CRISPR screening as a service

A pooled CRISPR screen asks which genes matter for a chosen phenotype, by disrupting each gene in a different cell within one mixed population and then letting the phenotype change the population’s composition. The read-out is not microscopy or assay signal; it is counting. Guide sequences are amplified from genomic DNA before and after selection and sequenced, and a gene’s importance appears as the change in abundance of its guides. Everything difficult about screening follows from that being a counting experiment.

One guide per cell, and enough cells per guide

Two quantities set whether the experiment can answer anything.

The library is delivered by lentivirus, and viral infection is a random process. At a high multiplicity of infection many cells receive several guides, and the resulting phenotype cannot be attributed to any one of them. Screens are therefore infected at low multiplicity — commonly targeting a regime where most transduced cells carry a single integration — and the untransduced majority is removed by selection. Accepting that most of the culture is discarded is the price of interpretability.

Coverage is the second quantity: how many cells carry each guide. Because the read-out is a change in relative abundance against sampling noise, a guide represented in too few cells produces a count that fluctuates for reasons unrelated to biology. Screens are planned in cells per guide and that number must be maintained at every step — infection, selection, each passage and the final harvest — since it is the smallest bottleneck, not the starting culture, that fixes the statistics. Multiple independent guides per gene, plus non-targeting and safe-harbour controls, are what let a gene-level call be separated from a single guide’s idiosyncrasy.

Dropout and enrichment are different experiments

A viability or dropout screen simply lets the population grow. Guides against genes required for proliferation become rarer; the signal is negative selection, and it identifies essential genes and, in the presence of a drug, genes whose loss confers sensitivity. The limits are intrinsic: it can only see phenotypes that change growth rate, it needs enough population doublings for depletion to accumulate, and a slow-growing hit is hard to distinguish from noise.

A positive selection screen — survival under a lethal drug or a toxin — is statistically much easier, because survivors are enriched from near-zero background rather than depleted from a large one.

A reporter-based sort measures something a viability screen cannot reach: expression of a specific gene, a pathway readout, a surface marker. Cells are sorted into bins by fluorescence and each bin sequenced, so the phenotype is molecular rather than fitness-based. The cost is that sorting imposes its own throughput limit, and coverage must survive it.

What the result is

A screen returns a ranked list of candidates, and the hit list is a hypothesis. Validation with individual guides, and ideally an orthogonal perturbation, is part of the experiment rather than a follow-up. The nuclease’s own behaviour — repair outcomes, off-target cutting — is treated on the genome editing page; here it enters only as a source of false positives, since cutting itself is toxic in proportion to copy number, and amplified regions drop out for reasons that have nothing to do with the gene.