Eight virtual knockout methods predict different effects of gene silencing
Eight virtual knockout methods predict different effects of gene silencing
On 19 August, the authors of a preprint compared eight methods that use single-cell RNA data to predict the consequences of gene silencing. In a test using K562 leukemia cells, the direction predicted by the linear version of CellOracle matched the experimental result for 18 of 44 transcription factor and gene pairs.
Single-cell RNA data show which genes tend to be active together. These patterns can support a hypothesis about a regulatory network and help researchers select a gene for the next experiment. Silencing a gene addresses a different question: how the activity of each specific gene changes after the intervention.
The authors tested four transcription factors, which are proteins that regulate the activity of other genes, and 11 glycolysis genes. Glycolysis is the first stage through which a cell obtains energy from glucose. In Perturb-seq, researchers use CRISPRi to suppress the activity of a selected gene and then measure RNA in individual cells. Silencing each of the four factors reduced the mean activity of these 11 genes. The researchers then compared these measurements with the signs of the coefficients produced by the linear version of CellOracle. The sign of a coefficient describes the relationship between genes in the original RNA data, whereas CRISPRi measures their response to an intervention. The directions matched in 18 of 44 cases.
The original CellOracle has a broader design. It constructs a regulatory network from RNA data and chromatin accessibility data, which identify regions of DNA that are available for transcription, and models the shift in cell state after a factor is silenced. In the current comparison, the authors reproduced its regression step using relationships reported in the scientific literature. The implementation is described in the benchmark repository.
Different programs described as performing a virtual knockout return different types of results. One produces a list of candidates for experimental testing, another models a shift in cell state, and a third predicts the direction of change in a specific gene. The same RNA dataset can therefore produce different target lists because each system measures a different aspect of the intervention's effect.
In the Virtual Cell Challenge, in which models predicted cellular responses to gene silencing, the full RNA profile, the distinguishability of the effect, and the set of genes that changed were evaluated separately. The authors of the current benchmark define their evaluation with similar precision: they compare the predicted direction of change in a specific gene with the change measured after CRISPRi.