP3 after Leiden clustering using features selected using different p -value and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} cutoffs. Each point represents an individual cluster. Panel D shows the enrichment scores using solely a \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varphi^\prime$$\end{document} threshold (top) or solely an adjusted p -value threshold. Panel E displays the enrichment scores as the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} is increased. Only genes with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varphi^\prime$$\end{document} p -values < 0.05 were used. F UMAP of the CD4 + T cell dataset with genes \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime < 3$$\end{document} and p -value < 0.05 along with each cluster’s F OXP3 enrichment (the enrichment of clusters 0 and 10 were 0.04 and 0, respectively). G–H For each of six datasets, the percentage of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} that were significant was plotted against the number of cells (panel G) and the median sequencing depth (panel H). I for the datasets in panel H, as well as the full PBMC dataset (displayed in Fig. ), the percentage of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} that are significant was plotted against the total number of transcripts in the dataset. J–K The highest enrichment score for six datasets (displayed in panel F) as a function of features selected using different quantile thresholds for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} (abscissa) and different p -value cutoffs (by color). The Leiden clustering algorithm was run on the selected features using 40 different randomly chosen starting seeds. For each Leiden seed, the enrichment level in the cluster with the greatest enrichment was found. Each point on the plot indicates the median of this value over the different Leiden seed choices, and the bars show the inter-quartile range (25–75% of the data).

Article Snippet: To compare the speed and memory usage of HVGs, SCT, FvF and BigSur when varying numbers of cells, we used the 1M PBMC dataset from Parse Biosciences.

Techniques: Modification, Gene Expression, Sequencing

Journal: BMC Bioinformatics

Article Title: Statistically principled feature selection for single cell transcriptomics

doi: 10.1186/s12859-025-06240-y

Figure Lengend Snippet: BigSur improves clustering performance using fewer total features. For each of four datasets (two PBMC datasets, a skin dataset and a macrophage dataset, see main text), 20 “semi-synthetic” datasets were generated (see methods). A Bars show the cell numbers for each cell type in each semi-synthetic dataset. B The percent of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varphi^\prime$$\end{document} values with p < 0.05 is plotted as a function of the median UMI/cell per dataset. C The number of features selected by each method for each semi-synthetic dataset (see main text for BigSur’s cutoffs). D–H Purity scores of each semi-synthetic dataset as the feature selection method was changed. For each of the semi-synthetic datasets, Leiden clustering, using 40 different random starting seeds, was performed using features that were either a random selection of 2,000 genes, or those chosen by HVGs, SCT, FvF and BigSur. Purity scores were calculated (as in Fig. ). Purity scores of semi-synthetic datasets generated from the 10k T cell, 1M T cell, keratinocyte and macrophage datasets are shown in panels D, E, F and G respectively. Each line represents a different semi-synthetic dataset (for clarity, only 10 of the 20 are shown in each plot; see Fig. for the remaining data), and the markers and bars are the medians and interquartile ranges (IQRs) of the purities (over the 40 different Leiden seeds). The purities from selected semi-synthetic datasets are colored for discussion purposes (see main text). H Median purities of semi-synthetic datasets. The markers represent medians and the bars represent IQRs. I Median purities of 20 semi-synthetic datasets generated from four datasets included in the Tabula Sapiens (TS) human cell atlas (see main text). Purities were calculated using the same procedure as for the data shown in panel H. J Purity scores of semi-synthetic datasets generated from the macrophage (TS) dataset with varying percentage of rare to total cells, holding the number of total cells constant. 20 semi-synthetic datasets were generated for each percentage.

Article Snippet: To compare the speed and memory usage of HVGs, SCT, FvF and BigSur when varying numbers of cells, we used the 1M PBMC dataset from Parse Biosciences.

Techniques: Generated, Selection