Fig 1.
(A) The schematic illustration of PanoView algorithm. (B-D) A toy model for the illustration of OLMC algorithm. (B) 500 random points in 2D space. Gray numbers represent the number of neighbors for each point. Colored numbers are three local maximum densities. (C) The histograms represent the distance to local maximums. The heights of colored bars are used for constructing the first convex hull for each local maximum. (D) Color-enclosed circles represent the convex hulls constructed by colored bars in (C) during OLMC algorithm.
Fig 2.
The performance of PanoView in comparison to other existing methods using ten simulated datasets and published scRNA-seq datasets.
(A) The ARI results of 8 different computational methods in 1,200 simulated datasets. Error bars indicate standard deviation of ARI. (B) The ARI result of 10 clustering methods in 11 published single-cell RNA-seq datasets. The order of legend is based on the number of single-cells in descending order. Dots are the calculated ARI values for each dataset. SC3 and pcaReduce did not produce usable clustering results for Campbell dataset. The dataset is missing for these two methods in B and C. (C) The ARI result of 4 datasets that contain more than 3,000 cells. (D): The ARI result of 7 datasets that contain fewer than 3,000 cells.
Fig 3.
Computational times for selected clustering methods.
The X-axis represents the number of cells in 8 datasets. Note that SC3 and pcaReduce did not produce usable clustering result for Campbell dataset.
Fig 4.
The results of PanoView for 10 scRNA-seq datasets with 500 random parameter sets.
(A) Boxplots of 500 simulation results. We ordered the 500 parameter sets based on median values of ARI in ascending order. The blue line indicates the median values of 10 ARI values for each parameter set. The vertical pink line represents the result of PanoView with the default parameters. (B) The distribution of median ARI in 500 simulation results. The default value ranked 98.2 percentile. (C) Boxplots of 500 clustering results in 10 scRNA-seq datasets. Red stars are the ARI results with current default parameters.
Fig 5.
The evaluation of detecting rare cell types.
(A) The recovery rate and false positive rate in detecting rare cell types in 260 simulated datasets. SC3 was not included in the comparison because it did not produce usable results in our simulation (B) The ground truth of the selected simulated data. Cluster 999 represents the predefined rare cell type and the t-SNE coordinates of three rare cells were adjusted for better visualization. (C, D, E, F) We selected one of the simulated datasets to visualize the performance of different computational methods (PanoView, GiniClust, Seurat, SCANPY). RaceID2 did not produce clustering result in this simulated dataset.
Fig 6.
Single-cell clusters of mouse embryonic hypothalamus identified by PanoView.
(A) Visualization of identified cell types from embryonic hypothalamus using t-SNE. (B) Relative expression of 12 selected marker genes in embryonic hypothalamus.