How strongly must a gene be knocked down before cells respond? Enter an essential gene to see its CRISPRi dose–response curve in K562 cells, its ED50 and how reliable the estimate is.
Examples:
About the atlas and how to read a result
Dose
The cell's own target mRNA relative to control cells (1 = median of control cells). CRISPRi knocks the gene down to a different depth in every cell, so one screen works as a titration.
Response score (PS)
How far a cell's transcriptome moved towards that perturbation's typical response (0 = like a control cell).
ED50
The dose at which the response reaches half of its range. Song et al. call a gene buffered when ED50 ≤ 0.5 (cells tolerate a partial loss) and sensitive when ED50 > 0.5 (a mild knockdown already triggers a response).
Per-fold refits
The curve refitted on the training cells of each fold. The three training sets overlap (each uses two thirds of the gem groups), so agreement is a stability check, not an independent replication.
Binary, switch and dial
Binary: every knocked-down cell shifts the same. Switch: some cells respond fully and the rest not at all. Dial: each cell responds partly. They are compared on cells held out from fitting.
Confidence
High when every reliability check passes, moderate with at least one caution, low with at least one failed check. It summarises the checks; it is not a statistical test.
Groups of perturbations that are densely linked in the network (Louvain communities), with their best-connected genes and the dose–response classes of their members. Descriptive only: for many genes the class is unstable.
Limitations
One cell line (K562, a leukaemia line) and only essential genes. Curves in other cell types may differ.
Cells with the deepest knockdown of an essential gene may have died before capture, so the deep-knockdown end of each curve is under-represented.
The per-cell dose comes from a few mRNA molecules. For weakly expressed targets it is mostly noise.
The curves are associations. Cells with less target mRNA respond more, but only a designed titration shows that the knockdown causes the response.
Labels are fragile: many ED50 values are extrapolated, labels change between refits, and Song et al.'s published buffered/sensitive split did not reproduce from the matrix.
CRISPRi lowers mRNA; a drug inhibits protein. The analogy with partial drug inhibition is approximate.
The atlas shows measured perturbations only. It does not predict genes or combinations that were not in the screen, and it is not for clinical use.