Control cells

The challenge asks every team to predict what silencing a gene does to six cell lines, named only by letters: A, B and C were the validation round, and D, E and F decide the final ranking. For each line Arc released 18,400 control cells, cells that carry a guide designed to silence nothing (46 such guides, 400 cells each). They are the only cells of those lines anyone gets, so every prediction starts from them. This page runs the quality checks a single-cell study opens with on those controls.

Maps and QC. Each dot is one cell, and cells with similar expression sit together. Validation, Test and All six put those contexts on one set of axes, so contexts separate by identity and by assay; a single letter shows that context on its own map, which is where what varies inside one context shows. Regressed first removes each cell’s depth and its two cell-cycle scores from every variable gene, so what still forms an island is where a clone, a sub-population or a misbehaving guide would show. Violin shows the same cells as one QC measure, one violin per context, cluster or phase. A card in the rail shows its context in every figure. Cluster numbers are local to each map. In the cluster table, a clone would show as a cluster drawn from a few guides (the largest guide’s share far above 1/46, about 2.2 %), and a group of damaged or shallow cells as a cluster spread over every guide with a fraction of the context’s UMIs.

The measures behind the colours:

• UMIs: molecules counted per cell, which is sequencing depth. Arc states a median near 20,000 for its lines.
• Genes: genes detected per cell. It rises with depth; at equal depth, fewer genes means a library made of few transcripts.
• Mito UMI %: the share of a cell’s UMIs on mitochondrial genes. In droplet data a high share marks a stressed or broken cell. These cells were fixed and read with probes (10x Flex), the share sits far lower, and the usual 10–20 % cut-off does not apply.
• Top-50 UMI %: the share held by the cell’s own 50 most-counted genes, high in a cell whose library is made of few transcripts.
• Complexity: log10 genes over log10 UMIs, how quickly new genes appear as depth grows.
• S score and G2/M score: how strongly a cell expresses the DNA-replication and the mitosis genes of Tirosh et al. (2016), against its own context. Phase is the call the two scores make, G1, S or G2/M. The call compares cells within a context, so a context where nearly every cell is cycling gets few G1 calls: read the scores rather than the labels.
• IFN score: the interferon response, an antiviral programme (ISG15, the IFIT genes, MX1 and others).
• Stress score: immediate-early and heat-shock genes (FOS, JUN, EGR1 and others) a cell switches on within minutes of handling.

Guides and clones. If one of the 46 non-targeting guides still moves expression (an off-target site), or a context holds a clone that grew out, its cells would sit apart from the rest. Each dot is one guide: how far its 400 cells' centre sits from the context’s centre, as measured (x) and regressed (y), in standard deviations of 500 shuffles of the guide labels, so 0 is what chance gives. Dots in ink pass q < 0.05 after the Benjamini–Hochberg correction over the 46 guides. The bars are each guide’s median UMIs: a guide whose cells were all shallower would show there first.

Highest-expressed genes. The 20 genes holding the most UMIs, as each cell’s share (box 25–75 %, whiskers 5–95 %), and every gene’s mean count against the share of cells detecting it, with the 2,000 variable genes the context’s own map uses in its colour.

How the numbers were made. Control cells only, 18,400 per context. Counts were scaled to 10,000 per cell and log-transformed; each context's map uses its 2,000 most variable genes (scanpy, Seurat flavour), scaled, a 50-component PCA, a 15-neighbour graph on the first 30 components, UMAP and Leiden clusters (resolution 0.5). The joint map takes the 2,000 genes variable in the most contexts. The guide checks compare each guide's centre with the context's in the first 20 components against 500 shuffles of the labels. The assay is 10x Flex (probe-based) read on an Ultima UG100, per Arc; reads and sequencing saturation are not in the released files, and the Flex panel has no ribosomal-protein genes.

Data: the control cells Arc Institute released for the Virtual Cell Challenge 2026, computed 2026-10-21. Every chart downloads as SVG from its toolbar.