Cells growing in a culture dish are not uniform entities; they vary both stochastically and as a consequence of external signals. One source of external variability is the microenvironment of the cells, and as Lucas Pelkmans, Berend Snijder and their colleagues at the University of Zurich have previously shown, microenvironmental properties such as cell density—what they call 'population context effects'—can strongly affect biological processes.

Such population effects could also influence phenotypes that result from a perturbation, for instanct, gene silencing by RNAi. Pelkmans and colleagues now systematically analyze population context effects on siRNA screens in cells and describe methods to understand and counter these effects.

The researchers began with image data from 34 small-scale (about 50 genes) and 7 large-scale (about 7,000 genes) siRNA screens that examined the infection of four different HeLa cell 'strains' with 17 viruses. As in their previous work, they observed that population context can affect the levels of viral infection. They also found that a model of this relationship could predict a large portion of the measured viral infection levels following siRNA treatment. This indicates that some phenotypes in the analyzed screens were in fact due to indirect effects of the siRNA on the cells. “To put it simply,” explains Pelkmans, “if you have a virus that prefers to infect very densely growing cells, and you now silence a gene that reduces cellular growth, you get a population where there are less cells in the right state for infection, and the model can predict that.” Pelkmans emphasizes, however, that population effects are complex and cannot be predicted by the number of cells alone.

In the case of siRNA treatments for which the model fails to fully explain the phenotype, the siRNA is likely to perturb viral infection via direct cellular effects. Following this reasoning, Pelkmans and colleagues used their model to distinguish between siRNAs that cause a phenotype via direct versus indirect effects. They also used Bayesian modeling to learn the causal relationships that lead to the observed phenotypic changes.

What consequence does separating direct and indirect effects have on the outcome of siRNA screens? Most strikingly, for both small- and large-scale viral infection screens, the top-ranked genes changed substantially when hit lists were corrected to remove indirect, population-based effects. For infection by the SV40 virus, removing indirectly acting siRNAs yielded hit lists enriched for genes known to be involved in viral entry.

Correcting for indirect siRNA effects improved the overlap between screens performed on different cell lines and even, in some cases, in different laboratories; these improvements were noted for screens examining several biological properties (organelle abundance, cell size or cellular cholesterol levels) in addition to viral infection. Removing siRNAs with indirect effects also increased the consistency of phenotypes produced by multiple siRNAs targeting the same gene. Pelkmans and Snijder emphasize, however, that the main consequence of implementing their corrections is that the results and their ranking change for a given screen. “What this does is it basically focuses you on a different set of genes,” says Pelkmans, “and as far as we can tell, it focuses much more on the direct regulators of the cell biology.”

Considering that cellular heterogeneity is a widespread phenomenon and that correction for indirect effects improved data quality for several types of image-based screens, Pelkmans suggests that other scientists conducting siRNA screens should consider that population context could affect the processes they are studying as well. If so, methods such as those used in this work could help to refine the resulting gene lists and allow researchers to focus their follow-up efforts on those genes in which they have the most interest.