×
PATTERNS viewer -
Explore deep-learning derived embeddings of the cells in the PATTERNS dataset. The Swin-based embeddings are 768-dimensional and SubCell-based are 1536-dimensional. The embeddings have been projected to 2 dimensions using UMAP for purposes of visualization and manual inspection of the data. Available embedding models:
Swin-tuned - Swin transformer v2 (Liu et al., 2022) with custom fine-tuning on part of the data, only on protein channel (best performing model)
SubCell-tuned - Subcellular localization model (Gupta et al., 2025) with custom fine-tuning on part of the data, leveraging protein and DNA channels
Swin-raw - Swin transformer v2 (Liu et al., 2022) pretrained on ImageNet by the original authors
SubCell-raw - Subcellular localization model (Gupta et al., 2025) pretrained by the original authors
Randomized - Randomized baseline (embeddings from Swin-tuned model, randomly shuffled in each dimension individually)
All of the 261200 cells in the dataset have been captured in three channels: protein (mNeonGreen), DNA (nuclear stain), and cell membrane. The Swin-based models operate on the protein channel alone, while the SubCell-based models leverage the nuclear channel as well.
PATTERNS viewer - tab
PATTERNS data
Usage
- Protein UMAP
- The Protein UMAP section displays a UMAP plot of embeddings of PATTERNS data samples. The map is zoomable and draggable - you can easily zoom in onto parts of the plot. Hovering over points in the plot displays detailed information on individual samples.
- Landmark cells
- Landmark cells are defined as a small sample of cells per gene belonging to the largest cluster of the given gene, provided the gene is concentrated enough in the given cluster. They are useful for exploring the meaning of the different areas of the UMAP, by sampling random cells from the zoomed in area.
- UMAP options
- The paragraph above the UMAP plot lets you choose how the points in the plot are colored (colored by) and which points are highlighted (filter by). Importantly, the coloring by point density allows to visualise the local density of the UMAP, with a custom filtering available for individual classes of the Combined annotations coloring. Below the plot there are controls that allow you to change the point size and choose the data splits to explore. When viewing Combined annotations coloring, you can choose a standard palette for contrast, or custom palette - with colors assigned based on meta-compartments (e.g., nuclear locations get shades of blue/violet), also you can view only genes with a confidence above a chosen threshold. Clicking on a point in the UMAP selects the corresponding cell.
- Selected cell and its neighbors
- The selected cell is identified with its unique ID. Also a cell can be selected by editing the ID directly.
- When a cell is selected it appears in the top left corner in the Selected cell and its neighbors section, highlighted with a box. Next to it, previews of 25 cells that have most similar embeddings are revealed. Note these are the nearest neighbors in the high-dimensional (768- or 1536-dimensional) spaces, not the UMAP projections. Hovering over the cells reveals their metadata, clicking on a cell selects it.
- Selection of random cells in visible UMAP area
- After clicking Shuffle random cells, this section provides a preview of randomly chosen cells which are highlighted (if a filter is chosen) and are currently visible in the UMAP plot. This way, you can zoom in to a specific section of the UMAP and browse a random selection of cells from this area.
- Preview channels
- Allows control of which channels are visible in the cell previews.
The viewer is a result of a collaboration between the JSB Lab of the University of Bonn and Appsilon.
Protein UMAP colored by
Combined annotation
Filter cells by